Over the last few months, AI risk has burst into the public and political consciousness in a big way. A combination of real cyber incidents, high-profile resignations, and statements from lab leaders has brought the discussion of catastrophic risk to a level of prominence that just a year ago seemed highly improbable.

While there are plenty of catastrophic risk estimates out there (including from our forecasters), we know much less about what people think about how policy can address risk from AI. Is AI risk addressable by policy? If so, which are the right policy levers to pull, and what is the likelihood of them being implemented?

The Longitudinal Expert AI Panel (LEAP) brings together forecasts from the kinds of experts that policymakers look to when considering the future of AI: computer scientists, AI industry experts, top economists, policy experts, and AI risk experts. In Wave 12 we asked our panelists to consider how far various policies would go toward addressing catastrophic risks from AI. We also asked panelists which policies they supported, and which ones they thought had the highest chances of being implemented.

We’ve also recruited a new expert sample, composed of people who work on addressing the risks posed by transformative AI. In this and subsequent work, we’ll report results from this “AI risk experts” group as well as our other experts, superforecasters, and our representative sample of the US general public.

The results from our policy forecasts are encouraging. LEAP panelists believe that specific policies can substantially reduce the risk of AI-caused catastrophes, and experts largely agree on which policies should be implemented, despite their disparate professional backgrounds and views on AI. Across groups, clear majorities support establishing an international body with pre-release authorization power over frontier models and expect doing so to have among the largest risk-reducing effects. Majorities of most groups also support policies to slow AI progress and strengthen AI labs’ liability for harms caused by their models, policies most forecasters also expect to meaningfully reduce risk.

However, when it came to implementing these policies, forecasters were much more pessimistic. The policies most expected to reduce risk were also the ones that respondents thought were least likely to be implemented. The one policy that was associated with an increase in risk—federal preemption of state AI laws—was seen as the most likely to pass. Panelists also noted throughout that although bilateral (US–China) versions of policies were more likely to reduce catastrophic risk from AI, they were also less likely to be implemented.

In essence, this wave highlights a contradiction at the heart of the AI policy debate. Forecasters believe that policy really can reduce risks from AI, but are pessimistic that governments will put in place the policies that have the highest impact.

Read on for more detailed insights from our wave of policy forecasts. You can also read the full report, with a question-by-question breakdown and forecaster rationales on the LEAP website.

Policies

We asked forecasters to consider the following policies:1

• P1: Federal preemption of state AI laws

  • A US federal statute or action barring states from regulating AI more strictly than the federal government.

• P2: AI capability slowdown

  • Unilateral decision by US policymakers or a bilateral/multilateral decision made jointly by policymakers in China and the United States to slow the development of frontier AI systems.

    • P2.a: US enacts a policy to slow AI development, with no qualifying measure binding the US and China in force.

    • P2.b: US and China agree to a single bilateral or multilateral measure to slow AI development.

• P3: Pre-release authorization for frontier models

  • A binding requirement that frontier models be vetted and authorized before release, imposed by a US federal body or by an international body.

    • P3.a: A US federal body holds pre-release authorization power.

    • P3.b: An international body whose member jurisdictions include the US and China holds pre-release authorization power over frontier models.

• P4: Strict liability

  • A US federal statute treating frontier model training and deployment as an abnormally dangerous activity subject to strict liability (i.e., a regime under which a developer is liable for harms its models cause without the injured party having to prove the developer was careless or the model defective).

• P5: Bundle of policies (P2.b, P3.b, and P4)

  • Assume all three policies are in force: a binding US–China measure slowing frontier AI development, an international body (with member jurisdictions including the US and China) holding pre-release authorization power over frontier models, and federal strict liability for frontier AI harms.

Key insights

Respondents disagree significantly about how risky AI is, yet still largely agree on what to do about it

We started by asking forecasters to give us their baseline all-things-considered forecasts of the likelihood of an AI-caused catastrophe that leads to the deaths of more than 10% of humans by 2050. Here’s what they forecast:2

• Experts: 2%

• Superforecasters: 1.6%

• AI risk experts: 5.2%

• General public: 5%

As these forecasts show, there is wide disagreement about the baseline level of risk. AI risk experts give risk estimates that are more than double those of our general expert sample, and three times higher than superforecasters.

Despite this disagreement about the level of baseline risk, panelists largely agreed about which AI risk policies should be implemented. A majority of forecasters in each group agreed that the following four policies should be implemented:3

• Pre-release authorization requirement (international body, including US and China)

• Pre-release authorization requirement (US-only)

• Strict liability for developers

AI capability slowdown (US–China)

The only policy that did not win majority support in any group of forecasters was a federal preemption of state AI laws. This was supported by just 29% of experts, 14% of superforecasters, 20% of AI risk experts, and 50% of the general public. The most divisive policy was a US-only capability slowdown—34% of experts and 35% of superforecasters thought this policy should be implemented, compared with 60% of AI risk experts and 65% of the general public.

Respondents across groups believe policy can substantially reduce catastrophic AI risk

After they gave us their baseline forecasts of AI-caused catastrophic risk, we asked forecasters to predict the same outcome but this time to assume that each of the policies described above was implemented immediately by the relevant governing body.

Experts forecast lower catastrophic risk under five of seven policies in our policy menu, indicating strong belief across forecasters that policies can address a substantial portion of AI-caused catastrophic risk.4 Experts and superforecasters expected an international pre-release authorization body for frontier models to be the most effective single policy at reducing risk, while AI risk experts predicted the largest effects from a US–China slowdown agreement.

All groups expected large reductions in risk from an international pre-release regime or a US–China slowdown agreement, forecasting that both policies would reduce risk by between 20% and 50% compared with the baseline. The three-policy bundle of US–China capability slowdown, international pre-release authorization, and strict liability was associated with the largest decrease in risk of any policy we tested—a forecast 72% reduction in risk from experts, 62% reduction from AI risk experts and a 56% reduction from superforecasters.

Forecasters strongly support the implementation of an international body with pre-release authorization power over frontier models

All groups strongly supported a policy that would give an international body whose member jurisdictions include the US and China pre-release authorization power over frontier models. In terms of support for implementation, this was the most-favored policy by experts (75% in favor), superforecasters (86% in favor), and the general public (84% in favor). For AI risk experts this was the second-most-favored policy (86% in favor), narrowly behind a US–China capability slowdown, which 87% of that group favored.

Forecasters also agreed that an international pre-release authorization regime would significantly reduce the risk of an AI-caused catastrophe by 2050. Compared with their baseline risk probabilities, forecasts of an AI-caused catastrophe fell by at least 23% when conditioned on implementation of this policy:

• Experts: 2% → 1% (-50%)

• Superforecasters: 1.6% → 1% (-38%)

• AI risk experts: 5.2% → 4% (-23%)

• General public: 5% → 3% (-40%)

For experts, superforecasters, and the general public, only the policy bundle was associated with a greater decrease in AI-caused catastrophic risk.

US-only policies are forecast to significantly reduce AI risk, but are still seen as less effective than bilateral policy

For two policies, we asked forecasters to predict the effect of a version implemented by the US only, and a version implemented bilaterally or multilaterally by the US and China. These were a decision by policymakers to slow the development of frontier AI systems (P2), and a binding requirement that frontier models be vetted and authorized before release, imposed by a US federal body or by an international body (P3).

The bilateral version of both of these policies was associated with a larger reduction in catastrophic risk compared with a US-only version of the policy. For example, experts gave a baseline 2% chance of an AI-caused catastrophe by 2050. This risk was unchanged if a US-only AI capability slowdown was implemented, and fell to 1.2% if this slowdown was implemented in both the US and China. For the pre-release authorization policy, the chance of an AI catastrophe fell to 1.5% (US-only version) and 1% (international version). This same pattern holds true across forecaster groups and when forecasters assess the chance of a smaller “sub-catastrophe,” which causes at least 1,000,000 deaths within a five-year period.

Despite bilateral policies being viewed as the most impactful, forecasters thought they were less likely to be implemented when compared with US-only versions of the same policy. Experts assigned a 20% probability to a pre-release authorization policy being implemented in the US before the end of 2030, compared with a 5% chance of an international version of the policy being implemented by the same date. For the AI capability slowdown policy, experts gave an 8% chance of a US-only version being implemented and a 5% chance to a bilateral version being implemented.

Experts strongly oppose federal preemption and think its passage would increase AI risk. But they also think it's the policy most likely to pass

A federal preemption of state AI laws was the only policy associated with an increase the risk of an AI-caused catastrophe by 2050 by all forecaster groups. These are the median forecasts of each group in the unconditional scenario, and then conditional on a federal preemption of state AI laws being passed, with the increase in risk in such a scenario shown in brackets:

• Experts: 2% → 3% (+50%)

• Superforecasters: 1.6% → 2.5% (+56%)

• AI risk experts: 5.2% → 8% (+54%)

• General public: 5% → 6.9% (+38%)

Federal preemption also had the lowest support for implementation of any of the policies we tested. Just 29% of experts supported implementation, with 14% of superforecasters, 20% of AI risk experts, and 50% of the general public in support. Despite this, experts assigned this policy the highest likelihood of passing—forecasting a 26% chance that such a policy is implemented by the end of 2030, rising to 40% by the end of 2050. This policy was also forecast as the most likely to pass by 2030 by the general public, with superforecasters and AI risk experts placing slightly more likelihood on pre-release authorization (US only) being implemented by 2030.

In their rationales, forecasters noted that companies often choose to comply with the strictest state standard, setting a national "floor" for safety standards, which a federal preemption would remove. Several forecasters argued that preemption's impact on catastrophic risk would be minimal, as much of the state-level regulation that might be preempted was unlikely to focus on catastrophic risk, and instead would address other concerns such as discrimination, bias, data exploitation, and so on.

The policies expected to reduce risk the most are those respondents think least likely to be implemented

We asked forecasters to predict the probability that the relevant governing body/bodies will implement each policy by the end of 2030 and 2050. Across all groups, most respondents place the probability of implementation by the end of 2030 at or below 25% for almost all policies.

Experts predict the largest reductions in risk under international pre-release approval, US–China slowdown, and strict liability. They also think these policies are least likely to be implemented, giving all three a 5% chance of being in place by 2030.

Conversely, experts give the highest chance of implementation to federal preemption (26%), under which they expect risk to increase.

Despite placing a high probability of implementation of federal preemption and US pre-release authorization, forecasters expect these policies to be among the least effective at reducing catastrophic AI risk. In contrast, the bundle of three policies, which is ranked as the most effective at reducing catastrophic AI risk compared to the baseline by all groups, is seen as the least likely to be implemented. For instance, experts expect the bundle of policies to reduce catastrophic AI risk by nearly 72%, but give only a 1% chance of these policies being implemented. This trend is seen across other policies where forecasters predict catastrophic AI risk will be lowered, including capability slowdowns (P2), pre-release authorization, and strict liability (P4).

Comparing EA-affiliated forecasters with other forecasters

Recent discourse about AI risk has focused on effective altruism (EA) and the differences in beliefs between EA-influenced individuals and other groups. We anticipated this discussion when designing LEAP, and we can explore this dynamic in our LEAP sample by splitting respondents into two groups:

1) Experts with connections to EA, including grant recipients and employees of EA-adjacent organizations, based on our sample definition last year5

2) Experts with no connection to EA

The figure below shows that these two groups hold virtually identical preferences about which policies they think should be implemented and have directionally similar beliefs about the impact of policies on catastrophic risk.

In other words, our headline results are not driven by experts connected to effective altruism. Instead, they represent a broadly shared view of a representative sample of experts.

Find out more

For full results from Wave 12, visit the LEAP website.

This is the second post in a short series on the accuracy of AI progress forecasts we’ve gathered at the Forecasting Research Institute (FRI). In the previous post, we explored the accuracy of the AI progress forecasts we’ve collected so far.

This post was authored by Josh Rosenberg, Matt Reynolds, Nadja Flechner, and Zack Devlin-Foltz.1

In the previous post we looked at AI progress forecasts from across FRI’s work. Broadly, we found that the median expert, superforecaster, and member of the public have consistently underestimated AI progress on near-term capability questions such as benchmarks. They also underestimated revenue growth at AI companies but have a more mixed track record on other adoption indicators. We don’t have enough evidence on forecasts about macroeconomics or other broad social impact forecasts to draw conclusions about their track record in those areas.

Our findings suggest that median forecasts from these groups are not reliable estimates for predicting the trajectory of AI on many questions—particularly those concerning benchmarks and AI capabilities.

This raises an obvious question: If we want to get a better grasp of the trajectory of AI and its impacts on society, whose forecasts should we listen to?

We’re working on three major changes to the Longitudinal Expert AI Panel (LEAP) to address this question:

1. We will share forecasts from a subsample of panelists who expect fast AI progress by 2040.2

2. We are adding regularly updated LLM forecasts to LEAP findings.

3. We will highlight forecasts from the most accurate LEAP participants as soon as we have sufficient data to make this judgment rigorously.

We’re also planning on putting more emphasis on the rationales of forecasters who make particularly insightful arguments as part of their forecasts, and will make some changes to our elicitation process, like allowing forecasters to update after seeing each other’s forecasts and reasoning.

In the rest of the post, we go into more detail on these changes and give a glimpse of what LEAP forecasts might look like in the future with these new additions.

Sharing forecasts from a “fast AI progress” worldview

As we saw in the last post, the majority of the LEAP panel so far has underrated AI progress on benchmarks and other capability indicators, as well as some near-term proxies of AI adoption such as revenue at top AI companies. This was highlighted recently with the news that OpenAI has likely settled the Navier–Stokes problem, surprising LEAP experts and superforecasters who gave median forecasts of 10% and 5.4% that a Millennium Prize Problem would be solved with AI assistance before the end of 2027.

However, a subset of the LEAP panel predicts much faster AI progress and, because of this, has been more accurate on the near-term questions we highlighted in the first post in this series. This subsample puts more weight on the possibility of transformative AI arriving soon than most LEAP forecasters, and we’re interested in what this group thinks about other AI progress questions. By breaking out this sample, we’ll address a core critique of typical LEAP experts’ forecasts: that they greatly underrate how fast AI progress is moving and therefore are giving an inaccurate picture of how the world will look going forward.

Perhaps the “fast AI progress” subgroup will be more accurate in general, or maybe they will be particularly good at forecasting capabilities while doing worse on societal impacts, for example. Our goal is to highlight this group’s forecasts for people who want to put more weight on them, and to explicitly track where this worldview is more and less accurate.

We explore several alternative ways to select this subsample in more detail in the brief appendix below. In our forthcoming work, we plan to share more about what this group forecasts about the future of AI and why.

LLM forecasts of LEAP questions

Having LLMs forecast LEAP questions could be a powerful way to make our forecasts more accurate and timely. We know from our forecasting benchmark, ForecastBench, that top LLMs are at parity with superforecasters on specific kinds of forecasting questions, and we expect LLMs’ forecasting abilities to increase as models become more capable.

LLMs also provide an advantage in forecasting areas with rapidly changing information, such as AI progress. Since we run LEAP surveys every month—and ask most questions only once—awkward timing can sometimes mean that our panel’s forecasts are based on information that quickly becomes out of date.

For example, in Wave 11 of LEAP we asked participants to forecast the annualized revenue of Anthropic and OpenAI. At the time of forecasting (July 13–August 11, 2026), Anthropic’s reported annualized revenue was $47 billion, but just a few days after the survey closed, this figure was updated to $65 billion. While it’d be tricky to re-ask questions of human forecasters, it is relatively trivial to ask LLMs the same questions and see how their forecasts change over time.

We plan to run regular LLM forecasts on key LEAP questions, allowing us to provide high-quality, up-to-date forecasts across key questions. In a later post, we’ll describe how we plan to prompt and ensemble our LLMs to provide the most useful forecasts.

We see these LLM forecasts as a supplement to—not a replacement for—human forecasts. LLM forecasts still have many limitations, and we plan to regularly assess their strengths and weaknesses, including by building new benchmarks that assess LLM skill at AI-focused questions, low-probability questions, conditional questions, and more.

Highlighting more accurate LEAP forecasters

At the moment, we’re unable to confidently identify the most accurate LEAP forecasters because of the small number of questions that have definitively resolved. We’ve made this easier by adding more questions that resolve in the near term, and by the end of 2026 we will have roughly 16 additional resolved questions that should enable us to start identifying more accurate forecasters.

In the new year we plan to publish a thorough analysis of LEAP accuracy so far, and as part of that analysis we will identify a group of more accurate LEAP forecasters. We’ll break out this more accurate sample in our LEAP reports to draw particular attention to their forecasts, and may also explore ways to increase the prominence of these more accurate forecasters within the LEAP sample.

Example of future LEAP forecasts

The charts below are examples of the kinds of forecasts we plan to share in future LEAP waves. The “fast AI progress” sample represented here is from a rough subsample that fulfills the inclusion criteria outlined in the appendix. The exact composition of this group and the criteria we use to select them may change, although the forecasts below are a good indication of how the views of this group differ from those of experts and superforecasters. The LLM forecasts are the median of 160 pooled runs from eight models, run in May 2026. As with the “fast AI progress” sample, the exact methodology we use to derive and display our LLM forecasts may change.

We’ve also recruited a new group of LEAP experts from people who work on mitigating catastrophic risk from AI. The forecasts from this group are not included in the graphs below as they have only forecast a small number of LEAP questions, but this subsample will be referred to in future LEAP reports. We include more information on how this group was selected in the appendix below.

Appendix on the “fast AI progress” subsample

There are several ways to select a subsample of LEAP panelists who place more weight on transformative AI arriving soon than other panelists.

One approach is to do this relatively. We could select a set of questions that measure expectations about AI progress, then select the top 20% or 10% of panelists based on how quickly they expect AI to progress. For example, we could select the 20% of panelists who gave the highest median forecasts on a set of benchmark or other capability questions across LEAP waves.

This approach is easy to describe and would capture a large subsample of LEAP participants. It would also make it easy to define comparison groups, as we could compare this group with the 20% of forecasters who expect the slowest progress, and so on, through the different quintiles. However, this approach does not guarantee the inclusion of participants who believe that AI will be transformative on relatively short timescales and would not define the subsample in terms that believers in the relevant hypothesis would necessarily agree with. (E.g., if our sample happens to not include many very fast AI progress forecasters, choosing the top 20% could still lead to conservative estimates of AI progress.)

A second approach would be to define this group by relevant professional expertise. We could highlight the median forecasts of a group of experts who work in the field of responding to risks from transformative AI. Understanding this group’s worldview is an important task, particularly as these views are becoming more prominent in AI policy discussions. Because of this, we’ve recruited a subsample that represents these views, and we will present these as a distinct and separate subsample in future LEAP waves. We call this subsample our “AI risk expert” subsample.

However, this subsample is not necessarily a good proxy for people who believe that AI progress will be very rapid. For example, median AI progress forecasts of the AI risk expert subsample were not dramatically higher than those from other experts. They fell short of the implied forecasts from senior AI company staff members, for example, suggesting that this group does not fully capture a truly “fast AI progress” worldview. We plan to report median forecasts from the AI risk expert subsample, but will make it clear that this is a separate sample from our “fast AI progress” group. We’ll share more details on how we recruited this group and what they believe about the future of AI in a forthcoming post.

The final approach is to select forecasters in more absolute terms. In the case of this sample, we could include candidates whose forecasts meet all three of a set of criteria, for example:

• A median probability of at least 90% that by 2100 more than 50% of LEAP panelists will agree that AGI exists.

• Conditional on AGI existing, a median forecasts that it arrives in or before 2040.

• A combined probability of at least 50% of AI reaching either Level 9 (technology of the millennium) or 10 (technology of the epoch) on the Technological Richter Scale by 2040.

The absolute approach involves selecting somewhat arbitrary cutoffs for who belongs in vs. out of the subsample. But it also sets a higher bar for beliefs about AI progress and therefore provides a cleaner test of the “fast AI progress” hypothesis. We want to choose a group that people who hold a fast AI progress worldview would feel properly represents them.

We will share more details about our final decision on how to select this sample, how many forecasters are in the “fast AI progress” subsample, and what they believe about the future of AI in forthcoming work.

This is the first post in a short series on the accuracy of AI progress forecasts we’ve gathered at the Forecasting Research Institute (FRI). In the next post, we’ll explore how to improve the accuracy and relevance of AI forecasts in our studies going forward.

This post was authored by Josh Rosenberg, Matt Reynolds, Nadja Flechner, and Zack Devlin-Foltz.1

Summary

• Since mid-2022, FRI has collected forecasts on AI progress across many projects. In this post, we cover results so far on how accurate experts and superforecasters have been at forecasting AI progress across a range of indicators: benchmarks and non-benchmark capability tests, adoption and diffusion metrics, social impacts, and geopolitics and governance.

• In brief, across our studies, forecasters dramatically underestimate AI progress on benchmarks, have a more mixed track record on AI adoption and diffusion though often lean toward underestimation, and there is not yet enough evidence to assess their track record on predicting macro-scale economic and societal impacts of AI.

• There are some examples of forecasters overestimating AI-related progress. In a mid-2025 study they overestimated the impact of LLMs on improving amateurs’ ability to do dangerous biology tasks, and the median expert looks very likely to have majorly overestimated how widespread the use of autonomous vehicles would be by the end of 2027.

• Forecasts on some benchmarks (e.g., Chinese model performance on the Epoch Capabilities Index) and some diffusion-related questions (e.g., the rate of data center growth) appear on track to be broadly accurate.

• There are some important limitations to keep in mind with these findings. For example:

  • We discuss many questions that have resolved early because AI milestones were reached sooner than expected, but highlighting these early-resolving questions may introduce bias, since by definition we do not have data on late-resolving questions. This may lead us toward identifying more areas where forecasters have underestimated AI progress rather than those where they have been accurate or overestimated progress.

  • For some unresolved questions we have used LLM projections to draw tentative accuracy conclusions—by their nature, these projections are more speculative.2 We plan to continue to publish accuracy updates over time, particularly for our Longitudinal Expert AI Panel (LEAP).

In the next post in this series, we’ll discuss how we might improve the accuracy and relevance of LEAP forecasts.

Introduction

There is major disagreement and uncertainty about the trajectory of AI progress, with thoughtful and engaged researchers looking at similar evidence and drawing very different conclusions. For example, Ajeya Cotra and Timothy B. Lee disagree by about 40 years on when they expect self-sufficient AI systems to exist. Andrej Karpathy has argued that achieving AGI would have little to no impact on GDP, while Epoch researchers have modeled that AI automation could lead to economic growth in excess of 30% a year.3 These disagreements are large, but even they only capture a relatively narrow section of the huge range of views that exist on AI progress.

At the Forecasting Research Institute, we want to bring more rigor to the AI progress discussion and help policymakers understand where we’re headed. We have been eliciting forecasts on AI progress since our Existential Risk Persuasion Tournament in mid-2022 and have continued to gather AI progress forecasts in specific domains, including biorisk and cyber. In mid-2025 we launched the Longitudinal Expert AI Panel (LEAP)—a monthly survey that gathers AI progress forecasts from a panel of leading experts and superforecasters. We want to clarify what’s likely to happen with the future of AI, help people understand where forecasts are more or less reliable, and identify particularly accurate forecasting approaches that might give us more insight on the future.

In this post, we’ll provide an update on how accurate forecasters—both domain experts and generalist forecasters with a strong track record (”superforecasters”)—have been across FRI studies. In the next post in this series, we’ll take a look at a few different approaches to improving the accuracy and relevance of LEAP forecasts.

Methodology

In this post, we considered forecasts across all relevant FRI studies from mid-2022 to August 2026: our Existential Risk Persuasion Tournament (XPT) and initial XPT follow-up, AI adversarial collaboration, LLM-biorisk study and Active Site RCT forecasting study, AI-cyber risk pilot, Forecasting the Economic Effects of AI, and the Longitudinal Expert AI Panel (LEAP).4

In our summary of findings below, we highlight a subset of forecasts that we believe are broadly illustrative of the core trends we have observed across studies. The number of actual AI forecasting questions across all of our studies is much larger than what we cover in this post. We focused on a reasonably representative subset of questions to keep this post simpler and more concise, which meant making subjective choices about which forecasts to include or exclude. In general, we included questions where there were either known resolution values or other data that gave us some indication of whether forecasters were on track to be accurate or not.

You can view all of the forecasts that informed this analysis in this table: Forecaster Accuracy on Selected AI Progress Questions and compare those to the full set of AI progress questions we’ve elicited across projects.

We split our AI progress forecasts into four main categories: capabilities, adoption, impacts, and geopolitics and governance:

• Capabilities: Predictions about AI performance on benchmarks or other clearly defined tasks. For example: performance on FrontierMath; when AI will solve a Millennium Prize Problem; when a robot will successfully complete the “Coffee Test.”

• Adoption: Predictions about the diffusion and adoption of AI in “real-world” settings. For example: the fraction of work hours assisted by AI; the proportion of adults reporting AI companion use; the global supply of industrial robots.

• Impact: Predictions about the impact of AI on key outcomes. For example: entry-level tech hiring; GDP growth and labor force participation rate; life expectancy; AI’s place on the Technological Richter Scale.

• Geopolitics and governance of AI, including on autonomous cyber operations, frontier AI chip manufacturing, and governance of AI models.

AI progress is moving quickly, so we believe it is useful to provide an interim update on accuracy, but many questions in our studies have not yet resolved. To allow us to assess the quality of forecasts beyond those that have already resolved, this preliminary accuracy assessment also compares forecasts to:

• Observed current values (prior to question resolution dates): for some questions, particularly benchmarks, we compare forecasts to the latest data that meets our resolution criteria, even if a question has not yet resolved.

• LLM projections: We used GPT-5.5 with “high” reasoning to project the resolution of questions using last known data points, current reporting, and reasonable extrapolations.5 We ran these LLM projections in August 2026.

In this analysis, we mostly compare median forecasts from experts and superforecasters to resolution data, observed values, and LLM projections. (Note that the group of “experts” providing forecasts varies between studies.6) This approach allows us to identify where median forecasts majorly diverge from observed data and possible resolutions, but it does have some major limitations:

1. Focusing on median forecasts may obscure more accurate predictions from subsets of forecasters.

2. Our methodology is biased toward identifying underestimates of AI progress rather than overestimates. As soon as a real-world value crosses the median forecast threshold, that forecast is revealed to be too low. In contrast, a forecast can only be confirmed as too high once its resolution date has passed. Therefore, interim checks on accuracy (of the kind we do in this post) highlight underestimates while overestimates are harder to identify.

3. For most questions, we are not able to draw firm conclusions on accuracy in either direction.

Finally, a note on the “domain experts” discussed below. Our goal in our research was to recruit the kinds of experts that policymakers would typically turn to for advice on these topics. So, we typically include highly cited and well-credentialed domain experts.

Some illustrative examples are:

• Our initial LEAP sample included 339 experts: 76 computer scientists, 76 AI industry experts, 68 economists, and 119 AI policy researchers. Twelve experts were AI TIME 100 honorees. Of the 76 computer scientists, 30 were professors at top-20 institutions (according to CSRankings.org) and 10 were among the 200 top-cited authors in AI (according to OpenAlex). Of the 76 AI industry employees, 26% worked for one of five leading AI companies (OpenAI, Anthropic, Google DeepMind, Meta, and Nvidia), and the median respondent had 9,100 citations (for the 59% of panelists for whom data is available).

• In our study on the economic effects of AI, we recruited 69 economists. Of these, 58% had at least 1,000 citations and 37% had at least one “Top Five” journal publication.

• In our study on LLM-enabled biorisk, we surveyed 46 subject-matter experts in biology and biosecurity. Of the experts, 27 (59%) reported expertise in both biosecurity and wet lab biology research. The expert group’s median number of years of experience was seven years for biosecurity work and eight years for wet lab research. Most experts had a doctorate (78%).

Capabilities

Across all of our AI progress work, forecasts on benchmark performance give us the clearest picture of forecaster accuracy so far. This is largely because more benchmark forecasts have resolved than any other type of forecast.

A clear trend is apparent from these resolved forecasts: the median expert and superforecaster tend to dramatically underestimate AI progress on benchmarks. This pattern is clear in earlier forecasts (from 2022) and those elicited as recently as spring 2026.

Below, we walk through several core examples that illustrate this pattern.

Existential Risk Persuasion Tournament (XPT)

In our mid-2022 Existential Risk Persuasion Tournament (XPT) we collected forecasts about AI progress on the MATH, MMLU, QuALITY, and IMO Gold Medal benchmarks. The chart below shows the implied probability the median domain expert and superforecaster assigned to the observed benchmark results. Overall, superforecasters assigned an average probability of just 9.7% to the observed outcomes across these four AI benchmarks, compared to 24.6% from domain experts. Forecasters were particularly surprised by AI achieving International Mathematical Olympiad gold-level performance in July 2025—this happened five years earlier than the median expert prediction (2030) and 10 years earlier than the median superforecaster prediction (2035).7

There is an obvious hypothesis for why XPT forecasters underestimated benchmark progress: they gave their forecasts between June and October 2022, before the public release of ChatGPT. But if we look to more recent forecasts, this pattern of underestimating benchmark progress continues.8

AI-bio and AI-cyber benchmarks

In our study on LLM-enabled biorisk, which collected forecasts between November 2024 and February 2025, superforecasters and subject-matter experts in biology and biosecurity substantially underestimated AI progress on biorisk-related benchmarks. For example, the median expert predicted that AI models would match a top virologist team on a troubleshooting benchmark (”VCT”) in 2030 while the median superforecaster thought it would take until 2034.9 In fact, AI models likely matched this baseline as of April 2025, several months after the survey was conducted. We also saw similar underestimation of AI progress on a cybersecurity benchmark (Cybench) in our study on LLM-enabled cybersecurity risk.

Several related benchmark questions from these studies remain unresolved, such as, AI reaching a 90% success rate at helping non-experts to acquire synthetic DNA, AI enabling plausible bioweapon attack plans, and AI enabling novice hackers to match an elite team.10 To see all relevant forecasts from these studies, see the appendix table of all relevant AI progress questions.

Longitudinal Expert AI Panel (LEAP)

Since mid-2025 we’ve collected AI progress forecasts from a broad sample of experts (computer scientists, economists, AI industry experts, and AI policy experts) and superforecasters as part of the Longitudinal Expert AI Panel (LEAP). Both groups have substantially underestimated progress on math and coding benchmarks.

We asked forecasters to predict the highest performance on the benchmarks FrontierMath and LiveCodeBench Pro (Hard).11 The FrontierMath question resolved at the end of 2025, with the median forecaster in both groups underestimating progress. The LiveCodeBench Pro question resolves at the end of 2026, but the current state-of-the-art is so far ahead of forecaster medians that it is already clear that they have massively underestimated AI performance on this benchmark.12

The median expert predicted that the leading AI model on FrontierMath Tiers 1–3 by the end of 2025 would achieve 31%, while the median superforecaster predicted 30%. In reality, this question resolved at 40.7%.13 On LiveCodeBench Pro (Hard), the median expert forecast state-of-the-art performance of 14% by the end of 2026, while superforecasters predicted 12%. In May 2026 the actual state-of-the-art on this benchmark was 53.8%.

Forecasters also significantly underestimated performance on a question we asked in the fall of 2025 about the performance of closed-weight models across three benchmarks (FrontierMath Tiers 1–3 and 4, SWE-Bench Verified, and ARC-AGI-2). Forecasts for the open-weight portion of this question were much more accurate than those for the closed-weight portion, with the median of both groups predicting 20% mean benchmark performance against a resolution of 22.7%.14

In April–May 2026, we asked LEAP panelists to forecast the longest METR 80% time horizon listed for an AI model by the end of 2026. Experts and superforecasters gave forecasts of 3.4 and 3.5 hours, respectively. On May 8, toward the end of the survey period, METR updated the benchmark to show that Anthropic’s Mythos Preview model had achieved an estimated 80% time horizon of 3 hours and 6 minutes (3.1 hours), already within the range of median panelists’ forecasts for the end of 2026. If the historical trend of progress on this benchmark continues, forecasters’ median predictions will have substantially underestimated progress by the end of 2026.

While we were drafting this post, the news broke that OpenAI had likely solved one of the Millennium Prize Problems—the Navier-Stokes Problem–using an internal model significantly more capable than GPT-6 Astra. LEAP experts, originally forecasting in August to September 2025, gave just a 10% chance of AI solving a Millennium Prize Problem by the end of 2027, while superforecasters gave even lower odds of 5.4%. Although it is not completely clear whether the current situation meets our resolution criteria, it appears likely that these forecasts are another significant example of forecasters underestimating AI progress.

Exceptions: more accurate forecasts

Accuracy on some unresolved benchmark questions looks more promising. For example, according to our LLM projections, near-term forecasts on Chinese model performance on the Epoch Capabilities Index (ECI) may be fairly accurate, although later resolution dates diverge more strongly from LLM forecasts.15

Overestimation of a non-benchmark capability: AI-biorisk uplift study

We have fewer examples of non-benchmark capability forecasts, but one notable overestimation of AI capabilities came when forecasters predicted how much mid-2025 LLMs could help non-experts with biorisk-relevant laboratory tasks.

In mid-2025, we asked for predictions about what fraction of STEM undergraduates would be able to complete three biorisk-relevant tasks with the help of an LLM in a randomized controlled trial run by Active Site. Biosecurity experts (at median) predicted that 22.5% of the sample would complete the tasks, virologists predicted 40%, and superforecasters predicted 16.2%. In actuality, 5.2% of the sample completed all of the tasks.

We discuss the potential implications of these findings about AI capabilities forecasting accuracy in the discussion section below.

AI adoption and diffusion

Adoption and diffusion forecasts attempt to track the uptake of AI through a range of proxies, including economic indicators, infrastructure rollout, and usage indicators. We have fewer resolved questions in this category and less observed data, so we are more reliant here on comparisons to LLM projections. As such, preliminary conclusions in this section are more speculative than our findings on benchmarks.

The forecasts below primarily come from LEAP, where we have collected the most adoption-related forecasts. In some highlighted cases we also have relevant data from the XPT.

Overall, the picture for forecasting accuracy on AI adoption and diffusion indicators is more mixed, though forecasters often lean toward underestimation:

• On economic proxies for adoption, such as revenue growth of AI companies, median experts and superforecasters majorly underestimated progress.

• On broad adoption indicators, such as what percentage of work hours in the US are assisted by AI, the evidence of forecaster accuracy is uncertain.

• Forecasts for resources relevant to AI growth that are proxies for adoption and diffusion (electricity used by AI and data center growth) were much closer to LLM projections than other adoption indicators. However, in our XPT study, forecasters substantially underestimated how quickly spending on the largest models would grow.

• It appears that experts overestimated the rate of autonomous vehicle adoption: when surveyed in June–August 2025, the median expert predicted that 7.3% of US ride-hailing trips would be completed by autonomous vehicles by the end of 2027. Based on current trends, LLMs predict that the actual resolution value will be 2.5%. Superforecasters appear more accurate, with a median forecast of 2%.

• We have collected forecasts for many adoption and diffusion indicators that we cannot yet judge the accuracy of, such as what fraction of people will use AI for companionship over time and what share of drug sales will come from AI-discovered drugs.

Economic proxies for adoption: major underestimation of AI revenue growth

Broadly, forecasts of economic proxies for adoption seem to have significantly underestimated AI uptake.

In our study “Forecasting the Economic Effects of AI,” we asked participants for the highest annual recurring revenue (ARR) for an AI-focused company at the end of 2026. AI experts gave a median ARR of $20 billion, economists predicted $16 billion, and superforecasters predicted $25 billion.16 It is likely that the current value for this question (as of September 22, 2026) is $100 billion, indicating that forecasters are on track to significantly underestimate adoption.

We posed a similar question to our LEAP panel in summer 2026, asking them to predict the combined annualized revenue run rate of Anthropic and OpenAI at the end of 2026. Experts and superforecasters gave median forecasts of $70 billion and $90 billion, both of which are substantially below the likely current value for this question, $140 billion.17 Our LLM projection (as of August 18, 2026) estimates that this question will resolve at $165 billion, although this does not take into account the latest reporting on Anthropic’s revenue.

Although revenue projections seem on track to be very inaccurate, the picture for other AI-relevant economic indicators is more mixed. In the XPT, the median forecaster accurately predicted the lowest price of 1 GFLOPS by the end of 2024, and was reasonably close to predicting US computer R&D spending at the same resolution date.18

Broad adoption indicators: uncertain

In September–October 2025 we asked LEAP panelists to forecast the share of US work hours that would be assisted by generative AI in 2025. Experts and superforecasters gave median predictions of 4% and 3.6%, respectively, and the question resolved at 5.7%. However, in the question information we gave to forecasters we cited a historical baseline of 2% from an early version of a study from the Federal Reserve Bank of St. Louis. An update to this paper revised this value to 3.35%, so while the median forecasters appear to have underestimated here, it may be that they were mistakenly anchoring on out-of-date baseline information.

Resources relevant to AI growth: a mixed track record

A question we asked in the XPT in 2022 about the cost of the largest AI experiment run by the end of 2030 is also very likely to result in a dramatic underestimate from forecasters: expert and superforecaster median forecasts of $180 million and $100 million, respectively, have already been overtaken by Grok 4, which cost an estimated $388 million to train (published in mid-2025). Forecasts for the cost of the largest AI experiment before the end of 2024, however, were more mixed, with the experts predicting $65 million and superforecasters $35 million against a resolution of $45.8 million.

Forecasts about the resources related to AI growth appear to be on track to be much more accurate. In LEAP we’ve asked questions about the share of US electricity going toward AI, and the buildout of hyperscale data centers globally at various resolution dates. For the closer resolution dates, the majority of forecaster medians appear to be reasonably close to LLM projections (see table below).19

Autonomous vehicles: experts overestimate adoption

We asked LEAP forecasters to predict the share of US ride-hailing trips that would be delivered by autonomous vehicles during 2027. The expert median for this question was 7.3%, which is on track to be a large overestimate compared to our LLM projection of 2.5%. Superforecasters appear to be notably more accurate on this question, giving a median prediction of 2%.

Future indicators

Many adoption forecasts from our studies have not yet resolved—for example, forecasts on drug discovery and diffusion across science. We’re also tracking forecasts on the share of US adults who self-report using AI for companionship daily by the end of 2027. The median expert and superforecaster both gave a forecast of 10%, which is above the LLM projection of 8% on this question.

Impacts

Our studies also capture a wide range of forecasts on AI’s impacts. These include impacts on employment and economic growth, but also AI’s effect on cybercrime, global catastrophic risk, and improving life expectancy. Not enough forecasts of this kind have resolved to draw any meaningful conclusions about accuracy, but we cover some preliminary data below.

Economic and labor impact

In the XPT we asked forecasters to predict the share of US GDP that would come from software and information services in 2024. They slightly overestimated this share, but their median forecasts were close to the resolution figure of 3.25%.20 In the same study, forecasters slightly underestimated the OECD labor force participation rate for 2024, but were again pretty close to the true resolution value.21

In LEAP, we’ve asked for forecasts on a range of economic and labor outcomes, including those in the table below. It is too early to assess the accuracy of these forecasts.

We also asked for many more macroeconomic forecasts in our study on the economic effects of AI.

Other impact measures

In the XPT, we asked forecasters to predict the share of Americans who would say they had a negative opinion of AI by the end of 2024. Here, the median expert and superforecaster predictions of 33% were extremely close to the resolution of 32.9%.

In our study on AI cyber risk we asked for the year in which an AI-enabled attacker causes a grid blackout causing $100 million or more in damages. Experts and superforecasters gave median predictions of 2031 and 2035, respectively, though this question is yet to resolve. In LEAP we asked for predictions of the annual losses from cybercrime reported to the FBI in 2026, 2028, and 2031, so we should have some resolution data on this question relatively soon.

In many of our studies we’ve elicited forecasts on catastrophic risk, AI-enabled pandemic risk, large-scale harm events, and other forms of major risk. We have also elicited forecasts on major potential benefits from AI such as improvements in life expectancy. These are some of the most important forecasts across our work, but it’s too early to say if high or low risk and benefit estimates are more accurate. For more information, see our LEAP waves on risks from AI and benefits from AI.

Geopolitics and governance

Although we only have a small number of questions in this category, it’s worth highlighting one recent question that appears to have surprised forecasters.

In LEAP Wave 9 (May–June 2026), we asked forecasters to predict the calendar year by which either the US, UK, or EU governments would issue a binding directive, regulation, court order, or injunction to delay, seriously restrict, or condition the public release of an AI system, based on safety risks. The median expert and superforecaster gave a 50% chance this would happen by 2030.

In fact, this question may be very close to resolution already. In late June, OpenAI restricted availability of its unreleased model GPT-5.6 Sol to a small group of partners approved by the US government. It is not clear that this meets our resolution criteria, as this does not appear to be the result of a binding directive, but this and the Claude Fable post-release restrictions suggest that resolution on this question is potentially much closer than forecasters anticipated.

Other unresolved geopolitics and governance questions include:

• The share of frontier AI chips produced by the US, China, Taiwan, and other regions;

• The year in which NATO or a member state will publicly authorize the use of fully autonomous offensive cyber operations or cyber weapons;

• Whether a top-four frontier lab will exist within a non-democracy by 2030.

Limitations

This post is intended as a brief overview of where median AI progress forecasts appear to be inaccurate across our studies, but there are several major limitations to our approach and caveats worth bearing in mind, including:

• Our methodology is biased toward identifying underestimates of AI progress as it is easier to identify forecasts where the current value has already overshot the median forecasters’ predictions. If an AI capability predicted to occur in 2030 is actually achieved in 2035, we will need to wait many years to report on this overestimation.22 The vast majority of our AI progress forecasts are unresolved and it may be the case that over time the picture of forecaster accuracy changes significantly.

• In this post we have concentrated almost exclusively on median forecasts, which may obscure more accurate predictions from subsets of forecasters, as well as information from forecasts covering other percentiles.

• In some cases, issues with our question specification or our elicitation format may have led to inaccurate forecasts. For example, for LiveCodeBench Pro (Hard) forecasts, it may be that some forecasters did not understand that the benchmark allows models to go back and test on questions from earlier quarters.23

Discussion

Overall, AI capabilities and some important measures of adoption (such as AI company revenue) are moving much faster than experts and superforecasters expected.

One potential implication of this is that forecasters should expect much faster AI progress, faster AI adoption, and much larger AI impacts going forward.

We have seen some evidence that forecasters in LEAP are updating in this direction. In summer 2025 and spring 2026, we asked panelists to forecast the likelihood of AI in 2040 reaching various levels of impact on the Technological Richter Scale (TRS, a scale devised by Nate Silver that ranks technologies by their broad societal impact). Over the course of nine months, experts and superforecasters have increased the likelihood they assign to AI’s impact being comparable to “technology of the millennium” (e.g., agriculture, Level 9), while continuing to assign the greatest probability to AI being comparable to “technology of the century” (e.g., electricity, Level 8):24

If forecasters are updating their views toward faster AI progress, we might expect them to perform more accurately on benchmark questions asked more recently. However, we have not yet seen evidence of improved accuracy on more recent questions. Forecasters likely majorly underestimated progress on the METR time horizon benchmark (asked in April 2026) and underestimated OpenAI and Anthropic revenue growth in our July 2026 survey. We will try to understand what is driving forecasters’ continued accuracy errors and will take various steps to highlight more accurate and relevant forecasts, which we’ll expand on in the next post in this series.

An alternative response to the evidence in this post is to say that benchmarks have moved much faster than forecasters expected in part because it was easier to “teach to the test” than forecasters expected and therefore that benchmarks are less meaningful for making sense of real-world AI capabilities than forecasters might have expected. Perhaps significant benchmark progress just doesn’t lead to real-world impacts in many areas. In which case, we might expect forecasters to underestimate AI progress on benchmarks, but turn out to be more accurate when it comes to forecasting real-world impacts such as macroeconomic impacts.

This debate highlights the importance of getting accuracy data on additional adoption and impact measures, which we expect to have soon as more questions resolve.

Next steps

In this post, we have covered the accuracy of median forecasts across our studies, often highlighting places where these median forecasts have been inaccurate. However, subsets of forecasters in our studies have been much more accurate than the median forecaster.

In the next post in this series, we’ll look at a few ways we can improve the accuracy and relevance of LEAP forecasts and highlight predictions from more accurate forecasters. These involve three core updates to LEAP:

• Highlighting forecasts from a subsample of panelists who expect very fast AI progress by 2040.

• Adding continuously updated LLM forecasts to LEAP findings.

• Identifying our most accurate LEAP participants and highlighting their forecasts as soon as we have sufficient data to make this judgment.

The Automated AI Risk Outlook (AIRO) is a new dashboard where an ensemble of frontier LLMs forecasts the probability of catastrophic risk events. To learn more, view the AIRO dashboard or read our launch white paper.

Forecasts about the likelihood of catastrophic risks from AI vary wildly. Dario Amodei has estimated a 25% chance of human extinction within the next few decades. Experts, meanwhile, put the chance of an AI-related catastrophe at 0.3% by 2030 and 2% by 2050. Understanding the level of risk and taking appropriate action is one of the most consequential challenges facing humans today.

Forecasts from frontier AI models could be an important input into this debate. The best models now approach superforecaster levels of accuracy, and models are rapidly becoming more capable. To take advantage of these capabilities, we’re introducing the Automated AI Risk Outlook (AIRO)—a fully automated dashboard of catastrophic risk outcomes as forecast by frontier AI models. AIRO enables us to track in close to real time how the likelihood of catastrophic outcomes changes, and how these risk forecasts relate to advances in model capability. The accompanying white paper also presents preliminary forecasts conditional on policy scenarios.

The AIRO dashboard currently puts the likelihood of an AI-related catastrophe that kills at least 10% of the population at 0.47% by the end of 2030 and 6% by the end of 2050. In addition to this headline forecast, AIRO tracks risks from AI-specific catastrophes and incidents, biorisk, cyberrisk, and misalignment risk.

What is AIRO?

AIRO is a regularly updated dashboard of catastrophic risk forecasts from an ensemble of the four highest-scoring eligible models on the Epoch Capabilities Index (ECI), keeping one model per family. At the time of launch, these were GPT-6 Astra, Fable 5.1, Opus 5, and GPT-5.5 Pro. We report separate forecasts from these frontier models as well as an ensemble forecast calculated as their unweighted median.

We define a catastrophic outcome as an event that kills at least 10% of the global population, and ask for the likelihood of such an event happening across various time horizons, from within six months to 2100. As well as a general catastrophe, we ask LLMs to forecast the likelihood of an AI-caused catastrophe and the probability of AI-enabled human disempowerment. AIRO also forecasts cumulative harm from AI-related incidents, including human-caused epidemics, cyber incidents, and misalignment incidents. Each severity threshold can be reached through deaths, or through equivalent morbidity or economic damages—for example, 100,000 deaths or $220 billion in damages. These are not forecasts of deaths alone.

We also ask LLMs to forecast these outcomes conditional on the score of the top AI model on the ECI, giving us a sense of how AI capabilities are linked to the probability of catastrophic outcomes. Models forecast the frontier score on the ECI six months from now, then forecast risk outcomes conditional on higher- or lower-than-median scores being achieved.

Automating risk forecasts has some big advantages. LLMs can forecast many outcomes across many time horizons and severity thresholds without growing tired, allowing us to elicit many more forecasts than is possible with human forecasters. AIRO also provides us with risk forecasts that update as real-world events change and model capabilities advance, giving us forecasts based on the very latest information.

AIRO risk forecasts

The following forecasts are the unweighted medians of the four included models’ forecasts, elicited on September 10, 2026. This post only includes a subset of forecasts; for the full forecasts visit the AIRO dashboard.

AI catastrophe

Probability of a global catastrophe in which AI was the proximate cause:

• Within 12 months: 0.01%

• By 2030: 0.47%

• By 2050: 6%

• By 2100: 12.3%

Across the four models at the 2050 horizon, forecasts range from 1.5% (Opus 5) to 7.5% (GPT-6 Astra). Forecasters from our Longitudinal Expert AI Panel (LEAP) who were asked a similar question in May–June 2026 gave lower probabilities. Experts gave a 2% probability of an AI-caused global catastrophe by 2050, and a 5% probability by 2100.

General catastrophe

Probability of a global catastrophe that causes the deaths of at least 10% of the population:

• Within 12 months: 0.04%

• By 2030: 1.1%

• By 2050: 8.5%

• By 2100: 19%

We asked our LEAP respondents a version of this question in May–June 2026. Experts gave a 4.2% probability of a global catastrophe by 2050, and a 10% probability by 2100.

Catastrophic risk conditional on AI capabilities

AIRO also tracks how LLM risk forecasts change conditional on AI capabilities. We ask models to forecast the frontier score on the ECI six months from now and then forecast risk outcomes conditional on where the state of the art falls relative to that median forecast.

As of September 10, the median of the models’ forecasts for the frontier ECI score on March 10, 2027, was 176. When each model instead assumed its own 75th-percentile ECI outcome, its forecast of an AI-related catastrophe by 2030 rose by an ensemble factor of 1.42 relative to its unconditional forecast.

Why trust LLM risk forecasts?

AI forecasting abilities have improved alongside other AI capabilities. In July 2026, an AI model reached parity with superforecasters on FRI’s forecasting benchmark, ForecastBench. This gives some basis for believing that AI forecasting may now be state of the art.1

The forecasts shown on AIRO are different from those evaluated in ForecastBench in two important ways. First, they concern very rare events. It is difficult to measure accuracy on forecasting rare events given that they, by definition, happen so rarely. However, unlike human forecasters, we can ask AI models to make thousands of predictions on events in a simulated world environment. When we do this, we find that more capable models—as assessed by their score on the ECI—are more accurate at forecasting rare events.

Second, AIRO includes many conditional forecasting questions. The dashboard includes forecasts conditional on model capabilities; the white paper also explores preliminary policy scenarios. We can again test performance on this type of conditional forecasting in a simulated world environment, and again we find that accuracy correlates with general capabilities.

There are risks associated with relying on automated AI forecasters of catastrophic risk. For example, AI forecasts may be influenced by strategic behavior: a highly capable but misaligned model trying to evade human suspicion may lower its estimates to avoid causing alarm. This “sandbagging” behavior may also arise if companies, concerned about regulation or other threats to profit, intentionally post-train models to underplay the risk of AI-caused harms.

How AIRO works

Our forecasting questions are drawn from earlier FRI work, allowing us to compare forecasts from the LLM ensemble with those from experts and superforecasters. We elicit forecasts from the four highest-scoring eligible models on the ECI, keeping one model per family.

Each model answers all questions—across time horizons and severity thresholds—in one forecasting session. Each model conducts its own research via an agentic harness. A prompt asks it to investigate recent developments, expert reports, and published estimates for each cause of catastrophe, and use what it learns to formulate follow-up queries. The model chooses searches, receives results, and decides what to investigate next. All models receive the forecasting date and questions and use the web-search and page-reading tools supplied through Tavily.

Models are allowed to submit their forecasts after at least 10 research-tool calls have returned. The model is instructed to stop researching when further searches would no longer change its estimate. At the turn limit, a safeguard requests final submission, with a bounded continuation for endpoints that cannot force a tool call. At the end of the session, the model submits its forecasts together with a short rationale and the key sources it relied on.

What’s next for AIRO

AIRO is still a work in progress. We plan to keep validating and extending our dashboard by refining ensemble forecasts and developing a larger set of intermediate questions related to catastrophic risk outcomes. These intermediate forecasts will serve as validation tools and provide inputs that future models can use to inform their risk assessments. We’ll also continue to test conditional forecasts to better understand the costs and benefits of proposed policies to mitigate AI risk.

If you have questions or feedback about AIRO, please email airo@forecastingresearch.org.

That is a big if. Other companies have had periods of very rapid growth only to tail off before crossing $100 billion in revenue. The AI industry also faces headwinds that may slow down infrastructure buildout. In July, New York State passed the first statewide moratorium on hyperscale data centers, while slow electricity interconnections and supply constraints continue to threaten buildout within the US.

Is all of this heading toward a boom or a bust? That is the theme of our latest set of forecasts from the Longitudinal Expert AI Panel (LEAP). We asked our panel of AI experts and superforecasters to forecast the following:

• The value of semiconductor and software stock indices

• Data center investment in the US

• Investment in IT equipment, and electrical and communication structures

• The combined annualized revenue run rate of OpenAI and Anthropic

Forecasters expect data center investment to continue growing quickly, despite the prospect of moratoria. They expect more moderate growth in investment in IT equipment and electrical infrastructure, both of which are related to AI buildout but are much broader. The LEAP panel predicts gains for a bundle of stocks covering software and another linked to semiconductors, though it predicts the latter to grow at a somewhat slower pace than the rapid rise of recent years. On frontier-AI revenue, meanwhile, superforecasters predict OpenAI and Anthropic to continue their impressive rise (assuming they exist as independent companies), while we think our elicitation method may have tripped up many experts and members of the public.

Read on for more on these insights. Visit our website for more on Wave 11, including question details and analysis of rationales.

Insight 1: Experts expect private investment in data center structures to grow by 76% in real terms from 2025 to 2028

Despite public opposition to data centers, and the recent statewide moratorium on hyperscale data center construction in New York State, experts expect private investment in data center structures to grow by 76% in real terms by the end of 2028 compared with 2025 levels.

In 2025, annual US private fixed investment in data center structures—the physical shells that house data center computing hardware—was $40.8 billion. Experts expect investment in data center structures to rise by 29% by the end of 2026, 76% by the end of 2028, and 120% by the end of 2035, compared with the 2025 level. These translate to total private investment of $52.6 billion in 2026, $71.8 billion in 2028, and $89.8 billion in 2035, all in constant 2025 dollars.

Superforecasters predict a similar trajectory. They expect private fixed investment in data center structures to grow by 30% by the end of 2026 and 70% by the end of 2028, compared with the 2025 level. By 2035 superforecasters expect investment to be 250% of its 2025 level. This translates into expected investment in 2035 of $102 billion for superforecasters versus the $89.8 billion predicted by experts. The public expects considerably lower levels of investment, forecasting a 48% increase by the end of 2028 and an 84% increase by the end of 2035.

In their written rationales, high-investment forecasters argued that announced data center projects are likely to translate into large increases in investment by 2028. Low-investment respondents cited public opposition and state-level moratoria on data center construction, such as the recent moratorium imposed in New York State, as factors that may slow down data center buildout and thus investment.

Insight 2: Experts predict significant increases in investment in IT equipment, and electrical and communication structures, although slower growth rates than data centers

We also asked forecasters to predict total annual US private fixed investment in electrical and communication structures, and IT equipment. These cover the increased electricity generation and transmission needed to power data centers and the servers, networking gear, and other electrical infrastructure that keep them running. Both of these categories are much larger than data center structures: private investment in IT equipment in 2025 was $514.8 billion while for electrical and communication structures it was $146.4 billion.

Experts forecast considerable growth in investment in IT equipment compared with a 2025 baseline of $514.8 billion. They expect this to grow 20% by 2026, 48% by 2028, and 90% by 2035—implying real-terms investment of $978 billion by 2035. Superforecasters have similar forecasts, edging above $1 trillion by 2035.

In contrast, both experts and superforecasters expect slower growth in total annual private investment in electrical and communication structures. From a baseline of $146.4 billion in 2025, experts forecast 5% growth in investment in 2026, 15% growth by 2028, and 32% growth by 2035. Superforecasters forecast similar levels of growth.

In their written rationales, forecasters noted that the electrical and communication structures category is economy-wide and thus less sensitive to a data center boom, and that investment in this sector has been relatively flat historically. By contrast, forecasters expected the short life of hardware to drive investment in IT equipment, pointing out that data center structures last for decades while servers become outdated after a number of years.

Insight 3: Experts forecast the value of semiconductor equities to grow more slowly through 2028 compared with the last 2.5 years

We asked forecasters to predict the change in two exchange-traded funds (ETFs) that track the share prices of semiconductor stocks and software companies. The VanEck Semiconductor ETF (SMH) tracks the 25 largest US-listed semiconductor companies, a proxy for the AI compute buildout. The top five SMH holdings by weight as of July 8, 2026, were Micron Technology, AMD, TSMC, ASML Holding, and Broadcom.

Over the past 2.5 years, SMH has more than tripled in value (rising from $180.31 on January 18, 2024, to $607.73 on July 9, 2026). A simple extrapolation of historical trends would suggest that by the end of 2028, SMH would have a value of roughly $2,048 (3.37 times its level on July 9, 2026).

Instead, the median expert forecasts SMH to reach only 1.5 times its mid-2026 value. Superforecasters had similar expectations, predicting SMH would reach 1.36 times its mid-2026 value by the end of 2028. On average, experts and superforecasters forecast less than a 10% chance (8% and 7% respectively) that SMH will reach the trend-extrapolated value of 337 by the end of 2028. In fact, they predict that it’s more likely (experts: 17%, superforecasters: 27%) that SMH will decrease in value than that it will remain at or above its current trend.

Respondents who predicted high SMH growth argued that near-term data center buildout (and thus demand for chips) was showing no signs of slowing down, with much future demand already locked in. Forecasters who predicted slower growth thought that future returns may not match the recent increase in stock prices and that large-scale data center expansion could slow down by the end of 2028.

Insight 4: Experts expect the value of North American software equities to grow at a slightly faster rate through 2028 compared with the last 2.5 years

We also asked forecasters to predict the change in the iShares Expanded Tech-Software Sector ETF (IGV). This fund tracks roughly 110 North American software companies. The top five holdings by weight as of July 8, 2026, were Palo Alto Networks, Palantir Technologies, Microsoft, CrowdStrike, and Oracle.

Over the past 2.5 years, IGV has increased roughly 1.1x in value (the price of a share rising from $82.19 on January 18, 2024, to $93.88 on July 9, 2026). A simple extrapolation of historical trends would suggest that by the end of 2028 IGV would have reached a value of $107.23 per share, 1.14 times its mid-2026 value.

Instead, the median expert forecasts that IGV value will grow faster than its recent historical trend, reaching 1.2-1.3x its mid-2026 value by the end of 2028. Superforecasters predict IGV will reach 1.2 times its mid-2026 value over the same period. In essence, forecasters expect lower absolute growth rates for software stocks compared with semiconductor stocks, but also expect software stocks to show a relative increase over their historical growth trend while growth in semiconductor stocks slows.

The average expert and superforecaster give a 63% and 58% chance, respectively, that IGV will surpass its trend-extrapolated value by the end of 2028. They also give a 25% and 29% chance, respectively, that the index declines from current values.

In their written rationales, forecasters considered that the key dynamic influencing the value of IGV was whether AI is used to sell software or to replace it. Low-valuation respondents tended to argue that AI is at risk of rendering many software companies obsolete while high-valuation respondents noted the ability of software companies to turn technology threats into profitable opportunities. Forecasters also pointed to IGV’s recent underperformance as a reason to expect better forward returns.

Insight 5: The average superforecaster predicts a 34% chance that the combined annualized revenue run rate of OpenAI and Anthropic exceeds $400 billion by the end of 2030 while forecasting an 18% chance that it’s less than $100 billion

OpenAI and Anthropic are reporting staggering revenue run rates. In February 2026, Bloomberg reported that OpenAI expects revenue to exceed $280 billion by 2030, and in August 2026, Reuters reported that Anthropic is forecasting revenue in 2028 to reach $190 billion. Reaching these figures would immediately put either firm in a very small club: only 30 companies on the Fortune Global 500 list reported more than $200 billion in revenue in 2026.

We asked respondents to forecast the combined annualized revenue run rate of OpenAI and Anthropic, assuming at least one of them exists as an independent company. If one or both companies exist, the median superforecaster expects their combined annualized revenue run rate will be $300 billion in 2030, or roughly four times the $72 billion annualized revenue run rate reported in July 2026 at the time of forecasting. The median superforecaster predicts this will rise to $770 billion by 2040. Since the survey closed, the reported combined annualized revenue run rate has surpassed $105 billion.

Using the 10th, 50th, and 90th percentile forecasts provided by each respondent, along with their forecasted probabilities of at least one company existing independently, we construct an average distribution across all superforecasters. On average, superforecasters predicted a 34% chance that the combined OpenAI and Anthropic annualized revenue run rate in 2030 will exceed $400 billion while forecasting an 18% chance it’s less than $100 billion.

We’re focusing on superforecaster predictions here, as we suspect a number of our expert and public respondents may have erred in forecasting this question such that their responses may not reflect their best judgment. Median forecasts for both groups were below the latest published value available at the time of forecasting. While one might hold the view that either or both companies’ revenue will collapse before year’s end, rationales showed no evidence that a large number of these forecasters intended to predict a reversal of recent trends. Instead, we suspect our elicitation design may have made it easy for forecasters to ignore this most recent data point and anchor on earlier data. While superforecasters used the same interface, their predictions and rationales provide less indication that they missed the latest published value.

We are currently running an accuracy analysis of LEAP forecasts, which will include exploring the effects of elicitation designs and attentiveness on forecast accuracy.

More from LEAP

This post covers key highlights from the Wave 11 LEAP survey conducted between July 13 and August 11, 2026. Recent waves covered robotics, the economic effects of AI, AI risks and AI benefits.

Summary

We ran a forecasting study to better understand expert views on the effectiveness of Anthropic’s AI Safety Level 3 (ASL-3) Deployment Standards at reducing biosecurity risk based on AI models from early 2026. We surveyed 22 biosecurity, national security, and/or terrorism experts, and 20 superforecasters between March 4 and March 29, 2026.

We asked respondents to forecast the total expected financial damages from large-scale human-caused outbreaks between April 2026 and the end of 2028 in four scenarios:

• A world without general-purpose AI (GPAI)

• A world where no GPAI models had any safeguards

• A world where all GPAI models were protected by ASL-3

• A world where frontier GPAI models were protected by ASL-3

In all scenarios, we asked respondents to assume that AI model capabilities remain fixed at the level of Claude Opus 4.6. This is because we wanted to understand expert views on how effective ASL-3 is at mitigating risks from the early 2026 generation of publicly available models, rather than how AI capabilities are likely to progress.

Experts and superforecasters generally believe that implementing ASL-3 safeguards across all GPAI models would substantially reduce the risk of large-scale human-caused outbreaks compared to a world with no safeguards. They also estimate that ASL-3 safeguards—if implemented across all GPAI models—would capture most of the risk reduction it is possible to achieve by safeguarding. In an “ASL-3 for frontier models” scenario, forecasters generally think that variations to ASL-3 could increase effectiveness, with the largest change in expected damages associated with a 5x faster speed of patching jailbreaks.

Introduction

In May 2025, Anthropic announced that it had activated its AI Safety Level 3 Deployment and Security Standards as a precautionary measure in response to Claude Opus 4’s chemical-, biological-, radiological-, and nuclear-related knowledge and capabilities. The deployment standards include the implementation of real-time classifier guards, offline monitoring, access controls, a bug bounty program, threat intelligence, and rapid response.

Although Anthropic can evaluate aspects of its mitigations (for example, the time taken to patch an identified jailbreak), it is unclear how effective these mitigations are at achieving the overarching goal of reducing biosecurity risks posed by GPAI models. To investigate this, we conducted a forecasting survey of experts in biosecurity, national security, and/or terrorism studies, as well as superforecasters. Respondents were given information on Anthropic’s ASL-3 standards, including confidential data from red-teaming efforts, the bug bounty program, and other safeguard testing.1

The main outcome we asked participants to forecast was:

The total worldwide damages attributable to human-caused outbreaks that start between April 1, 2026, and December 31, 2028, and that each individually cause at least $100 million in total worldwide damages (in 2026 USD).2

Although we specified that outbreaks must start between April 1, 2026, and December 31, 2028, we did not specify a cutoff date for damages.

Forecasters considered the following scenarios:

• No GPAI: A world where GPAI never existed and will not exist. All other features of the world remain the same, so the absence of GPAI does not imply that technological progress is slower across all domains.

• No safeguards: A world where no safeguards are applied to any GPAI model.

• ASL-3 safeguards for all GPAI models: A world where all GPAI models—regardless of size or capability—are protected by Anthropic’s ASL-3 safeguards.

• ASL-3 safeguards for frontier GPAI models: A world where “frontier” GPAI models are protected by ASL-3 safeguards.3

How effective are ASL-3 safeguards?

When respondents assumed all models had ASL-3 protections, the median forecasted probability of human-caused outbreaks causing at least $100 million in damages dropped by roughly 40% relative to the “no safeguards” scenario (from 16% to 9.2%). For comparison, when respondents were asked to consider a hypothetical world where GPAI was never developed, the median forecast was 9.3%.

The scenario that asked respondents to assume GPAI models did not exist can provide an upper bound on the risk reduction achievable by safeguards against GPAI model misuse, as it demonstrates the risk of human-caused outbreaks that would persist regardless of GPAI models. We can also express these results in terms of the share of GPAI-attributable risk mitigated by ASL-3.4 The median participant’s forecasts imply that ASL-3 for all models would mitigate 71.7% of GPAI-attributable risk.

This comparison has an important limitation: the “no GPAI” scenario is an imperfect proxy for perfect safeguards. Perfect safeguards would eliminate the risk of misuse while preserving the defensive benefits of GPAI, such as contributions to pandemic surveillance and vaccine development. The “no GPAI” scenario removes all of these. This means the “no GPAI” world is likely more vulnerable to outbreaks (including human-caused ones) than a world with perfectly safeguarded GPAI, making it a lenient benchmark.

The practical consequence is that the 71.7% efficacy figure—the share of maximum risk reduction achieved by ASL-3 for all models—is likely biased upward. If GPAI’s defensive contributions are small relative to the misuse risk, the bias is minor. If they are large, the bias could be substantial.

The impact of protecting more models with ASL-3

To understand the potential impact of all frontier models being protected by ASL-3 safeguards, we asked respondents to forecast the main outcome under this scenario.

Under the “ASL-3 for frontier models” scenario, the median probability of human-caused outbreaks occurring between April 2026 and the end of 2028 and leading to at least $100 million in damages (see Figure 1 above) was 10.6%. Compared to the “no safeguards” scenario, the “ASL-3 for frontier models” scenario was associated with an 18% reduction in this risk.

We also assessed how the calculated expected damages under the two ASL-3 scenarios and the “no GPAI” scenario compared to the “no safeguards” scenario.

The median respondent thought that “ASL-3 for frontier models only” would be associated with a 20% reduction in expected damages due to large-scale human-caused outbreaks relative to the “no safeguards” scenario. Most respondents placed the “ASL-3 for frontier models” scenario’s risk between “ASL-3 for all models” and “no safeguards;” but disagreed on where, with some treating it as nearly equivalent to “ASL-3 for all models” and others as only marginally better than “no safeguards.”

The impact of variations in ASL-3

Respondents were asked to forecast the main outcome conditional on frontier models being protected by the following hypothetical variations on the ASL-3 scenario:

Time required to identify jailbreaks

• A1: The average time required to find a universal jailbreak is 5x longer.

• A2: The average time required to find a universal jailbreak is 5x shorter.

Time to jailbreak patching

• B1: It takes a fifth (0.2x) of the time to patch all universal jailbreaks.

• B2: It takes up to 2x as long to patch all universal jailbreaks.

Capabilities loss associated with jailbreaks

• C1: All universal jailbreaks significantly reduce the model’s GPQA score. Compared to jailbroken models’ current performance relative to the “no jailbreak” baseline, jailbroken models’ relative performance decreases by roughly 30%.

• C2: All universal jailbreaks only minimally reduce the model’s GPQA score. Compared to jailbroken models’ current performance relative to the “no jailbreak” baseline, jailbroken models’ relative performance increases by roughly 30%.

Figure 4 below shows the median change in expected damages relative to “ASL-3 for frontier models” under the variant scenarios. The largest median change—an 8% decrease relative to the “ASL-3 for frontier models” scenario—was associated with a faster time to jailbreak patching.

The impact of ASL-3 safeguards on biosecurity threat pathways

We asked which type of actor is most likely to be the primary cause of a human-caused outbreak that causes at least $100 million in damages, and how this would change depending on whether GPAI models had no safeguards or were all protected by ASL-3 safeguards. Both groups of respondents thought that state actors were the most likely cause of such an event and would be an even more likely cause under the ASL-3 safeguards scenario. Among the expert respondents, the next most likely group was individual expert actors, although this group’s probability of being the primary cause of a large-scale human-caused outbreak fell under the ASL-3 safeguards scenario.

When asked about the probability of unauthorized frontier model use contributing to a human-caused outbreak that causes more than $100 million in damages, the median expert forecast a 30% probability, and the median superforecaster a 23.5% probability. We then asked respondents to assume that such a scenario had occurred, and to say how likely they thought the following pathways were to have contributed to that unauthorized model use:

• Own jailbreak

• Public jailbreaks

• Nonpublic jailbreak (e.g., black market)

• Trusted user exemption

Each of the pathways we asked about had a median probability of 15% or higher, suggesting that most respondents found all pathways plausible. Relative to experts, superforecasters generally thought that a trusted user exemption was more likely to be used, and that the actor generating their own jailbreak was less likely.

Conclusion

This study asked 22 domain experts and 20 superforecasters to estimate how ASL-3 safeguards would change the expected financial damages from large-scale human-caused outbreaks between April 2026 and 2028. Four findings stand out.

1. Respondents believe ASL-3 safeguards demonstrate high efficacy. If applied to all GPAI models, the median respondent’s forecasts imply that ASL-3 would mitigate roughly 70% of GPAI-attributable risk.

2. Protecting only frontier models with ASL-3 reduces risk, but not as much as protecting all GPAI models. Protecting only frontier models captured less of the risk reduction than protecting all GPAI models, though respondents disagreed substantially on the size of the gap. This disagreement generally hinged on how capable and accessible near-frontier models—particularly Chinese open-weight models—are judged to be.

3. Making jailbreaks harder to find, patching them faster, or further degrading the performance of jailbroken models were generally expected to improve ASL-3 performance. The largest effect was associated with faster patching, consistent with the logic that bioweapon development requires sustained iterative access. The variants were tested only under the frontier-only scenario, where unsafeguarded near-frontier models may dampen the marginal impact of any frontier-specific improvement.

4. ASL-3 shifts the threat landscape toward better-resourced actors and accidental pathways. Respondents expected ASL-3 to disproportionately screen out nonexpert and lone-wolf actors, shifting relative probability toward state actors, well-resourced organizations, and accidental releases from laboratories.

For more information, see the full report.

These forecasts hint at the ways that rapid AI progress could change the world for the better: over a million fewer deaths from currently non-eradicable diseases, a 10-year jump in life expectancy, and a doubling of GDP growth in advanced economies. Yet even as forecasters expect AI to deliver large, measurable gains in health and wealth, many doubt these gains will change how satisfied people feel with their lives. Forecasters also argued that the benefits of rapid AI progress are unlikely to be evenly distributed across the globe.

We asked our forecasters for predictions on the following subjects, conditional on different levels of AI progress by 2030 and at a range of time horizons:

• Global deaths from non-eradicable diseases

• U.S. life expectancy

• GDP growth in advanced and developing economies

• How much users value AI

• Self-reported life satisfaction

More information about Wave 10, including question details, full AI progress scenario descriptions, and analysis of rationales, is available here.

Insight #1: If AI progress is rapid, experts forecast 500,000 annual deaths from tuberculosis, HIV/AIDS, and malaria in 2050, compared with 1.65 million deaths if AI progress is slow

In 2023, about 2.5 million people died from tuberculosis, HIV/AIDS, and malaria—three prominent diseases currently considered non-eradicable. If AI capabilities progress slowly, the median expert forecasts 1.65 million combined deaths from these diseases in 2050. But if AI progress is rapid, this number falls to 500,000—a difference of about 1.1 million deaths a year.

In the slow progress scenario, experts predict deaths in 2050 from TB, HIV/AIDS, and malaria to fall by about 34% relative to 2023 levels. In the rapid scenario, they predict deaths to fall by about 80% relative to 2023 levels. The same pattern holds across superforecaster and public predictions, although with smaller gaps between the rapid and slow AI progress scenarios.

In their written rationales, many forecasters were skeptical that AI-discovered treatments would be a key factor in reducing deaths from non-eradicable diseases in the near term. They cited inadequate health systems and poverty as the reasons high death levels persist. On the other hand, several forecasters pointed to the potential for AI to alleviate poverty and, by extension, reduce deaths from disease in the regions where these deaths are highest.

Insight #2: Experts forecast that rapid AI progress would increase U.S. life expectancy by almost 16 years in 2100, compared with just six years if AI progress is slow

U.S. period life expectancy rose about 13 years between 1946 and 2023, to its current level of 79 years. If AI progress is slow, the median expert expects life expectancy to increase to 85 years by 2100, but if AI progress is rapid, they expect life expectancy by 2100 to reach 95 years. This gap between a slow-AI and a rapid-AI world is a full decade of life expectancy—comparable to most of the gains the U.S. has made since 1950.

Although almost all forecasters thought period life expectancy would rise between 2050 and 2100, many expressed through their written rationales that the gains would be unevenly distributed across the income spectrum. Optimists, on the other hand, thought that AI would address the top sources of mortality in a way that was broadly accessible to all Americans. Others pointed to the life expectancy in countries such as Monaco and San Marino, arguing that the U.S. could catch up to the existing frontier.

Insight #3: Experts expect developing economies to see smaller GDP gains from rapid AI progress than advanced economies

We asked forecasters to predict annualized real GDP growth in advanced economies, and emerging markets and developing countries. Their answers predict a sharp asymmetry, with larger growth gains in richer countries and smaller changes in the developing world.

Advanced economies have grown about 2% a year for the past two decades, but if AI progress is rapid, the median expert expects growth in 2050 to reach 5%—more than double the recent trend. The picture for emerging and developing economies is different. They averaged nearly 6% annual growth in the 2000s and grew 4.4% in 2025, but the median expert expects growth in 2050 to remain close to 6% even under rapid AI progress.

Some respondents argued that AI would destroy the traditional development paths of emerging economies, preventing workers from making the transition from low-productivity to high-productivity jobs. They also argued that robotics and AI could bring manufacturing back to advanced economies, to the disadvantage of emerging economies.

Respondents also thought that much higher rates of growth were possible under a rapid AI progress scenario. The median expert assigned a one-in-ten probability that annualized GDP growth in advanced economies would be at least 10% in 2050. For emerging markets and developing economies, the median expert assigned the same probability that GDP growth in 2050 would be at least 11%.

Insight #4: Experts and superforecasters expect U.S. users to value a month of AI access at about $450 in 2026

Willingness to accept (WTA) measures the payment a person would require to give up a technology for a given period. We asked our panelists to predict the mean WTA among U.S. adult chatbot users. Unconditionally, the median expert and superforecaster put the 2026 value at about $450 a month, roughly $5,400 a year. For comparison, willingness-to-accept experiments in 2019 valued search engines at about $17,530 a year, email at $8,414, and digital maps at $3,648.1

The predicted value of generative AI climbs steeply with time and AI progress. If AI progress is rapid, experts and superforecasters predict that the mean WTA in 2030 will reach $1,500 a month, or roughly $18,000 a year.

Respondents noted that the mean WTA is dominated by a minority of heavy users with an exceptionally high WTA, with high-WTA respondents predicting that a growing subset of people would pay extremely high amounts to access models. Low-WTA respondents instead tended to predict that as AI use expands, new users would draw down the mean over time.

Insight #5: Forecasters expect global life satisfaction to remain flat, even if AI capabilities progress rapidly

The global population-weighted score on the Gallup World Poll’s 0-to-10 Cantril ladder, a measure of how people evaluate their lives overall, was 5.44 in 2025 and has barely moved in over a decade. Forecasters expect that near-flat path to continue regardless of how fast AI advances. Across most forecaster groups and AI progress scenarios, median forecasts stay within a narrow band, rising only to about 5.5–5.6 by 2040, and no median in any scenario exceeds 6.1.

Even as forecasters expect AI to deliver large, measurable gains in health and economic benefit, they doubt these gains will change how satisfied people feel with their lives. Respondents frequently pointed to the stability of the historical trend line, noting that the 2008 financial crisis, the COVID-19 pandemic, and recent inflation had done little to shift self-evaluations of life satisfaction globally. Respondents forecasting flat or lower life satisfaction also noted that wealth stemming from AI adoption would not be broadly shared, and that AI could cause a “crisis in human purpose.”

More from LEAP

This post covers key highlights from the Wave 10 LEAP survey that we conducted between June 16 and July 7, 2026. Recent waves covered robotics, the economic effects of AI, security and geopolitics, AI R&D, and AI timelines.

• Data-damaging worm attacks similar to WannaCry and NotPetya that could cause at least $10 billion in economic damages.1

• Cyberattacks against the U.S. electrical grid that cause a large-scale blackout leading to at least $10 billion or at least $100 billion in economic damages.2

Participants were surveyed between July and August 2025, and some responded to a follow-up survey between December 2025 and January 2026. This study was designed as a pilot, so the results should be interpreted as initial evidence for areas that merit further research, not as definitive estimates.

Because AI cyber capabilities are changing rapidly, the results should not be interpreted as a current assessment of frontier-model cyber capabilities. In particular, at the time of the survey, Claude Mythos Preview/Mythos 5, Claude Fable 5, GPT-5.3-Codex, GPT-5.5/GPT-5.6/GPT-5.5-Cyber, and other cyber-specialized models had not been deployed. We’re working on a follow-up study that will incorporate beliefs about Claude Mythos and other recent AI models, as well as the Hugging Face incident.

The forecasts we elicited in the present study show a consistent pattern: participants assessed baseline risks of catastrophic cyber harms in 2026 as low but non-negligible, and they expected some AI capabilities to substantially increase those risks, especially when AI lowers barriers for moderate-sophistication actors. It is difficult to rule out the possibility that an open-weight, safeguard-free version of a Mythos-level model would reach the level of another capability we described: elite exploit development for moderately skilled individual hackers. If that is the case, then our findings suggest that wide access to these capabilities could substantially increase the risk of a major cyber event, but that mitigation measures could significantly reduce that risk.

This post summarizes the key results from the pilot. You can read the full report here.

Key results

Data-damaging worms were assessed as the higher-risk pathway

Participants estimated a 5–8% probability of at least one data-damaging worm attack causing at least $10 billion in damages in 2026. The median participant’s forecast of expected annual damages from data-damaging worms was approximately $10–15 billion.

Electrical grid cyberattack risks were assessed as substantially lower

Participants estimated a 1% probability that a cyberattack against the U.S. electrical grid would cause at least $10 billion in damages in 2026, and a 0.1% probability that such an attack would cause at least $100 billion in damages. Expected annual damages were roughly $0.2–1 billion, more than an order of magnitude lower than for worms.

AI-enabled elite exploit development produced the largest risk increase

Under a hypothetical scenario in which AI models enable 25% of moderately skilled individual hackers to find vulnerabilities and write elite exploits (zero-click exploits that allow remote code execution with high privileges—the kind that led to the WannaCry and NotPetya attacks), and models with this capability are available open-weight, worm attack risk estimates increase by 3–3.5x. The median expert forecast of a data-damaging worm attack causing at least $10 billion in damages rose from 8% to 41%, while the median superforecaster forecast rose from 5% to 15%.

Forecasters likely underestimated cybersecurity-relevant AI progress

The median expert thought it would take until 2028 for an AI model to solve more than 90% of tasks on the Cybench benchmark, while superforecasters predicted 2030. This capability was most likely surpassed in February 2026, when Claude Opus 4.6 achieved 93% on a subset of problems on the benchmark. It is difficult to rule out the possibility that an open-weight, safeguard-free version of a Mythos-level model would reach the level of another capability we described: elite exploit development for moderately skilled individual hackers.3 Experts and superforecasters forecast this would be achieved in 2032 and 2030, respectively.

Current cyber benchmarks may be informative, but they are imperfect signals of real-world cyber risk

Participants generally saw AI performance on Cybench as informative of technical progress, but not as the best standalone indicator of whether capabilities most relevant to catastrophic cyber outcomes, such as elite exploit development or real-world grid attacks, had been achieved. Respondents viewed current (as of mid-2025) benchmarks as incomplete measures of operational constraints such as zero-day discovery, targeting, coordination, and persistence. Current cyber evaluations may track important technical progress without fully measuring the forms of AI-enabled uplift most relevant to real-world cyber risks.

Model access controls were seen as meaningful but incomplete mitigations

Nearly half of the participants thought that keeping the relevant models—those that could enable moderate-sophistication actors to develop elite exploits—proprietary and protected by anti-jailbreak measures would at least halve the risk of a large-scale worm attack. However, participants also noted that such measures may be less effective against sophisticated actors or exploit-as-a-service markets.

In a follow-up survey conducted four months later, forecasts remained largely unchanged

A subset of participants revisited some of their forecasts between December 2025 and January 2026, after reviewing the original study results and Anthropic’s November 2025 report on an AI-assisted cyber espionage campaign. The median forecast for a data-damaging worm attack causing at least $10 billion in damages in 2026 remained unchanged. Many participants thought the Anthropic cyber espionage report suggested that actors were more capable than they had originally expected, but most did not view it as direct evidence that AI could develop elite exploits or execute a large-scale data-damaging worm attack by 2026.

Additional detail

Rapid advances in artificial intelligence capabilities have introduced new dimensions of uncertainty into the cybersecurity landscape. A central uncertainty is whether AI will strengthen defensive capabilities overall, disproportionately benefit malicious actors, or affect different threat actors in different ways.

This pilot study is an initial effort to systematically forecast how AI might affect large-scale cyber risks over the near term, with a particular focus on 2026. We used structured forecasting methods to elicit judgments about two high-impact cyber threat scenarios: data-damaging worm attacks and cyberattacks against the U.S. electrical grid.

Data-damaging worms

A “worm” is malware that can spread autonomously between systems without an attacker needing to infect each system individually. In this survey we concentrated on “data-damaging worms” that cause damage by wiping, encrypting, or corrupting data on a large number of systems. The WannaCry and NotPetya attacks, both released in 2017, were data-damaging worms that caused roughly $1 billion to $10 billion in damage after infecting hundreds of thousands of systems.

Worm attacks can exploit vulnerabilities: bugs in software or hardware that create security weaknesses in the design, implementation, or operation of a system or application. An exploit is malicious code that takes advantage of one or more software vulnerabilities to infect, disrupt, or take control of a computer without the user’s consent and typically without their knowledge. Following Halstead and Righetti (2026), we define exploits (or exploit chains) as “elite” if they satisfy all of the following criteria:

• Zero-click: Infection requires no user interaction, such as opening emails, clicking links, or visiting a webpage.

• Remote code execution: Allow attackers to execute arbitrary code on a system without the user’s knowledge, and without the attackers needing physical access to the system.

• High privileges: Have administrator privileges or higher.

• Targets widely used software: Effective against more than 10 million systems.

Elite exploits are especially well-suited to worm attacks as they enable autonomous spread and significant damage on a large number of systems. The 2017 leak of elite exploits initially developed by the NSA quickly led to the two major data-damaging worm attacks cited above: WannaCry and NotPetya. Developing elite exploits may require an order of magnitude more skilled researcher time than the other tasks involved in developing a data-damaging worm.

The effect of AI on the risk of worm attacks is uncertain and likely to vary over time. AI systems that can discover vulnerabilities, develop exploits, or automate attack steps are dual-use: the same capabilities could help both defenders and attackers. The net effect depends on several uncertain factors, including which actors gain access to the most capable models, how quickly vulnerabilities are disclosed and patched, and whether patch deployment keeps pace with vulnerability discovery.

Cyberattacks against the U.S. electrical grid

The power grid consists of generators, transmission and distribution networks, and substations. Grid operations, like other infrastructure and industrial processes, rely on two broad types of computer systems:

• Information technology (IT) systems: systems that handle most business operations like billing, email, and administration, and are typically connected to the internet.

• Operational technology (OT) systems: programmable systems that interact with the physical environment or manage devices that do. In grid operations, these systems monitor and control equipment like generators, circuit breakers, and transformers.

Grid cyberattacks involve compromising OT systems, as this is a prerequisite for directly interfering with grid behavior. OT environments, when compared to IT environments, present additional challenges. They typically are—or should be—segmented from IT and internet-facing networks, and they use more niche software and protocols that require more specialized knowledge; OT devices may have relatively individualized configurations to a given environment.

The most relevant prior OT cyberattacks have required substantial time, resources, and specialized expertise, often from state-level actors. Beyond the significant resource and time investments, prior OT cyberattacks have required the integration of diverse capabilities, including long-term reconnaissance, knowledge of the target OT environments, tailored malware, realistic test environments, and stealth.

These requirements make the role of AI difficult to assess. AI could plausibly assist with some components of a grid attack, such as reconnaissance, code generation, vulnerability discovery, and operator decision support. It is less clear how much AI would help with other important bottlenecks, such as avoiding detection, and orchestrating complex, dynamic operations.

Data-damaging worms: main takeaways

The median expert forecasted an 8% probability that a large-scale data-damaging worm attack would cause at least $10 billion in economic damages in 2026, while the median superforecaster forecasted a 5% probability. These baseline forecasts were anchored by the low historical frequency of cyberattacks of this magnitude, although participants noted that the 2017 WannaCry and NotPetya worm attacks demonstrate that worms can spread rapidly, and that future attacks could have large impacts in a more digitally dependent economy.

AI-enabled elite exploit development was seen as a significant driver of increased risk. If an open-weight AI model enabled 25% of moderately skilled individual hackers to find vulnerabilities and write elite exploits, the median expert forecast rose to 41%, and the median superforecaster forecast rose to 15%. Participants viewed this as an important risk pathway because individual hackers are relatively numerous and may be more willing to cause damage, but they are usually capability-constrained; more sophisticated actors are generally more capable but constrained by escalation and retaliation risks.

Participants expected model access controls and safeguards to reduce this risk, especially for actors below state-level sophistication, but not to eliminate it. In particular, having models with the relevant capabilities protected with anti-jailbreak measures was believed to significantly decrease the probability of a large-scale data-damaging worm attack. Model-weight theft, exploit-as-a-service markets, and slow patching of vulnerabilities remained important limitations.

Electrical grid cyberattacks: main takeaways

Participants assessed large-scale cyberattacks against the U.S. grid in 2026 as substantially less likely than data-damaging worm attacks. The median expert and superforecaster both put the probability of a cyberattack against the grid causing at least $10 billion in damages at 1%; for a larger attack, causing at least $100 billion, both medians were 0.1%. Expected annual damages were also much lower, at approximately $0.2–1 billion.

Participants emphasized that grid attacks face additional barriers beyond cyber capabilities, including expertise in industrial control systems (ICS), operational coordination, physical infrastructure constraints, and the risk of geopolitical escalation. They generally viewed a large-scale grid attack as more likely in the context of state-level conflict than in the context of ordinary cybercrime.

Forecasts of large-scale cyberattacks against the U.S. electrical grid were less sensitive to AI capability scenarios than forecasts in the data-damaging worm threat model. AI uplift in an ICS/OT capture-the-flag style competition only modestly increased the median forecast from 1% to 1.5–2%. A real-world AI-enabled warning shot causing at least $100 million in damages produced a larger increase in risk, raising the median forecast to 15% among experts and 4% among superforecasters. Overall, results from this pilot suggest that AI may increase the risk associated with this threat model, but large-scale grid attacks remain constrained by operational complexity and geopolitical considerations.

Conclusion

The study results suggest that AI capabilities could significantly increase some cyber risks in the near term. Both experts and superforecasters predicted substantial increases in the likelihood of a large-scale data-damaging worm attack when open-weight AI enables moderately sophisticated individual hackers to develop elite exploits. Median forecasts suggest that AI-enabled vulnerability discovery could increase the risk of large-scale worm attacks by 3–3.5x, with expected annual damages rising from roughly $10–15 billion to $33–67 billion. Our respondents, answering in mid-2025, believed these capabilities would arrive within the next five to six years, but it is plausible that Claude Mythos has already achieved this capability.

At the same time, structured mitigation measures appear to matter. Participants believed that maintaining proprietary model access with strong anti-jailbreak and monitoring controls could more than halve the probability of such catastrophic outcomes. Limited early access for defenders was considered less effective, though still potentially useful.

Given the small sample size and the speculative nature of some scenarios, these results should be interpreted as directional rather than definitive. Nonetheless, they illustrate the value of structured forecasting for anticipating technology-driven security risks. Future work could explore a wider range of possible cyber harms and include a larger, more representative sample of relevant subject-matter experts.

Further information

For more from this study, including forecasts of threat actors' capability and willingness to carry out cyberattacks, and more details on expected damages from attacks, see the full report.

Several models are already statistically indistinguishable from superforecaster-level accuracy

On the ForecastBench tournament leaderboard, several models are statistically indistinguishable from superforecaster-level accuracy.1 The top current system is from Cassi AI, with models from xAI and Google DeepMind also ranked as indistinguishable from superforecaster-level accuracy. The tournament leaderboard is intended to capture the frontier of LLM forecasting ability, so submissions may make use of tools, added context, fine-tuning, ensembling, or other methods.

We described the approach Cassi is taking in a previous blog post back in January. In essence, the approach uses a multi-stage pipeline that generates sub-questions and search queries, retrieves up-to-date context, and filters for relevance and recency. The system then generates forecasts using an ensemble of models, with another LLM analyzing the reasoning to produce a final forecast.

Although the overall architecture remains similar, Cassi has made some incremental improvements to the system since our January post. These include refinements to the forecasting pipeline, such as how the agent gathers, organizes, and synthesizes information, and better integration of relevant data sources.

Other systems have also likely reached parity with superforecasters on ForecastBench. These include external submissions from xAI and Google DeepMind.

An AI system has now ranked higher than superforecasters on market questions for the first time

On market questions, Cassi’s system also ranks higher than the superforecaster median forecast—the first time that any model has achieved this.2

For context, ForecastBench is divided into two different kinds of questions:

• Dataset questions: Automatically generated from real-world time series (ACLED, DBnomics, FRED, Yahoo! Finance, and Wikipedia) using pre-specified templates.

• Market questions: Drawn from leading prediction platforms (Manifold, Metaculus, Polymarket, and RAND Forecasting Initiative).

AI systems matched superforecasters on dataset questions back in May. However, in some ways market questions are a better test of human-like forecasting skill. LLMs may have an advantage over humans on dataset questions because they involve looking up lots of data, establishing base rates, and so on—tasks for which LLMs are uniquely well suited. Market questions often require judgment about novel, one-off events, so it’s significant that an AI system now outranks superforecasters on these questions.3

Many AI systems likely outperform superforecasters on dataset questions

17 submissions now rank above superforecasters on our preliminary leaderboard. The preliminary leaderboard ranks models based on their performance on resolved dataset questions. Given that until very recently Google DeepMind’s green-tree was the only submission ranked higher than superforecasters on this leaderboard, this is a significant development.

Our preliminary leaderboard provides early results before a model appears on the tournament leaderboard. While it takes 50 days after submission for a system to appear on the main tournament leaderboard, submissions appear on the preliminary leaderboard after at least 225 dataset questions have resolved (typically around 10 days after the initial submission). These rankings are not final, but our stability analysis suggests that these early results are robust. So, we expect many of these systems to rank higher than superforecasters on dataset questions once they appear on the tournament leaderboard.

Three submissions from Torchcast are at the top of the preliminary leaderboard. Torchcast told us their approach uses a proprietary multi-agent architecture to analyze each question, gather relevant evidence, generate forecasting perspectives, and produce a final probability estimate. They use a combination of Torchcast-trained models and commercially available LLMs.

Caveats

There are a few caveats that should be considered when interpreting all of these results:

• We last elicited superforecaster predictions in 2024. Our comparison relies on a statistical extrapolation that grows less reliable over time.

• Results depend on realized questions and are inherently stochastic. As more questions resolve, the results may shift.

• The 95% CIs for many of these submissions overlap substantially; the results are more consistent with superforecaster parity than with outperformance.

• These findings don’t mean that ForecastBench is saturated. The superforecaster median is a human reference point, and it is possible that AI systems may exceed human superforecaster accuracy.

Next steps

We’re working on a few ways to improve ForecastBench. Upcoming changes include:

• A fresh superforecaster round (fall 2026)

• Updated dataset questions (fall 2026)

• Quantile questions (fall 2026)

Some of the most consequential forecasting questions are also the hardest to evaluate. Many high-stakes policy and business decisions depend on expectations about long-term changes, low-probability events, or the causal effect of interventions—all of which are challenging to measure in the near term. How do we know if forecasts on questions like these are accurate?

One approach is to look for proxies: easy-to-collect signals that predict accuracy on the questions we really care about. In February, we published the results of a forecasting study in which we asked experts and superforecasters to predict the outcome of a randomized controlled trial (RCT) testing whether LLMs increased novices’ ability to complete molecular biology tasks in a wet lab. Since the RCT itself provided a measurable outcome, this study gave us a chance to test whether faster, lower-cost signals track performance on the forecasts that really matter.

You can read more about the RCT forecasts here, but in short, we asked superforecasters, biosecurity experts, and virology experts to predict how well novices would perform at various molecular biology tasks and whether access to LLMs improves (”uplifts”) performance.

Alongside the main outcomes, we asked participants to answer 10 calibration questions (brief factual questions answered without prior research), and to share intersubjective forecasts (asking each respondent to predict the group’s median forecast).1 In this post, we examine whether performance on either of these proxies predicted accuracy on the RCT forecasts.

Calibration questions

Before they began forecasting the RCT results, we asked survey participants to answer 10 quick calibration questions. These questions asked how often rare events occur or what proportion of a population shares a rare characteristic. They were asked to do this without research and were shown the correct responses after they answered.

Here are two examples of calibration questions we asked:

• What is the probability that a randomly chosen person in the world is a millionaire (total wealth of $1 million or more)?2

• What is the probability that a randomly chosen person in the U.S. is a physician?3

We scored performance on calibration questions using mean squared error (MSE). Figure 1 shows the correlation between MSE on the calibration questions and s-scores on forecasts of the main RCT outcome.45\

The correlation is 0.267, so we do select better individual forecasters by focusing on those who were most calibrated. But does that mean calibration questions are useful to a decision-maker? One way to assess the usefulness of calibration testing is to ask the following question: “If we selected the best forecasters according to calibration performance, would that give us a clearly superior group of RCT forecasters?” To test this, we selected both the top and bottom 20% (n=15)6 of forecasters based on calibration performance and compared these groups’ accuracy on the RCT outcome forecast with sets of forecasters chosen at random. If calibration performance is useful for identifying better and worse forecasters, it should at least do so better than random selection.7

As we see, the group of top 15 respondents selected by accuracy on calibration questions has a higher average RCT accuracy (lower s-score) than 99 of 100 groups of 15 drawn at random. Calibration scores are even more effective at identifying poor RCT forecasters: the worst 15 calibration performers have a mean RCT accuracy worse (a higher s-score) than all 100 randomly drawn groups.8 This suggests that calibration scores provide a promising signal for identifying stronger and weaker forecasters, albeit still some distance from selection based on actual RCT s-scores.

Intersubjective forecasts

Intersubjective questions are another possible proxy for forecaster accuracy. In these questions, respondents were asked to predict the median responses of other groups taking part in this study. These forecasts are useful because they are immediately resolvable, and they test how well respondents understand the perspectives and reasoning of different communities in addition to their general forecasting skill. Plausibly, those who do well at that task would also do well at the RCT forecasts.

We asked respondents to forecast the median forecasts of superforecasters, biosecurity experts, and virology experts on two questions:

1. The difference in the success rate between the LLM-plus-internet and the internet-only groups (the main RCT outcome).

2. The likelihood of a human-caused large-scale pathogen outbreak in 2028 that leads to at least 10,000 deaths or $100 billion in damage within three years.

For each participant, we calculated the distance of their predictions from the actual group medians, averaging the absolute error across all three groups for each question. Participants were then ranked from most to least accurate. Figure 3 below applies the same approach used for calibration scores, comparing the mean s-score of the top and bottom 15 intersubjective forecasters with that of randomly selected groups of 15 from the same sample.

Here, the mean RCT s-score of the top 15 intersubjective accuracy group falls at the center of the random distribution. The worst 15 forecasters by intersubjective accuracy distinguish themselves somewhat more than the top 15 forecasters, with a mean RCT accuracy worse than that of 85 of 100 random groups. Overall, intersubjective accuracy appears to be a weak proxy for RCT forecasting performance but is perhaps more promising for identifying the worst forecasters.

Implications

Our objective in this work was to determine whether quick-to-collect signals—accuracy on calibration questions and intersubjective forecasts—served as good proxies for accuracy on the main outcome of interest: predicting the result of an RCT measuring LLM uplift on molecular biology tasks.

The results we have so far suggest that brief proxy assessments have potential but remain limited tools for identifying the best forecasters for more consequential questions. When we selected forecasters by calibration score, the top and bottom 15 outperformed or underperformed all or nearly all randomly selected groups. Selection based on the intersubjective forecast scores was less distinguishable from the randomly selected groups, especially for the top 15 group. However, the two proxies still fell well short of the best possible selection based on actual RCT forecast accuracy.

There are a few useful lessons here:

1. Scores on calibration questions worked well at identifying a group of people who had more accurate aggregate forecasts on the main outcome. This suggests that calibration questions may be a promising way to quickly identify accurate forecasters.

2. Calibration scores and intersubjective scores seem to be more useful for identifying the worst forecasters than they are for identifying the best. This echoes earlier FRI work that found that cognitive tests generally provide more information about people with lower forecasting proficiency than about higher-performing forecasters. Thus, calibration questions in particular could be a relatively easy way to filter out lower-proficiency forecasters.

3. Calibration scores had a stronger predictive signal for accuracy on RCT outcome forecasts than did intersubjective forecast accuracy. Intersubjective scoring has shown promise in other settings—for example, incentivizing accurate forecasts for unresolvable questions—but in this study, it did not seem to capture the same skill as predicting an RCT outcome.

Of course, all the usual caveats apply: this is a single study with 77 forecasters, nine calibration questions, two intersubjective questions, and one main prediction of interest. But the broader takeaway is consistent with our wider research: identifying skilled forecasters efficiently is difficult, but potentially feasible and high-value.9 That’s one reason we are tracking expert and superforecaster predictions on AI progress over many years through our Longitudinal Expert AI Panel (LEAP). In time, we hope that the resolved forecasts from LEAP will help us identify the best forecasters of AI progress and perhaps reveal broader lessons about how to identify good forecasters. In the meantime, we will continue to explore proxy metrics and forecasting proficiency screens as a way of identifying the most (and least) accurate forecasters before questions have been resolved.

We also asked for forecasts on the likelihood of a smaller-scale AI harm event, financial loss from cybercrime, the chance of top-ten economies backsliding into authoritarianism, and the probability of a major government restricting an AI model release. For many of these questions, we asked forecasters to consider specific AI progress scenarios, allowing us to partially identify the additional risk posed by rapid near-term AI progress.

The next LEAP wave will focus on the benefits of AI. We’ll release those forecasts in our next LEAP wave report and expect them to pose an interesting counterpoint to the forecasts below.

This post covers key highlights from the Wave 9 LEAP survey that we conducted between May 19 and June 10, 2026. Recent waves covered robotics, the economic effects of AI, security and geopolitics, AI R&D, and AI timelines. More information about Wave 9, including question details and analysis of rationales, is available here.

Insight 1: Experts and superforecasters think a major AI-driven harm event is more likely than not by 2050

We asked forecasters about the probability of an AI-driven event causing the deaths of at least 50 people or $100 billion in damages before the end of 2050. The median expert gave a 62% probability that this would occur by 2050, while the median superforecaster predicted 70%. Both of these forecasts are considerably above the general public’s expectations—they assigned a 35% probability to this outcome.

Conditional on a major AI-driven harm event occurring, both experts and superforecasters gave 50% odds that this would happen by 2035. Both groups gave a one-in-four chance of such an event happening by 2031. Most commonly, forecasters expected such events to arise from AI-introduced errors in critical infrastructure systems and cyberattacks, followed by autonomous or AI-assisted military systems. Below these, transportation accidents and deliberate human weaponization of AI form a secondary tier.

Insight 2: If AI capabilities progress rapidly, forecasters predict a roughly 5x increase in the probability of a global AI-related catastrophe, compared with a world with slow AI progress

We also asked forecasters to predict the likelihood of a global AI-related catastrophe in which more than 10% of the people alive at the start of a five-year period die by the end of that period. If AI capabilities progress slowly, the median expert assigns this a 0.08% probability by 2030, 1% by 2050, and 2% by 2100. In a scenario where AI progress by 2030 is rapid, those forecasts rise to 1%, 5%, and 10%.1

Our slow progress scenario describes a world in which AI is merely a capable assistant to human workers, while our rapid scenario depicts AI that can match or surpass the best humans across cognitive and physical tasks. Expert forecasts, then, predict that moving from the slow world to the rapid world also implies a five-fold increase in the risk of a deadly AI-related global catastrophe.

In their written rationales, both optimistic and pessimistic forecasters agreed that faster progress raises risk because it compresses the time available for policy and society to adapt. However, forecasters disagreed on how likely it was that a catastrophe would meet the threshold of killing 10% of the global population, with some expecting warning events to trigger protective mandates before this threshold was reached.

Insight 3: If AI capabilities advance rapidly, experts attribute roughly two-thirds of all global catastrophic risk this century to AI

We also asked respondents for the probability of a global catastrophe from any cause—again defining catastrophe as the death of more than 10% of the people alive at the start of a five-year period. Conditional on AI capabilities progressing rapidly by 2030, the median expert forecasts the probability of an AI-related catastrophe by 2100 at 10% compared with 15% for a catastrophe from any cause.

We computed each respondent’s ratio of AI catastrophic risk to total catastrophic risk. The median expert attributes about two-thirds (67%) of total global catastrophic risk to AI in a world where AI capabilities progress rapidly, but roughly 30% in a slow progress world.

In their rationales, forecasters who assigned a higher share of total catastrophic risk to AI argued that the technology is becoming so entangled with every catastrophic pathway that a genuinely AI-free catastrophe is hard to imagine. Those who saw AI catastrophic risk as a smaller percentage rejected the premise that AI is entangled with everything, treating the dominant non-AI pathways as substantial standalone risks.

Insight 4: Experts and superforecasters give a 50% chance that a major government restricts an AI release by 2030

The median expert and superforecaster assign a 50% chance to the U.S., UK, or EU restricting the initial public release of an AI system on safety grounds by 2030. The median expert gives 25% odds of this happening by 2028 and 75% by 2034.

On June 12, shortly after our survey closed, the U.S. government issued an export-control directive suspending access by any foreign national to Anthropic’s Mythos 5 and Fable 5 models. As this happened after the models were released, this did not meet our resolution criterion that such a directive must restrict the system’s release. We will monitor ongoing news surrounding GPT-5.6 and related models to see if near-term government actions meet the criteria.

Early-restriction forecasters read Anthropic’s voluntary withholding of its Mythos model as evidence the dangerous-capability threshold has arrived, and pointed to the EU AI Act’s Article 93 powers that take effect in August 2026 as the earliest plausible trigger. Late-restriction forecasters argued that voluntary industry restraint substitutes for state action and that governments typically regulate deployment and use rather than pre-release publication.

Insight 5: Experts predict a roughly 25% chance that two more of the world’s current eight largest democracies will backslide into authoritarianism by 2040 if AI progress is rapid

Two out of the world’s ten largest economies are currently classified as “authoritarian” on the Economist Intelligence Unit (EIU) Democracy Index: China and Russia. In an unconditional scenario, the median expert and superforecaster expect that count to stay at roughly the same level through 2040.

However, if AI progress is rapid, the median expert assigns a roughly 25% chance of two of today’s eight largest democracies (the U.S., Germany, Japan, UK, India, France, Italy, and Canada) backsliding into authoritarianism, bringing the total to four or more authoritarian states among the top-ten economies by 2040, and a 5% chance the total reaches five or more. Those assigning higher counts argued that rapid AI progress amplifies the tools of surveillance, censorship, and repression and widens the range of political outcomes, disproportionately adding probability to less democratic results.

Insight 6: Experts predict that cybercrime losses by 2031 would be roughly 52% higher in a world where AI capabilities progress rapidly compared with a world with slow AI progress

In 2025, the FBI reported nearly $21 billion in financial losses from cyber-enabled crime—more than double the figure in 2022 ($10.3 billion). If AI capabilities progress slowly through 2030, the median expert predicts that losses reported to the FBI will increase to $25 billion in 2026, $35 billion in 2028, and $46 billion in 2031. If AI progress is rapid by 2030, they predict reported losses of $31 billion, $50 billion, and $70 billion, respectively. In a rapid-progress world, experts give a one-in-four chance that losses are $93 billion or higher by 2031.

Many respondents reported that AI-enabled deception would be a primary driver of future losses, citing deepfakes, automated targeting, and personalized messages as factors that would increase the rate of fraud. Forecasters who predicted high cybercrime losses argued that AI advantages attackers, with defenders structurally disadvantaged and legacy systems and human behavior remaining highly vulnerable to attacks.

Find out more about LEAP

To read more about Wave 9 of LEAP, including further questions and rationale analysis, visit the LEAP website.

Good forecasts often come with a rationale—the reasoning behind the number. In our Longitudinal Expert AI Panel (LEAP), for example, forecasters write hundreds of thousands of words each month, explaining their logic, citing evidence, and weighing competing considerations.

That data lets us report not just that AI experts expect a robot to reliably make a cup of coffee in less than a decade, but why they think so. For example, these are two LEAP rationales from respondents who expect robots to be able to reliably make a cup of coffee soon (source):

But here’s what we don’t know: which features of a rationale predict forecasting accuracy? Is statistical reasoning an indicator of forecasting accuracy? What about causal analysis based on domain expertise, as in the rationales highlighted above?

That’s where our latest research comes in. In a new working paper, led by Chris Karvetski and Sheldon Huang, together with Simas Kučinskas, Nadja Flechner, Jingyu Hu, Phil Tetlock, and Ezra Karger, we study which rationale features are correlated with forecasting accuracy.

Our approach

First, we defined sixty Explanation Quality Markers, or EQMs for short, inspired by the judgment and decision-making literature. Each EQM was described in natural language together with a directional hypothesis: is the EQM expected to help accuracy, or hurt it?

Some EQMs are “good habits,” like statistical or fact-based reasoning. Others are “warning signs,” such as guessing, confirmation bias, or extreme confidence. We then used an LLM (GPT-4o for most of the analysis) to score each rationale against the sixty EQMs on a 0/1/2 scale. For example, a rationale that explicitly cites a base rate might score a 2 on Statistical Reasoning, while one relying on a gut feeling would get a score of 0. For interpretability, we collapsed those sixty scores into a single composite number via a statistical model. Finally, we correlated that composite score with forecasting accuracy. The full scoring pipeline is shown below.

We ran this scoring pipeline on more than 55,000 forecast–rationale pairs from the ACE geopolitical forecasting tournament, the IARPA-funded competition (2011–2015) that produced the Good Judgment Project and the original work on superforecasters. The methodology is cheap and scalable. It cost roughly $0.007 per rationale, or about $400 in total, to score the ACE rationales.

Insight 1: Statistical reasoning is rare

EQMs let us systematically analyze the reasoning behind the forecasts.

The first thing the EQMs reveal is that forecasters in the ACE tournament typically expressed their reasoning in causal, not statistical, terms: Causal Reasoning shows up in 77% of rationales, while Statistical Reasoning appears in just 19%, a fourfold difference.

The single most common EQM is Forecast and Rationale Align (97%); that is, forecasters’ attitudes—as expressed in their rationales—usually match their quantitative forecasts. However, problematic reasoning patterns are also common. Speculative Terms (55%), Simplification Bias (34%), and Gut Based (31%) all turn up frequently.

Insight 2: EQMs predict accuracy, with caveats

OK, but can the EQMs actually predict forecasting accuracy? And do they outperform previous methods?

We measured accuracy in two ways. Forecast-level accuracy asks whether an individual prediction was more accurate than the crowd median on that question. Forecaster-level accuracy measures whether a person was accurate across a whole season of questions. The EQM composite score correlated more strongly with forecasting accuracy than a pre-LLM benchmark did, at both the forecast level (r = .19 vs. .06) and the forecaster level (r = .51 vs. .39); both differences were statistically significant. For context, the pre-LLM benchmark consisted of a rich set of 100+ features, including a comparison-class classifier (Karvetski et al., 2022), automated measures of reasoning complexity, and a standard psycholinguistic dictionary (LIWC 2022). So, EQMs cleared a real bar.

Which EQMs are doing the work?

The graph below shows where each EQM lands in terms of both forecaster-level accuracy (x-axis) and forecast-level accuracy (y-axis).

EQMs in the upper-right quadrant (positively associated with both accuracy metrics) include Forecast and Rationale Align, Fact Based, and Concrete Reasoning. The lower-left quadrant (negatively associated with both accuracy metrics) is dominated by the bias family, including Gut Based, Simplification Bias, Confirmation Bias, and Extreme Confidence. In particular, Forecast and Rationale Misalign had the strongest negative correlation with forecast-level accuracy. Statistical Reasoning correlated more strongly with accuracy than Causal Reasoning for both metrics (forecast level: r = .04 vs. .01; forecaster level: r = .26 vs. .05), despite being much less prevalent. In total, over 90% of the statistically significant correlations matched the hypothesized direction. No pre-LLM feature reached the predictive strength of the top EQM patterns, as shown by the gray points clustering around zero.

However, EQMs flag bad forecasts and forecasters more reliably than they pick out excellent ones. In the graph below, we sort the data into nine bins by the EQM composite score. At the forecast level, accuracy jumps by 0.09 across the bottom third but only by 0.01 and 0.02 across the middle and top thirds. At the forecaster level, the biggest gain (0.24) is also in the bottom third. Intuitively, a low EQM score appears to be a “red flag,” but a high one is a much weaker “green flag.”

Insight 3: What looks good to humans isn’t always what’s accurate

During the ACE tournament, forecasters also rated each other’s rationales for quality.1 That lets us ask a more pointed question: do EQMs outperform human ratings?

The answer is “mostly yes.” Human ratings correlate with forecast-level accuracy at just r = .07, versus r = .23 for the EQM composite score; at the forecaster level, human ratings underperform at r = .40 vs. .50.2

As shown in the figure above, human ratings strongly correlate with rationale length: rationale length correlates with average ratings at r = .62, yet it is essentially unrelated to forecast-level accuracy. Humans also strongly preferred fact-based reasoning (Fact Based). Human raters weren’t wrong directionally, but they appeared to place undue weight on some features. For example, as we show in more detail in the working paper, they underweighted “red flags” such as Extreme Confidence.

In an out-of-sample test (Team Dynamics dataset), human ratings performed similarly well at the forecast level, although they still lagged at the forecaster level. So, the head-to-head performance of EQMs vs. humans at the forecast level appears to be context dependent. As in the ACE dataset, human ratings remained strongly correlated with rationale length.

Takeaways

So, can we spot a good forecast by its rationale? Three findings stand out:

• EQMs carry real signal. EQMs predict accuracy, and they beat both older text-analysis tools and (often) humans’ own quality ratings. They do so cheaply, at scale, and from a single written rationale.

• EQMs are mostly a screen, not a talent detector. The predictive signal is asymmetric: EQMs flag weak forecasts and forecasters more reliably than they single out the best. We observe the same pattern when EQMs are used for forecast aggregation. Ranking forecasters by their EQM score helps about as much as ranking by past accuracy, while ranking individual forecasts barely moves the needle.

• What looks good to humans isn’t always what’s accurate. Human raters tend to reward length and a fact-based tone more than those features deserve, and underweight warning signs such as overconfidence.

Two caveats to keep in mind. First, rationales provide a limited view into the reasoning underlying forecasts. Second, our results are correlational, not causal. The ACE tournament included a randomized experiment in which some forecasters received the CHAMPS KNOW training, a structured program teaching probabilistic-forecasting skills. We use the randomized treatment to check whether the training shifted the EQMs it was meant to target. It did: trained forecasters scored higher on exactly those markers (e.g., Statistical Reasoning, Statistical Causal Blend, and Best Practices). However, the evidence is preliminary, and the training’s effect on forecasting accuracy itself was more mixed.

Nevertheless, the results are encouraging. The written reasoning that already accompanies many forecasts is a largely untapped source of information. EQMs are an exciting new method to begin extracting it.

Read the working paper for the detailed methodology and full results here.

We asked panelists for their predictions on AGI timelines,1 near-term progress on METR’s task-completion time horizon benchmark, the factors that will enable or block AGI by 2040, and timelines for a “rapid AI progress” scenario. We also asked panelists to assess AI’s overall impact on the U.S. over the next 20 years and its effects on people’s ability to solve problems, make difficult decisions, form meaningful relationships, and think creatively.

Every month, LEAP tracks the views of top AI scientists, industry leaders, policy researchers, economists, and high-performing forecasters on the trajectory of AI’s development and use. Recent waves covered robotics, the economic effects of AI, security and geopolitics, and AI R&D. You can see a full list of LEAP questions and forecasts here.

This post covers key highlights from the Wave 8 LEAP survey that we conducted between April 20 and May 11, 2026. More information about this wave, including question details and analysis of rationales, is available here.

Insight 1: Experts and superforecasters have updated their expectations of AI’s impact upward in the last nine months

Back in June/August 2025, we asked LEAP panelists to assess the likelihood of AI in 2040 reaching various levels of impact on the Technological Richter Scale (TRS)—a scale devised by Nate Silver to rank technologies according to their broad societal impact. Credit cards, for example, are ranked as a “technology of the decade” (Level 7) while the rise of humans ranks as a “technology of the epoch” (Level 10). You can view the forecasts from the first time we asked the question here.

We repeated this question in the current April/May 2026 wave, nine months after we first asked for forecasts. Experts and superforecasters have increased the likelihood they assign to AI’s impact being comparable to higher levels of technological importance, more akin to a “technology of the millennium” (e.g. agriculture, Level 9), but experts continue to assign the greatest probability to AI being comparable to “the technology of the century” (e.g. electricity, Level 8).

On average, experts assign a 35% chance of AI reaching TRS Level 8 by 2040, with substantial weight on Level 9 (24%) and Level 10 (11%). Superforecasters gave a near-identical distribution (34% to Level 8, 23% to Level 9, 8.5% to Level 10), while the public is meaningfully more conservative.

Among the 264 expert and superforecaster participants who completed both Wave 1 and Wave 8, the mean expected TRS levels rose for experts by +0.20 (from 7.86 to 8.06), with larger movement among superforecasters (+0.39, from 7.50 to 7.89) and minimal movement among the public (+0.07, from 7.18 to 7.25). 74% of participants remained within ±1 point of their Wave 1 forecast, while 15% increased and 11% decreased by more than one point.

This shift is also visible in participants’ modal forecasts—the level each person assigns the highest probability to. Among matched experts, the share whose modal forecast is “technology of the century” grew from 38% to 53%. Superforecasters showed a more pronounced change: “technology of the decade” was the most common modal choice in Wave 1 (38% of superforecasters), but this was replaced by “the technology of the century” in Wave 8 (43%).

Insight 2: Most experts expect “AGI” before 2100, with a median forecast that it will occur by 2050

We asked respondents for the probability that, before 2100, more than 50% of LEAP panelists will agree that artificial general intelligence (AGI) exists and—if such a scenario occurs—the year in which they expect this to happen. We defined AGI as a commercially available AI system that meets both of the following criteria:

• Can outperform the 90th percentile professional human employee in every primarily non-physical occupation (based on 2025 performance), across all sectors, on at least 90% of the economically useful non-physical tasks that they perform.

• Has an inference cost no more than 5x the cost of equivalent human labor.

The median expert forecasts an 80% probability that the majority of LEAP panelists will agree AGI exists before 2100 according to this definition. Conditional on this occurring, experts give a median forecast of this occurring in 2050, with a 25% probability by 2039 and a 75% probability by 2065. Superforecasters give a similar median probability (80%) and a slightly earlier median year of occurrence (2047).

Insight 3: Experts predict that AI will be able to complete eight-hour tasks with 80% success by 2030 on METR’s task-completion time horizon benchmark

When asked when an AI model will achieve 80% success on software tasks requiring 8 hours or more of human expert effort, the median expert gives a 50% probability of 2030 or earlier. The median superforecaster reports more aggressive timelines, forecasting 2028, while the public forecasts a significantly slower timeline of 2037 at the median.

We also asked forecasters to predict the longest 80% success time horizon achieved by the end of 2026, and all three groups’ medians were between 3 and 4 hours (experts 3.4, superforecasters 3.5, public 3), up from a baseline of 1.5 hours at time of survey launch (April 20, 2026).

On May 8, 2026 (toward the end of our survey period), METR updated its benchmark to include a preview of Anthropic’s Mythos model. The model achieved an 80% time horizon of 3 hours and 6 minutes, already within the range of the median expert and superforecaster predictions with more than seven months remaining in 2026.

Insight 4: Experts and superforecasters tend to view AI more optimistically than the general public

We asked panelists for their views on AI’s impact on the U.S. over the next 20 years. We asked for their broad predictions as well as more specific forecasts as to whether AI will make people better or worse at thinking creatively, making difficult decisions, solving problems, and forming meaningful relationships with other people.

When asked about AI’s overall impact over the next 20 years, 57.5% of experts and 69.8% of superforecasters predict a somewhat or very positive impact, compared to 42.0% of the public. The divergence is sharpest on problem-solving (72.7% of experts and 69.8% of superforecasters say AI will make people better at this, compared to 48.6% of the public) and on making difficult decisions, where 58.3% of experts expect improvement compared to just 36.9% of the public, 39% of whom expect people to become worse at this.

One area in which experts’ and the public’s views converge is about AI’s impact on forming meaningful relationships: 68.4% of experts and 66.7% of the public expect people to become worse at this, in contrast to just 50.9% of superforecasters expecting this.

Find Out More About LEAP

To read more about Wave 8 of LEAP, including further questions and rationale analysis, visit the LEAP website.

Every month, LEAP tracks the views of top AI scientists, industry leaders, policy researchers, economists, and high-performing forecasters on the trajectory of the development and use of artificial intelligence. Earlier LEAP waves covered AI research and development, security and geopolitics, using AI for science, and more.

This post covers key highlights from the Wave 7 LEAP survey that we conducted between March 9 and 31, 2026. More information about this wave, including question details and analysis of rationales, is available here.

Help Us Improve LEAP

We’re interviewing a small number of LEAP users to learn how we can make our work more useful for decision-makers. If you’re interested in taking part, please fill out this Google Form.

We’re offering a $50 voucher to everyone we interview.

Insight 1: Experts expect robots to be able to reliably make coffee in an unfamiliar environment by 2034

The “Coffee Test” requires a robot to make a cup of coffee in three different unstructured, unfamiliar environments of equivalent complexity to an ordinary home, without making any mistakes. This test is generally attributed to Apple co-founder Steve Wozniak, who in 2007 predicted that “we will never see a robot that makes a cup of coffee, never.”

LEAP forecasters are more optimistic than Wozniak, but generally do not expect an imminent breakthrough on the Coffee Test. The median expert forecasts that an autonomous robot will first successfully complete the Coffee Test in 2034, with superforecasters predicting 2032 and the general public predicting 2039.

In their written rationales, forecasters disagreed in their assessment of current robot capabilities. Respondents who predicted an earlier resolution tended to argue that the technical ability to pass such a test already exists, while those who had longer timelines typically saw a wide chasm between current systems and the level required to pass the test. Forecasters highlighted robot dexterity and generalization as key barriers to passing the Coffee Test.

Insight 2: Respondents expect fully robotic surgery to happen later than the Coffee Test

Experts forecast that a robot will first successfully perform an appendectomy without physical human assistance around 2040, with superforecasters predicting 2035 and the public 2039. The later timeline for surgery is notable given that, unlike the Coffee Test, the appendectomy test explicitly permits verbal coaching by a human. In rationales, forecasters expecting a later resolution date pointed to regulatory barriers, malpractice liability, and surgeon lobbying as potential multi-decade blockers of an AI appendectomy.1

As with responses to the Coffee Test, forecasts show a large amount of disagreement. The interquartile range for experts' 50th percentile forecasts on the autonomous appendectomy spans from 2035 to 2046, while the interquartile range for the Coffee Test spans from 2030 to 2040. Across the 50th percentile forecasts for both questions, superforecasters have the smallest interquartile range while the public has the largest.

Insight 3: Despite forecasting a world with far more industrial robots, experts forecast labor's share of the economy to fall only modestly

We asked forecasters to predict the U.S. labor share for the private, nonfarm sector in 2026, 2030, and 2040 as well as the number of industrial robots in use globally in those same years. Labor share is the proportion of economic output that goes to workers in the form of income and benefits, rather than to capital. Some AI thought leaders believe that AI will eventually replace many human workers, driving down labor’s share of economic output, and some recent economic studies suggest that automation may have reduced the labor share of economic output.

All participants groups expect the number of industrial robots in use globally to continue to grow from a 2024 baseline of 4.66 million units. The median expert predicts industrial robot stocks will grow to 16 million units by 2040, with superforecasters predicting a similar number.

Despite this, all groups forecast relatively modest declines in labor share. From a 2024 baseline of 58.3%, the median expert forecasts U.S. labor share declining to 51% by 2040—a roughly 7 percentage point drop over 16 years, compared to the roughly 10 percentage point decline that occurred over the previous 23 years.

Forecaster rationales suggest that experts see the growth in industrial robots as important, but many note that labor share changes slowly, and is embedded in institutional and demographic structures that change over decades rather than years. As one moderate-decline respondent put it: “labor share moves like a supertanker.” Even respondents who cite advances in AI in their rationales tend to anchor to the current downward trend in labor share rather than forecast an acceleration.

Insight 4: Experts forecast Amazon's e-commerce sales per distribution employee to rise

The median expert forecasts Amazon’s real e-commerce net sales per distribution employee to rise from an estimated 2024 baseline of $464,771 to $480,000 by 2026, $549,842 by 2030, and $800,000 by 2040.2 Superforecasters are more bullish, predicting over $1,000,000 by 2040.

Forecasters cited the suitability of warehouses for robotic deployment, Amazon’s track record in automation and a decoupling of sales from labor as key factors driving their predictions.

Find Out More About LEAP

To read more about Wave 7 of LEAP, including further questions and rationale analysis, visit the LEAP website. If you would like to participate in LEAP feedback, please fill in this Google Form.

In a new working paper, researchers from the Forecasting Research Institute and coauthors from the Federal Reserve Bank of Chicago, Yale School of Management, Stanford University, and the University of Pennsylvania present results from a large-scale forecasting exercise tracking the views of 69 leading economists, 52 AI industry and policy experts, 38 highly accurate forecasters, and 401 members of the general public. The survey ran from mid-October 2025 to the end of February 2026.

This post summarizes the key findings. For more details, refer to the full working paper.

We also built a tool so you can compare your own forecasts of AI’s economic impact to economists’ forecasts. When you submit your forecasts, you will get a shareable image that compares your predictions to economists’ predictions.

Executive Summary

1. Economists, AI experts, and superforecasters expect that, by 2030, AI will significantly exceed the capabilities of present-day systems; but economists predict that key indicators—including U.S. GDP, total factor productivity, and labor force participation—will remain close to historical trends.

2. Economists assign a 14% chance to an exceptionally rapid-progress scenario materializing, which they predict could lead to major increases in GDP growth and wealth inequality, along with a large decrease in labor force participation.

3. Economists support job retraining as a response to AI’s economic impacts, predicting it would increase labor force participation in a rapid AI progress scenario by 1 percentage point (roughly 2.76 million additional people in the labor force), while providing a 0.2 percentage point boost to GDP growth.

Key Findings

1. Respondents predict significant AI progress by 2030

As shown in Figure 1 below, economists gave a 47% probability that a moderate AI progress scenario would most closely match the world in 2030. This describes a world in which AI can run semi-autonomous research laboratories, complete coding tasks that would take a human five days, produce high-quality novels, manage complex business roles, and where robots exist that can navigate homes and perform basic tasks. This is a significant advance from present-day AI capabilities.

On average, economists think there is a 14% chance that in 2030, AI progress will most closely match a rapid scenario where AI outperforms all humans at many tasks, can run businesses at the level of a competent CEO, and where robots exist that can perform nearly all in-home and industrial tasks faster than humans. In total, the average economist assigned a 61% probability to either the moderate or rapid progress scenario. AI experts gave a similar combined probability (62%), while superforecasters (55%) and the general public (59%) were only slightly more pessimistic.

2. Despite predicting significant AI progress, economists do not expect major departures from key economic trends

We asked respondents to forecast key economic indicators in 2030 and 2050, assuming their current beliefs about the rate of AI progress. These unconditional forecasts generally clustered around existing trends, although a labor force participation rate of 58% has not been seen since the 1960s, and is well below the Congressional Budget Office's projections for 2050 (see below).

Key insights #1 and #2 seem in tension: Economists expect AI capabilities to improve significantly by 2030, but they do not expect this to translate into dramatically different economic outcomes.

What explains this disconnect between forecasts of AI progress and economic indicators? The following themes appear in the rationales economists wrote to accompany their GDP forecasts.

Uneven and time-lagged diffusion

Some economists argued that AI productivity gains would not be evenly distributed across all sectors, particularly where human labor is a bottleneck. Others pointed out that with other general-purpose technologies (electrification, automobiles, personal computers), there were multi-decade lags between widespread implementation and productivity improvements. Part of this delay is attributed to a shift in capital away from labor and toward compute, data centers, APIs, and so on, which would not manifest as an increase in GDP until productivity improvements set in. Economists agreed that AI boosts to scientific research and development had the most potential to contribute to GDP growth.

Geopolitical, structural, and demographic headwinds

Some economists expected demographic decline and geopolitical instability to offset some of the GDP boost from AI progress. They cited trade wars, global conflicts, climate change, economic competition in a multipolar world, political instability, an aging population, and declining immigration as factors that would limit GDP growth up to 2050. A number of economists were skeptical that double-digit GDP growth could be sustained in a developed economy, pointing out that there is no historical precedent for this.

Bottlenecks on AI infrastructure

Some economists argued that constraints on energy and chip supply, data center build times, and other commodities put a cap on the upper limit of GDP growth. Power plants, data centers, and robotic factory capacity take many years to build, and businesses take time to integrate new tools, which all limit the potential for explosive economic growth.

Possibility of extreme outcomes

Some economists argued that tail risks, particularly in a rapid AI progress scenario, raised the possibility of extreme outcomes that would impact GDP growth. These included existential risks from AI, societal unrest or collapse, and war. Others pointed out that the rapid scenario may lead to large numbers of people leaving the workforce (or even a post-work economy), which would likely negatively impact real GDP.

As you can see in Figure 4 below, economists' forecasts for GDP growth in 2030 are higher than most comparable forecasts used by government agencies and the private sector, which tend to be closer to 2%

3. Economists place a 14% chance on a rapid AI progress world where uncertainty about economic impacts is much higher

The rapid progress scenario describes a world in which AI in 2030 is extremely advanced compared to present-day capabilities: autonomously creating game-changing technologies, outperforming all software engineers, producing award-winning songs and novels, acting at the level of highly competent CEOs, and enabling robots that can navigate all homes with ease to perform tasks faster and more reliably than humans and perform all tasks in autonomous factories.

Economists assigned an average 14% probability to this scenario most closely matching the real world in 2030. When asked to forecast economic indicators in 2030 and 2050 in such a world, the median economist predicted substantial changes that parallel some of the most productive periods in recent U.S. history.

A small increase in annualized GDP growth leads to large differences in the size of the U.S. economy over time. The 3.5% growth rate in the rapid scenario would produce a real GDP of $54.7 trillion in 2050, 25% larger than the $43.7 trillion in the unconditional scenario (2.5% annualized growth). This is roughly equivalent to the difference in GDP (economic output) between 2016 and today.

In a rapid AI progress scenario, economists forecast a drop in labor force participation from its 2025 baseline of 62.6% to 59.1% in 2030 and 55% in 2050, with roughly half of this decline—equivalent to 10 million jobs—attributable to AI. By contrast, the Congressional Budget Office projects the labor force participation rate to be roughly 62.2% in 2050, close to current levels (CBO, 2026).

Forecasts for wealth inequality are particularly stark. In the 2030 rapid scenario, the median respondent expects 75% of U.S. national wealth to be held by the richest 10% of the population. In 2050, respondents expect this to hit 80%.

HISTORICAL CONTEXT

The above forecasts describe an economy that is substantially different from the U.S. today.

GDP: The U.S. has not seen a 10-year period with greater than 3.5% growth since 1992-2001 (FRED).

TFP: The U.S. saw average annual total factor productivity growth rates of around 2% in the post-WWII period (Shackleton, 2013, p. 5). The late 1990s “IT boom” was associated with a 0.6 percentage point jump in TFP (Jorgenson and Stiroh, 2000, p. 6).

LFPR: The annual labor force participation rate was last below 60% in the 1960s, when women’s participation in the labor force was much lower than it is today. The Congressional Budget Office projects the labor force participation rate to be roughly 62.2% in 2050, close to current levels (CBO, 2026).

Wealth inequality: The share of U.S. wealth owned by the top 10% wealthiest households last reached 80% in 1939 (World Inequality Database).

Unemployment: The forecasted unemployment rate of 6% is still below the rates of >7% seen in the years following the 2008 financial crisis (FRED).

Economists have substantial uncertainty about economic outcomes in the rapid scenario: their 80% credible intervals for GDP growth in 2050 range from 1.0% annualized GDP growth to 7.0%.1 The equivalent 80% credible interval for the labor force participation rate was between 45.5% and 61%.

Understanding the views of respondents who predict significant GDP growth

We filtered respondents by their unconditional forecasts for GDP growth by 2030. Respondents whose forecasts fell in the top 20% of this distribution tended to focus on the following themes in their rationales:

AI-driven scientific breakthroughs

Some respondents expect rapid progress to lead to major breakthroughs in science, technology, and governance that enable unprecedented levels of productivity before 2030. These respondents expect scientific gains to strongly outweigh any demographic, political, or structural headwinds.

Radically altered economics/geopolitics

Respondents argued that in a radically changed world, U.S. GDP may no longer be a relevant economic indicator. Some pointed out that it is unclear whether economic expansion to outer space would be counted as U.S. GDP, while others predicted that in such a world, the U.S. may no longer exist.

4. Economists support job retraining as a response to AI's economic impacts

Respondents forecasted the impact of six policy proposals implemented in isolation in both an unconditional and a rapid progress scenario. They were: retraining support, modernized unemployment insurance, universal basic income, a Manhattan Project for AI, a compute tax, and a job guarantee program.

Economists’ support for implementation was highest for a job retraining policy (71.8%), which they predicted would have a modestly positive effect on both GDP (+0.2 p.p.) and LFPR (+1 p.p.) in a rapid progress scenario in 2030. A +1 p.p. increase in LFPR would amount to an additional 2.76 million Americans in the workforce. Economists gave only a 10% chance that this policy would be implemented.

A Manhattan Project for AI, defined as a deployment of roughly $120 billion in annual federal spending to accelerate AI research and infrastructure funded by a 0.7% VAT, and a retraining support program for unemployed people leaving jobs at high risk of automation, had the joint largest forecast impact on GDP growth (+0.2 p.p.) by 2030. The retraining support program received support from 71.8% of economists, with 8.3% unsure and 19.9% opposed to implementation.

Economists viewed universal basic income the most negatively. It was the only policy to be associated with a forecast of a substantial projected LFPR decline (-2.0 p.p.). Over a third of economists (38.2%) opposed it, although the job guarantee program and the tax on compute both garnered even more disapproval, with 56.8% and 45.9% of economists rejecting the policies, respectively.

Economists and the general public disagreed on their support for some policies. 47.9% of the public respondents supported a universal basic income, compared to 37.4% of economists. 57.1% of the public supported a job guarantee program, compared to 13.7% of economists.

Next Steps

In this study, we elicited precise, falsifiable predictions about the future of the U.S. economy from leading economists, AI industry professionals, AI policy experts, top generalist forecasters, and the general public. Over time, we will be able to understand whose predictions about the impact of AI are most accurate. Will people who predicted rapid AI progress in 2030 be closest to the truth? Will AI experts prove more accurate than economists in their predictions for the U.S. economy in 2030?

As AI’s early economic effects become clear, this study can provide a roadmap for whose predictions we should trust more, and which policy responses may be most impactful based on the trajectory we are on.

Study Details

Economists were sampled from three target populations: economists working on AI-related topics, those working on the impact of technological change, and well-known economists, including Nobel Prize winners and members of the Clark Center U.S. Economic Experts panel. AI experts were sampled from frontier and highly funded AI companies, leading AI institutions, and U.S. think tanks focused on AI policy. Forecasters are designated “superforecasters” if they were either in the top 2% of the accuracy distribution in a given year of the Intelligence Advanced Research Projects Activity (IARPA) Aggregative Contingent Estimation (ACE) tournament or a highly accurate performer on Good Judgment Open. Responses were reweighted to the sample population to correct for non-response bias. The survey ran from mid-October 2025 to the end of February 2026.

For more information, read the full report.

TL;DR: The Brier score is hard to interpret because its units are squared probabilities. The Brier Index, defined as (1 − √Brier score) × 100%, fixes this problem, giving a 0–100% scale where 50% = always predicting 50% and 100% = perfect foresight. On ForecastBench, superforecasters score 70.6% and the best LLMs (CassiAI’s ensemble_2_crowdadj and xAI’s Grok 4.20, Preview) score 67.9%. We’ve adopted the Brier Index as the primary metric on the ForecastBench leaderboard.

To evaluate forecasting accuracy, we typically use the Brier score.1

The Brier score has many desirable features. Most importantly, it’s strictly proper, meaning that to achieve the highest possible score, it’s optimal to report your true probability forecast (i.e., it’s not optimal to lie).2

However, the Brier score has a major drawback: it’s difficult to interpret and communicate.

In this post, we untangle the problem, propose a simple metric—the Brier Index—that fixes it, and show what the solution looks like in practice on ForecastBench, our AI forecasting benchmark.

1. Why Brier Scores Are Confusing

Currently, superforecasters—highly accurate forecasters identified through prior geopolitical forecasting tournaments—lead the ForecastBench leaderboard with a Brier score of 0.086. The best LLMs, CassiAI’s ensemble_2_crowdadj and xAI’s Grok 4.20 (Preview), sit at #2 with 0.103.

Two natural questions arise:

• What does a Brier score of 0.086 mean? Are the superforecasters highly accurate? Are they not that great?

• How large is the difference between 0.086 (superforecasters) and 0.103 (best LLMs)? Is the 0.017 Brier-score gap “large” or “small”?

Both questions are difficult to answer because the Brier score has uninterpretable units. The Brier score is the mean squared error (MSE) for binary events: if f is your forecast and o is the outcome (either 0 or 1), then the Brier score is given by b = (f − o)². Averaged over N predictions, you get the mean Brier score.

The issue is the same one that plagues MSE generally: the units are squared. A Brier score of 0.086 means “0.086 squared probability units from the truth.”

Confused? You’re not alone.

2. A More Interpretable Metric: The Brier Index

Step 1: Take the Square Root

The first fix is straightforward: take the square root of the Brier score to recover the natural units (i.e., calculate the root mean squared error, RMSE). This is analogous to converting variance (hard to interpret) into a standard deviation (same units as the data).

The square root of the superforecasters’ Brier score is √0.086 ≈ 0.29. But what does an RMSE of 0.29 mean, precisely?

It’s tempting to say “on average, superforecasters are 0.29 probability units away from the truth.” Unfortunately, that’s not quite right. This interpretation is correct for the mean absolute error (MAE) but not RMSE.34

The precise interpretation of RMSE is slightly more indirect: an RMSE of 0.29 means that a hypothetical forecaster who always predicts 29% on questions that resolve to 0, and always predicts 71% on questions that resolve to 1, would achieve this exact same score.5 That’s a simplification: real forecasters don’t make constant-magnitude errors on every question. However, it provides a useful reference point.

Step 2: Flip the Scale

Taking the square root solves the “squared units” problem. However, as with the Brier score, lower values are better for RMSE. In our experience, decision-makers find this counterintuitive; for most metrics decision-makers encounter, a higher score indicates a better result.

Our proposal—the Brier Index—is defined as:

Brier Index = (1 − √Brier score) × 100%

The scale works as follows:

• 100%: perfect foresight (assigning probability 1 to events that happen and 0 to events that don’t).

• 50%: equivalent to always predicting 50%.

• 0%: maximally wrong (assigning probability 1 to events that don’t happen and 0 to events that do).

We can also connect the Brier Index back to the hypothetical forecaster interpretation from Step 1. The superforecasters’ Brier score of 0.086 corresponds to a Brier Index of 70.6%. This is the same score achieved by a hypothetical forecaster who predicts 70.6% on events that resolve to 1, and 29.4% on events that resolve to 0.

An important caveat: Brier Index is not a “percentage-correct” metric. A Brier Index of 70.6% does not mean the forecaster “got 70.6% of questions right.” Instead, it measures how far probability forecasts are from the truth, rescaled to a 0–100% range. For example, a forecaster who predicts 51% on events that happen and 49% on events that don’t would get the direction right every time, yielding a high percentage correct but a mediocre Brier Index (just 51%).

3. Brier Index on ForecastBench

Let’s return to our motivating questions.

Applying the Brier Index to ForecastBench:

Superforecasters achieve a Brier Index of 70.6%, and the gap to the best LLM is 2.7 percentage points.

Now, how large is that 2.7 percentage point gap?

One way to put this in context is to look at the progress of state-of-the-art (SOTA) LLM forecasting performance:

From Claude 3.5 Sonnet (October 2024) to Grok 4.20 (Preview, October 2025), the Brier Index increased by 2.4 percentage points in one year. The current gap between superforecasters and the best LLMs represents a little over one year of LLM progress. That gap is substantial, but if the current rate of LLM progress continues, LLMs may reach superforecaster-level performance relatively soon.

Switching from Brier score to Brier Index shifts the extrapolated LLM–superforecaster parity date slightly:

• Extrapolated Brier score parity (previous): March 2027 (95% CI: February 2026 – January 2028).

• Extrapolated Brier Index parity (updated): May 2027 (95% CI: April 2026 – May 2028).

This shift is a modeling artifact of how we extrapolate trends. In both cases, we extrapolate the parity date by fitting a linear trend over time. But a straight-line trend is not preserved under nonlinear transformations. Because the Brier Index is a monotone, convex transformation of the Brier score, the same absolute improvement in Brier score translates into a progressively larger increase in Brier Index at high accuracy levels:

Hence, a trajectory that looks linear in Brier score will generally look curved (accelerating) in Brier Index. While the leaderboard ranking is unchanged by the transformation, the intersection point implied by a simple linear extrapolation—and thus the parity date—shifts slightly.

A note on the extrapolation: These trend lines are fitted to a small number of SOTA models over roughly two years of data. The 95% confidence intervals reflect uncertainty in the linear fit, not uncertainty about whether LLM progress will in fact remain linear. We currently don’t have strong evidence that progress is “more linear” in Brier score than in Brier Index (or vice versa). For this reason, we keep the extrapolation deliberately simple and report parity under the same linear-trend approach as previously, using the Brier Index as our primary, more interpretable metric.

4. Updating the ForecastBench Leaderboard

We’ve adopted the Brier Index as the primary metric on the ForecastBench leaderboard. This change doesn’t affect the rankings. The Brier Index is a monotonic transformation of the Brier score, so the ordering is identical. However, we expect that the headline number will be easier to understand. We have also updated the linear trend extrapolation graph on the “Explore” page. The underlying Brier scores remain available on the "Datasets" page.

Addendum: Properness of the Brier Index (2026-03-05)

We made a technical error in the initial post. Specifically, we claimed that “Because the Brier Index is a monotonic transformation of the Brier score, it is also strictly proper.” This claim is incorrect: properness is preserved under affine transformations, but monotone nonlinear transformations need not preserve properness. For the special case of a single forecasting question (N = 1), the Brier Index equals

which is an affine function of the absolute error, |f - o|. As we discussed in the original post, the MAE is not a proper scoring rule for binary events (see Footnote 3). Hence, the Brier Index is not proper for N = 1. We have updated the original post to remove the properness claim.

However, as the number of questions grows, the incentive to deviate from truthful reporting shrinks under some baseline assumptions. As shown in the graph below (source code), already for N = 10 independent questions, the forecast that maximizes the Brier Index is very close to truthful reporting (in this example, we set the true probability to 70% on each question). While many forecasting questions are not independent, we specifically designed ForecastBench to cover many domains and time horizons, so that effectively the benchmark has many independent questions. In addition, ForecastBench evaluates each model on hundreds of questions, further reducing the distortion.

The core issue that prevents the Brier Index from being strictly proper is that the expectation operator commutes only with affine transformations: E[g(X)] ≠ g(E[X]) in general for nonlinear g. Denote the Brier score from N forecasts by B = B(f₁ , f₂,..., fₙ). A forecaster minimizing the expected Brier score solves E[B] → min. Denote the Brier Index transformation by g, so that the Brier Index is g(B). To maximize the Brier Index, the forecaster instead solves E[g(B)] → max. If B were non-random, the two problems would have the same solution. However, B is random (because the outcomes are random), and since g is nonlinear, the two optimization problems can differ.

With a large number of questions, however, B converges to its expected value, say, μ due to the Law of Large Numbers (under appropriate conditions). By the Continuous Mapping Theorem, g(B) then converges to g(μ). This suggests that, for large N, the forecasts that minimize the expected Brier score should also approximately maximize the expected Brier Index, consistent with the numerical results above.

Finally, this correction does not affect the ForecastBench leaderboard ranking. We apply the Brier Index transformation only after calculating the average Brier score across all questions, so the ordering of models by the Brier score is exactly the same as the ordering of models by the Brier index. Had we instead applied the transformation at the question level—i.e., computing (1 − |f − o|) * 100% for each question and then averaging—the result would be an affine transformation of the Mean Absolute Error, averaged across all questions. The ranking produced by this metric would generally deviate from the Brier score ranking.

We thank our readers (Ege Erdil and Jérémy Andréoletti) for their critical feedback in pointing out the error.

Every month, LEAP tracks the views of top AI scientists, industry leaders, policy researchers, economists, and high-performing forecasters on the trajectory of the development and use of artificial intelligence. Earlier LEAP waves covered AI research and development, broad adoption of AI, using AI for science, and more.

This post covers key highlights from the Wave 5 LEAP survey that we conducted between January 12 and February 2, 2026. Full details of this wave, including question details and analysis of rationales, are available here.

Insight 1: Experts and superforecasters expect the performance gap between U.S. and Chinese AI models to narrow by 2031, with parity anticipated by 2041

Experts and superforecasters forecast that Chinese AI systems will roughly match American ones on benchmarks by the end of 2040. At the time this survey was published (January 2026), the top Chinese models lagged behind U.S. models by nine points on the Epoch Capabilities Index. At the time of writing (February 2026), this gap is six points. The median expert expects this gap to narrow to five points by the end of 2030 and to zero by the end of 2040.1

Insight 2: Experts predict that Taiwan’s dominance in leading-edge chip manufacturing will erode, with the U.S. and China becoming significant leading-edge chip manufacturers

Experts project Taiwan’s global share of leading-edge chip production to fall from over 90 percent today to roughly 71 percent by 2030 and 46 percent by 2040, with the U.S. and China each capturing around 20 percent of the market in 2040.

Forecasters cited national security concerns as the major factor driving increased U.S. leading-edge chip manufacturing, arguing that onshore chip production could serve as insurance against blockade, conflict, or disruption. Forecasters also expected China to scale up its leading-edge chip manufacturing abilities, eventually reaching technical parity with the U.S. and Taiwan.

Insight 3: Experts expect a NATO member state to use autonomous cyber weapons by 2041, and that U.S.–China agreement on military AI is unlikely by this date

The median expert forecasts that NATO or a member state will publicly authorize fully autonomous offensive cyber operations in the coming decades (a median forecast of 2041), with superforecasters predicting this will occur slightly earlier, in 2038. A quarter of experts expect this policy shift to occur in 2035 or earlier.

In addition, experts gave only a 16 percent probability that the U.S. and China would sign an agreement on military AI usage by the end of 2030, but gave a 40 percent probability that such an agreement would be signed by the end of 2040. The median superforecaster agreed with these predictions, although there was widespread disagreement within each group about the likelihood of such an agreement.

Insight 4: AI systems are expected to surpass top human forecasters within the next few years, but the significance of that achievement is debated

Superforecasters themselves are the most bullish group on automated forecasting progress, with the median superforecaster predicting AI systems will beat their ForecastBench benchmark by 2028, which is earlier than both the median expert (2030) and the median public (2033) forecast.

Insight 5: Experts expect large (>$1 billion) regulatory fines on AI companies to total a cumulative $4 billion by 2030

The median expert predicts that the U.S. and/or EU would impose $1 billion in cumulative large-scale regulatory fines on AI companies by 2027, rising to $4 billion by 2030, and $10 billion by 2040. The public consistently expected higher fines, while superforecasters gave the most conservative predictions, estimating $8 billion in fines by 2040.

Forecasters cite AI-driven financial fraud, disinformation campaigns, sexual victimization through “deepfakes,” and copyright violations as potential trigger points for fines, but were skeptical that major fines would be levied by the end of 2027, pointing out that there have been few examples of such fines so far.

Find out more about LEAP

To read more about Wave 5 of LEAP, including further questions and rationale analysis, visit the LEAP website. If you have a suggestion for a question you’d like to see covered in a future LEAP wave, please submit a question through our online form.

Summary

Active Site (a nonprofit that evaluates AI-driven biology, formerly know as Panoplia Laboratories) ran a randomized controlled trial to test how well novices do at various molecular biology tasks and whether access to LLMs improves (“uplifts”) performance. To help contextualize these results, we surveyed 30 superforecasters, 30 biosecurity experts, and 17 virology experts. We asked respondents to predict the RCT results and how these would affect overall levels of biorisk. We then repeated this survey after forecasters learned the true RCT results.

The RCT results surprised most respondents: LLMs were less useful than expected. The median forecaster overestimated how many participants in the non-LLM (internet-only) group would succeed at three core lab tasks (median forecast 12% versus actual 7%) and how much LLMs would improve performance in mid-2025 (median forecast 27% versus actual 5%). Superforecasters were more accurate in their forecasts than biosecurity or virology experts. Most forecasters slightly revised down their predictions of a human-caused large-scale outbreak in 2028 after seeing these results.1

The RCT study design appears to capture an important dimension of biorisk. Forecasters predicted that if the study were to be repeated in the future and were to find a 5x LLM uplift, then that would be associated with a 2x increase in the likelihood of a human-caused outbreak. Lastly, additional information seemed to build consensus: once informed of Active Site’s mid-2025 RCT results and each other’s forecasts, respondents’ biorisk forecasts converged into a notably narrower range.

Introduction

In August 2025, Active Site reached out to the Forecasting Research Institute (FRI) to tell us about a randomized controlled trial (RCT) it was working on. The RCT tested how access to large language models (LLMs) affected the ability of novices to do molecular biology work in a wet lab.

Active Site’s RCT sought to better inform such discussions and help draw out specific biosecurity implications. Active Site wanted to obtain forecasts from superforecasters and experts about the results of its ongoing RCT, as well as to pre-register how their forecasts of biorisk would change depending on different results of the RCT. To answer these questions, we surveyed 30 superforecasters (people with a track record of accurate forecasts), 30 biosecurity experts, and 17 virology experts.

Since RCTs provide a gold standard measurement of causal effects, the Active Site study would produce a plausibly unbiased estimate of how much LLMs caused improvements in novices’ ability to perform biological laboratory processes in mid-2025. But how much do such RCT results matter for biosecurity? Answering this question requires making judgments about various factors outside the scope of Active Site’s RCT. Our study quantified and aggregated such judgments from superforecasters and experts.

In this post, we describe respondents’ predictions of the Active Site RCT, how these forecasts (and their accuracy) relate to predictions of biorisk, and how forecasters updated their thinking upon learning the true RCT results.

Background on the Active Site RCT results

Active Site’s RCT tested novices’ ability to successfully complete wet lab tasks involved in reconstructing a virus with minimal instruction. The three most important “core” tasks were:

1. Cell culture: reconstituting human cells and maintaining them over several generations

2. Molecular cloning: assembling pieces of DNA into a plasmid

3. Virus production: producing adeno-associated virus (AAV) by inserting DNA into human cells

The study included 153 participants and lasted eight weeks; participants were provided with all necessary laboratory materials but given minimal instructions. The RCT tested novices’ skills on these tasks in two experimental conditions, to which participants were randomly assigned. In the first group, participants had access to the internet but not LLMs. In the second group, they had access to both the internet and several LLMs available during summer 2025, with some relevant models having their biosecurity safeguards removed for the experiment.

In the internet-only condition, 6.6% of participants completed all three core tasks within the allotted time. In the group with access to LLMs and the internet, 5.2% of participants completed all three tasks. The difference in success rates between the two groups was not statistically significant.

How accurate were respondents at predicting the RCT results?

Before the results were in, we asked forecasters to predict the main outcome of the RCT: the percentage of novices in each condition that would be able to complete all three core tasks.

Most people overestimated how well the novices would do, both with and without access to LLMs. Across the full sample, the median forecaster predicted a success rate of 12% in the internet-only group (versus actual 6.6%) and 27% in the LLM plus internet group (versus actual 5.2%).

Generally, superforecasters were more accurate at predicting the RCT in absolute terms than experts, particularly compared to virologists. The median virologist expert predicted RCT outcomes of 30% in the internet-only group and 40% in the LLM plus internet group. Superforecasters were much closer to the actual RCT outcomes, with median predictions of 8% and 16%, though this still overestimated the effect of LLMs. The median biosecurity expert’s predictions fell between these two groups.

Readers should note that our estimates of these between-group differences are somewhat noisy given the relatively small sample sizes of groups of forecasters.2 Still, analysis suggests that the ordering of the forecaster subgroups is robust.3

How did the RCT predictions and results influence views on biorisk?

We were also interested in how predictions of the RCT might relate to views on biorisk. For example, there has been significant expert disagreement about the importance of tacit knowledge in biology and the extent to which a novice can complete wet lab tasks. We wanted to see if more accurate forecasts on the RCT results are associated with different views on biorisk.

To begin, we asked respondents to forecast the risk of a human-caused disease outbreak that caused at least 100,000 deaths or $1 trillion in damages in 2028. Although there was a large range of opinion among forecasters, there were only small differences in forecasts when grouping participants by their RCT prediction accuracy. The median forecaster in the full sample gave a 0.3% probability of such an outbreak (25th percentile: 0.005%, 75th percentile: 1%), while among those who forecasted the RCT result most accurately, the median was 0.18% (25th percentile: 0.01%, 75th percentile: 0.725%).4

Next, we wanted to know if people would change their views on biorisk after learning the true RCT results, especially since so many overestimated the results. Three months after the initial survey, 65 out of the 77 original forecasters completed a follow-up survey.

When they learned the results of the RCT, 26 participants reduced their forecast of the risk of a large-scale human-caused disease outbreak, 30 left it unchanged, and nine increased it. The median across forecasters fell slightly from 0.3%5 to 0.2%. Among the 15 respondents who forecasted the RCT result most accurately (see footnote two), six decreased their forecast, eight left it unchanged, and one increased it, yielding another slight decrease in the median outbreak risk probability from 0.18% in the original survey to 0.10% in the resurvey.

How would future RCTs inform views on biorisk?

While the Active Site result did not find LLMs to have a statistically significant “uplift” in mid-2025, LLMs are becoming more capable, and we were interested in how people would react if future work discovered a larger effect.

We asked our sample to imagine a repeat of the Active Site RCT in 2026 with newer models and to assume that the future RCT produces various hypothetical results.6 Figure 3 shows how their estimates of biorisk would change given three hypothetical conditions of future LLM uplift—including a condition where we continue to see no large uplift.

The resulting risk curves sloped upwards for all forecaster groups, indicating that participants perceived the RCT as a relevant signal of increased outbreak risk. For example, when assuming that a future RCT now finds that 25% of the LLM group succeed at the core tasks, while still only 5% in the internet-only group do, (i.e., a five-fold biological skill uplift due to LLMs), the median outbreak risk forecast increased from 0.3% to 0.7% (i.e., a two-fold risk increase relative to before).7 Notably, we saw substantially narrower interquartile ranges on the resurvey, suggesting that seeing ground truth from the RCT and each other’s predictions led to greater consensus among forecasters about biorisk.

How likely is it that we might see such LLM-enabled wet lab skill uplift? We asked forecasters to predict the results of future repetitions of the RCT on both the initial survey and the resurvey. Most participants adjusted their LLM-group forecasts downwards after learning the 2025 results but still expected LLM success rates to increase over time. They predicted a median LLM-group success rate of 7% (25th percentile: 6%, 75th percentile: 7.5%) if the RCT were repeated in 2026 which increased incrementally to 11% (25th percentile: 8%, 75th percentile: 15%) if it were repeated in 2030. If such an RCT is repeated, we will be interested to see how these forecasts compare—and if this might differ from how we have seen benchmark performance outpacing most forecasters’ near-term expectations.

Key takeaways

• Forecasters overestimated the assistance that mid-2025 LLMs would provide to novices attempting to complete three wet lab tasks that are necessary to reconstruct a virus.

  • The median forecaster predicted a 12% success rate for the internet-only condition and 27% for LLM and internet access, compared to the actual results of 6.6% and 5.2%, respectively.

• Those with a strong forecasting track record made more accurate forecasts than domain experts.

  • As a group, superforecasters produced the most accurate RCT forecasts. Despite deep technical knowledge of wet lab work, virologists performed the worst out of the three groups.

  • This may suggest that forecasting expertise is more helpful than domain expertise at predicting RCT results, but we cannot draw definitive conclusions from a single, somewhat noisy result.

• The median forecaster predicted a baseline risk of 0.3% for a large-scale human-caused outbreak in 2028 in the original survey, and updated to 0.2% after the results of the RCT were revealed.

  • Among those who forecasted the RCT result most accurately, the outbreak risk forecast was 0.18% on the original survey and 0.10% on the resurvey.

• Risk estimates increased meaningfully when forecasters assumed higher novice wet lab capabilities than those found in the Active Site RCT.

  • When asked to assume that LLM access leads to five times higher success at wet lab tasks, respondents’ median outbreak risk forecast increased two-fold.

• Forecasters see a relationship between the RCT outcomes and outbreak risk, though it is weaker than the connection between RCT outcomes and virus rescue.

  • Forecasts of outbreak risk increased as forecasters assumed higher hypothetical RCT success rates. When asked to assume that the RCT showed that 50% of participants succeeded at core lab tasks, the median forecaster gave a 1.3% outbreak risk forecast, compared to 0.3% when only 5% succeeded at core lab tasks.

  • Median outbreak risk forecasts decreased after the RCT success rates (which were lower than most forecasters expected) were revealed, with the median risk forecast changing from 0.3% to 0.2% (among the sample that took part in both phases of the survey). These changes were notably smaller than the decrease applied to forecasts of novices’ capability to complete a virus rescue protocol.

The result? External submissions now hold #2 and #3 on our leaderboard, outperforming all our baseline LLM configurations. Superforecasters still hold #1.

Here are the current standings (lower Brier scores are better):

Two external submissions—xAI’s Grok 4.20 (Preview) and Cassi’s ensemble_2_crowdadj—are tied at #2, both ahead of our own runs of GPT-5, o3, and Claude Sonnet 4.5. Superforecasters are #1, leading state-of-the-art LLMs by 0.017 Brier points.

How large is the gap between LLMs and superforecasters? The current 0.017-point gap represents about one year of LLM progress. For comparison, Claude 3.5 Sonnet (released October 2024) achieved a Brier score of 0.117; a year later, Grok 4.20 (Preview, run in October 2025) achieved 0.102—an improvement of 0.015 points. If this trend continues, we extrapolate LLM-superforecaster parity in November 2026 (95% CI: January 2026 – November 2027). Dataset question parity is estimated for June 2026, and market question parity for August 2026.

We use difficulty-adjusted Brier scores to compare forecasters even when they predicted on non-overlapping question sets. For instance, our initial superforecaster sample and the latest models on ForecastBench have zero overlapping questions. For details, see our technical report.

To learn more, we asked the top-performing external teams to share their methods. Here’s what the two top teams told us about how they approached ForecastBench.

xAI took a minimal approach: give the model the question, provide access to standard Grok tools (X search, web search, Python REPL), generate eight forecasts, and average them. Notably, the team used an early preview version of Grok 4.20 and expects the full 4.20 release to perform even better.

In contrast, Cassi used a multi-stage pipeline built around retrieval and ensembling. The system generates sub-questions and search queries, retrieves up-to-date context via Tavily, and filters for relevance and recency. It then generates forecasts using an ensemble of models (o3 and GPT-5 in this run), with another LLM analyzing the reasoning to produce a final forecast.

For questions drawn from prediction platforms, Cassi adds an additional step: comparing the final forecast to the market price and, if they disagree significantly, giving an LLM a chance to review and adjust. This “crowd adjustment” helps: Cassi’s ensemble_2_crowdadj outperforms its ensemble_2 by nearly 0.01 Brier points. Cassi notes they do not use post-forecast calibration, finding that current LLMs already produce broadly well-calibrated forecasts. For more on Cassi’s work, see the company’s website and Substack.

More models are joining the leaderboard soon. We add models 50 days after their first forecast to ensure enough questions have resolved for stable rankings. In late February, GPT-5.2 and Gemini 3 Flash will appear on the leaderboard, while GLM-4.7 will become visible in late March.

Think you can do better? The Tournament leaderboard is open for external submissions. See our wiki for instructions. Explore the leaderboards at forecastbench.org and download full datasets at forecastbench.org/datasets.

Every month, LEAP tracks the views of top AI scientists, industry leaders, policy researchers, economists, and superforecasters on the trajectory of the development and use of artificial intelligence. As our chief scientist, Phil Tetlock, wrote in a recent essay for Bloomberg, the aim of LEAP is to provide the debate regarding the future of AI with precise, policy-relevant predictions rather than overconfident proclamations.

This post covers key highlights from the Wave 4 LEAP survey that we conducted between November 20 and December 20, 2025. Full details of this wave, including question details and analysis of rationales, are available here. For more information on earlier LEAP waves, read our launch blog or white paper.

Insight 1: AI model benchmark performance outpaced most forecasters’ near-term expectations

AI performance on hard coding tasks is a useful indicator of potential capability increases in self-improving AI R&D: where models take an active role in improving AI itself.

LEAP panelists gave their predictions of state-of-the-art (SOTA) accuracy on LiveCodeBench Pro (Hard), which is a benchmark that tracks performance on tough programming tasks. The median expert in our sample predicted SOTA accuracy of 14% in 2026 and an increase in accuracy to 33% by the end of 2030.

Recent benchmark performances suggest that the median expert is substantially underestimating the progress of development of SOTA AI systems. Since the survey closed, OpenAI’s GPT-5.2 achieved 33% accuracy on the Q3 2025 set of problems on LiveCodeBench Pro. While the set of questions on the platform is dynamic, and future performance on question sets may be lower than 33%, the resolution of this question will be at least 33%.

Experts’ underestimation of benchmark progress is also evident in earlier work from FRI. In a paper released in September 2025, we found that domain experts and superforecasters underestimated AI progress across the MATH, MMLU, and QuaLITY benchmarks.

Despite the low median forecasts, LEAP panelists still substantially disagreed with each other. For example, the top quartile of superforecasters predict at least 60% SOTA accuracy on LiveCodeBench Pro (Hard) by 2030. We are excited to identify the forecasters who were most accurate on questions like this to see what they believe about other topics.

Insight 2: Experts predict continual low levels of junior hiring, and some rationales point to AI being a driver of this outcome

We asked panelists to forecast the percentage of new hires at the top 15 tech companies with up to one year of experience in 2026, 2030, and 2040. The median expert expects this share to remain close to its current low of seven percent through 2040. This predicted share is well short of the 15% of hires seen in 2019 and 2023.

Experts and superforecasters give largely similar forecasts, but superforecasters expect stronger entry-level hiring by the end of 2040 than experts: superforecasters predict that 10% of new hires will be entry-level in the median end-of-2040 scenario, compared to the median expert’s central forecast of 7.8%.

Some respondents pointed to AI as a driver of continued low rates of entry-level hiring. For example, one forecaster wrote that “[t]he risk for entry-level hiring skews to the low end of the distribution as it will be plausible to use a combination of contractor-outsourcing and LLM/AI tools to skew a company’s workforce towards mid-level employees.” Other forecasters pointed to macroeconomic factors that may influence the share of entry-level hires rather than AI progress.

Insight 3: Experts expect AI and technology companies to become more valuable per employee

The median expert predicts that by the end of 2030, the valuation-per-employee for top AI companies will grow to $20 million; this is about a 40 percent increase over the Q3 2025 baseline of $14.2 million. By the end of 2040, the median expert’s central forecast rises to $32 million. However, experts express uncertainty in their forecasts. The median expert assigns a 25% likelihood that the valuation-per-employee ratio in 2026 will fall to $13 million or less and a further 25% likelihood that the ratio will increase to $21 million or more.

We also asked forecasters to predict the calendar year in which at least three companies with five or fewer full-time employees first report individual valuations of $10 billion (in 2025 USD) or more. The median expert gives a 50% chance of resolution by the end of 2032, the public by the end of 2031, and superforecasters by the end of 2033. The median forecaster in all three groups gives a 5% chance that three such companies will exist by the end of 2027.

Some public figures have opined on a similar question. OpenAI CEO Sam Altman has publicly stated that “[w]e’re going to see ten-person, billion-dollar companies pretty soon.” Economist Tyler Cowen, in an interview with Altman, predicted “I think you’ll have billion-dollar companies run by two or three people with AIs, I don’t know, two-and-a-half years.” Altman responded, “I agree on all of those counts. I think the AI can do it sooner than that.”

Insight 4: Experts forecast major increases in AI data center buildout

We asked our panelists to predict the total installed capacity (in gigawatts) of hyperscale data centers operating globally by 2026, 2027, and 2030. We follow the International Energy Agency’s (IEA) definition of hyperscale data centers as having 100 MW of capacity—that is, they consume as much electricity annually as 100,000 households. The IEA estimates that the total installed global capacity of hyperscale data centers in 2024 reached 36 GW.

The median expert’s central forecast of 92 GW of installed capacity by the end of 2030 is a 2.6x increase from the 2024 level and aligns with IEA projections of 62–108 GW of installed capacity by the end of 2030. Other groups give slightly lower forecasts than experts: the public predicts 83 GW and superforecasters predict 86 GW. Experts’ and superforecasters’ central forecasts for 2030, however, are statistically indistinguishable.

Insight 5: Experts and superforecasters tend to forecast similar levels of progress, but when the two groups disagree, experts tend to predict more progress than superforecasters

As we observed in previous LEAP waves, experts and superforecasters tend to forecast similar levels of progress across most questions. However, even when forecasters agree directionally on the impacts of AI, written rationales reveal deep disagreements about the reasons for their forecasts. Some forecasters argue that high valuations and the rapid buildout of data centers are the result of bubble dynamics. Others argue that improvement in AI agents will lead companies to dramatically increase their revenues without increasing the size of their workforces.

Find out more about LEAP

To read more about Wave 4 of LEAP, including further questions and rationale analysis, visit the LEAP website. If you have a suggestion for a question you’d like to see covered in a future LEAP wave, please submit a question through our online form.