Jump to content

Talk:ChatGPT

Page contents not supported in other languages.
Page semi-protected
From Wikipedia, the free encyclopedia

where we're at with NPOV

ok, if I can sum up, @Uhoj finds the article biased toward the negative. I see that point if you just count up negative sentences. I don't see why comparison to other reviews or articles is germain, let's get this one right. The "what is it used for" is a reasonable proposal to add some positive weight, we just need to figure out how to incorporate in a proper way, and if @Uhoj or any of the rest of us have any ideas, let's do it. I also see how some of the negatives are over emphasised.

So may I propose we all work to try to bring up the positive side to an appropriate weight? Some varient of usage may be a good start. WhaleFarm (talk) 02:47, 19 April 2026 (UTC)[reply]

This appears to be an argument for WP:FALSEBALANCE, so I have to reject the starting premise. We don't assume that the article needs to bring up the 'positive side'. Start from reliable, independent sources. Avoid hype and press releases. If those sources are unflattering, that means the article should be unflattering. That's what WP:NPOV means. If that's not what you're trying to say, than please start over and explain it differently. Grayfell (talk) 02:57, 19 April 2026 (UTC)[reply]
I agree completely with this. My intention with the Applications section was to cover what sources say ChatGPT is good at. I think adding a subsection on leisure with appropriate sources could be warranted and may satisfy all parties? Czarking0 (talk) 00:42, 20 April 2026 (UTC)[reply]
Wikipedia articles often lean more easily toward criticism than praise (which is one of the reasons why we have an article "Criticism of Google" but an article "Praise of Google" would be quickly deleted). Taking into account that tendency of Wikipedia articles, the ChatGPT article seems well-balanced in my view, perhaps even leaning a little on the "too positive" side. A full section doesn't seem warranted, but a paragraph on how people use ChatGPT could be interesting to add to the lead of the "Reception" or the "Applications" section if there is a good independent source. Alenoach (talk) 09:25, 19 April 2026 (UTC)[reply]
The article cited would probably suport something like. "researchers are starting preliminary study on what it's used for......".
There are a lot of good, solid sources for uses of GPT that are appropriate, and missing. Gardening to executive productivity. Comparison of writing quality. Coding help. Use as a better search facility. Stock picking. Contract drafting. Including the pitfalls of each of them. I'm not arguing that these need to be 50% of the balances, but I see the @Uhoj argument that these are underweighted, and that is in conflict with WP:NPOV.
@Alenoach - i don't think neutral is equivalent to "no more critical than other example" WhaleFarm (talk) 12:40, 19 April 2026 (UTC)[reply]
Most of those uses you mention boil down to either using it as a search engine replacement or for generating boilerplate/plagiarism. Most of the sources I have seen for these uses tend to be pretty bad, but a disproportionate number of those are from hunting for spam in Wikipedia articles, so that's a sampling bias. Since you know of good ones, please share them.
These purported uses feel similar to blockchain hype from a few years ago. What does it do? Lots of stuff. What does it do better than existing alternatives? Nobody seems to agree on that. That's a recurring criticism from many sources. So if ChatGPT can be used for something, but per sources it's not necessarily good at doing that, is it still worth mentioning as a "use"? Maybe, but not without a lot of context.
Regardless, all of these uses are still applications. That's what the word "application" means in this context. We do not need to include a preprint preliminary study from OpenAI affiliated researchers. The mere existence of an unreliable source isn't noteworthy in the article. Grayfell (talk) 19:26, 19 April 2026 (UTC)[reply]
https://www.science.org/doi/10.1126/science.adh2586 Was a very early productivity test. https://news.stanford.edu/stories/2026/04/digital-chores-productivity-boost-research
a recent report
https://hbr.org/2025/05/research-gen-ai-makes-people-more-productive-and-less-motivated
very interesting mixed results on motivation WhaleFarm (talk) 20:39, 19 April 2026 (UTC)[reply]
I appreciate that, but I don't think those sources do a good job of explaining what ChatGPT actually does. They talk about how the people using them are effected, or how their productivity changes, but that's very different from what ChatGPT is used for.
For the first one, regarding ChatGPT, 2023 is indeed very early. If secondary sources have discussed this, they might be useful, or they might not be.
The second is from Stanford's news site. These sites exist to promote the work of their schools' researchers, which is reasonable, but does undermine the reliability and impartiality of this source. It is not an independent source for this research.
The third one is a blurb from the authors about their research. The blurb doesn't mention ChatGPT, and while the study itself does use ChatGPT, it's debatable whether or not it is about ChatGPT, as the authors generalize this to be about 'GenAI' instead.
Harvard Business Review has produced a lot of similar work, and I'm not sure if any of it is really appropriate, but it's at least worth considering. For example:
  • Ranganathan, Aruna; Ye, Xingqi Maggie (9 February 2026). "AI Doesn't Reduce Work—It Intensifies It". Harvard Business Review. Retrieved 19 April 2026.
  • Bedard, Julie; Kropp, Matthew; Hsu, Megan; Karaman, Olivia T.; Hawes, Jason; Kellerman, Gabriella Rosen (5 March 2026). "When Using AI Leads to "Brain Fry"". Harvard Business Review. Retrieved 19 April 2026.
While we're at it, how about this one?:
  • Duong, Cong Doanh; Dao, Thanh Tung; Vu, Trong Nghia; Ngo, Thi Viet Nga; Tran, Quang Yen (November 2024). "Compulsive ChatGPT usage, anxiety, burnout, and sleep disturbance: A serial mediation model based on stimulus-organism-response perspective". Acta Psychologica. 251. doi:10.1016/j.actpsy.2024.104622.
Obviously, picking some primary sources while ignoring other is not going to work. We should have a plan before we start throwing even more sources into an already bloated article. Grayfell (talk) 22:11, 19 April 2026 (UTC)[reply]

The NBER paper has been cited 282 times. Does Grayfell wish to continue arguing that it's an unreliable source for these statements? Does anyone else think it's unreliable in that context?

I didn't say we should add some positive weight. I said An article is meant to explain what its subject is. This one falls flat. Rather, much weight is given to criticism. I came to this article to learn something about what ChatGPT is. I didn't find the explanation that I was seeking. Instead, I was disoriented by a litany of outdated news snippets.

Cynicism is a cheap substitute for neutrality. Neutrality is about far more than emotional valence. If ChatGPT is commonly used for practical guidance, but rarely used for ChatGPT-powered church service, then we should probably be saying a lot about practical guidance and very little about ChatGPT-powered church service. If other overviews focus more on what ChatGPT is and less on speculative statements that were made in the news a few months after it was released, then that's probably where we should be focusing as well.

What other reviews or articles say is in fact quite germane. How faithfully we reflect the current consensus is the essence of neutrality. Uhoj (talk) 21:51, 19 April 2026 (UTC)[reply]

Even at a glance, the Google Scholar list includes unreliable sources. So in isolation, the cite count for a preprint is mostly irrelevant to its reliability. AI researchers tend to ignore standard academic publishing practices and issue preprints without ever following through on the publication process. Others in the same field will then often cite those sources. Peer-review and independent editorial oversight are never even attempted. This is not something for the industry to be proud of, it's a red flag, and Wikipedia editors are not obligated to ignore our standards on sourcing just because OpenAI's researchers get a medium amount of attention within their own industry. If this has been properly published, tell me where, and we can evaluate that version.
We agree that the article should summarize what reliable sources say. Primary sources are less useful than secondary sources. That's a big part of the problem. Grayfell (talk) 22:05, 19 April 2026 (UTC)[reply]
WP:PREPRINT are typically not reliable sources. If you think an exception should be made I am open to the argument but saying it has a lot of citations is not a good one.
What did you come to learn about that you did not find? Czarking0 (talk) 00:46, 20 April 2026 (UTC)[reply]
Re: What did you come to learn about that you did not find?:
  • What ChatGPT is
    • How it works
    • What it does
    • If / how it's changed with time
  • Whether it differs in either technology or specialization relative to similar products/projects
  • Historical context: where it came from and how it affected subsequent events
    • Is it a software advance, or hardware?
  • Who the inventors are
  • Why it's such a hot topic in popular discourse
  • What effects it's actually had (not projected to have)
  • What people use it for
    • Whether it's a tool or a toy
Uhoj (talk) 17:03, 21 April 2026 (UTC)[reply]
I think arguably the article does answer some of these questions, but definitely not all of them.
The first subsection is 'Training', which is above even 'Features'. To me, this is a good demonstration of the organizational issues and general bloat of the article. That section includes sources ranging from the NYT to an opinion article in a non-notable journal. This kind of kitchen-sink approach is too common on Wikipedia, but it's much harder to fix for high-profile topics like this one. It also doesn't help that there exists a lot of very strong opinions and a flood of garbage sources.
For the article to be more useful, and to answer fundamental questions like these, some of this bloat should be cut, either with a scalpel or a hatchet. As I said, some of these questions are already answered, but good luck to anyone trying to find those answers here. Grayfell (talk) 22:41, 22 April 2026 (UTC)[reply]
I agree that some of the answers are hidden in the chaff. I'm in favor of cutting to get it into a coherent form. Moving Features above Training would help. But, the mere fact of a sprawling Features section is disturbing given that the concept of features is largely a marketing gimmick. What do you think about condensing that down into a brief functional description along the lines of Siemens_NX#Key_functions? Uhoj (talk) 18:50, 23 April 2026 (UTC)[reply]
I think the "Languages" subsection isn't very useful and could be trimmed, and there may be other sentences that don't provide useful information to readers to remove, but condensing down to something like Siemens_NX#Key_functions seems way too radical in my view. Alenoach (talk) 19:30, 23 April 2026 (UTC)[reply]

Usage chart from OpenAI

Looks like we also need to discuss a new chart since it was removed. What do other people think? Uhoj (talk) 22:26, 19 April 2026 (UTC)[reply]

As I said, I don't accept that the source is reliable. The source was not being cited either, and if cited, it would need to be attributed to OpenAI affiliated researches.
Further, the chart is more confusing than helpful. A reader who looks closely at this chart is likely to ask questions that are not answered in the article. For example, the difference between 'creative ideation', 'personal writing', 'write fiction' is not obvious at all. Likewise for the difference between 'greeting and chit-chat' and 'personal communication'. I'm not saying there is no difference, I'm saying the chart doesn't have enough room to explain the difference according to the authors. Grayfell (talk) 22:38, 19 April 2026 (UTC)[reply]
When I maximize the image link that you added here I see the full citation down at the bottom. Is that not showing up for you? Uhoj (talk) 22:49, 19 April 2026 (UTC)[reply]
The citation needs to be in the article instead of via in Commons. For WP:V, info on Wikipedia should be cited on Wikipedia, not on a sister-project. Further, because this is a biased, unpublished source, it would need attribution. Adding a citation is a very easy fix. The attribution and RS issues are not. Grayfell (talk) 00:39, 20 April 2026 (UTC)[reply]
I think the chart has information value and the source doesn't look too bad. Some of the text is tiny, though I don't know if there is an easy fix for that. Alenoach (talk) 23:51, 19 April 2026 (UTC)[reply]
The image presents ambiguous subjective claims with false precision and without context. If this data is significant, a reliable, independent source should be cited for it. Otherwise, it's yet more trivia coming from OpenAI. So it's informational value should be decided by RS.
Uhoj made this image, so presumably they could fix the text issue most easily. Otherwise, SVG files are pretty easy to edit. Grayfell (talk) 00:39, 20 April 2026 (UTC)[reply]
Agreed Czarking0 (talk) 01:39, 20 April 2026 (UTC)[reply]
I like the idea of having a graph on this, it's a good way to condense a lot of information; having the equivalent in prose risks being a little boring. Even if some categories are a little vague, there is still a lot of information to get from the graph, which would fit well in the "Applications" section. We could use this source from TheStreet, which includes the paper's graph containing the same information and makes verification easy. If we need clear attribution, we could have a caption like "How people are using ChatGPT, according to an NBER working paper by the OpenAI Economic Research team and David Deming" (similarly to the attribution this Gizmodo article makes), though a less verbose attribution might be better. Alenoach (talk) 09:49, 20 April 2026 (UTC)[reply]
I also like the idea of having a graph. However, I think this particular one has issues due to being primary and non-independent. I do not think a secondary source republishing the graph makes it independent nor secondary. At best it is a quote. Independent secondary graphs on this would be a better starting place. Czarking0 (talk) 15:15, 20 April 2026 (UTC)[reply]
The article in TheStreet is a start, but that source directly describes this info as coming from OpenAI. It also includes some important context that was missing from the proposal: However, it should be taken with a grain of salt: the new report only looks at the consumer version of ChatGPT, missing out on some of the enterprise uses or private instances. etc.
I understand the appeal of graphs for complicated information, but as I said, this looks like false precision. The graph introduces ambiguities about methodology and definitions, and a reader would need to dive into the preprint itself to attempt to find answers. Grayfell (talk) 19:19, 20 April 2026 (UTC)[reply]
Deloitte has a lot of data, but it’s on AI usage in general
https://www.deloitte.com/content/dam/assets-zone3/us/en/docs/services/consulting/2026/state-of-ai-2026.pdf
I didn’t see anything specific to openai, but there’s a lot of data on usage. WhaleFarm (talk) 19:41, 20 April 2026 (UTC)[reply]

Reliability of the NBER working paper was discussed at WP:Reliable_sources/Noticeboard#NBER_working_paper --Uhoj (talk) 17:06, 21 April 2026 (UTC)[reply]

How adults are using ChatGPT Consumer, according to a working paper from NBER by the OpenAI Economic Research team and David Deming.[1][2][3]

@Czarking0 Why was this version of the chart removed? The revert summary is See talk. What I see on this talk page is objections to sourcing from the working paper and from TheStreet. Attribution was suggested and incorporated. Do you dispute the reliability of all of the sources being used for this version? Uhoj (talk) 18:57, 23 April 2026 (UTC)[reply]

Text size has been increased where possible and false precision has been decreased per feedback above. Might take a bit for the new version to be displayed due to caching. Uhoj (talk) 01:54, 24 April 2026 (UTC)[reply]
My understanding of this talk section was that the current consensus was to not include this chart. I am sure that if I am mistaken others will correct me as well. Czarking0 (talk) 02:14, 24 April 2026 (UTC)[reply]
I don’t think you have any consensus on that chart but I would support a well formatted inclusion of that data. WhaleFarm (talk) 04:17, 24 April 2026 (UTC)[reply]
I do not think the chart is appropriate. While cutting off the numbers past the decimal point was an improvement, that was only a small part of the false precision problem I was attempting to describe. I could explain in more detail, but a much better approach would be to ignore the study itself and look only at what reliable, independent sources say about it.
So, is MediaNama reliable? I'm not sure, but that source includes the specific data as only one small part of the story, so citing it only for that data is disproportionate. The Ars Technica source doesn't focus on this data, instead focusing more on demographics of the users. Both sources include an image from the study itself of the usage data, but in both cases, the resolution is too small to read clearly, suggesting this is included more for decoration than for information. As a tertiary source, Wikipedia shouldn't highlight raw data in this way, and we shouldn't include images for decoration. We should instead use reliable secondary sources to provide an explanation of this data for readers. Grayfell (talk) 08:31, 24 April 2026 (UTC)[reply]
The caption would probably need to mention that only users who declared being aged over 18 and who are not on Business, Enterprise, and Education subscriptions are included in the study. Even considering that the study has these selection biases and that there is some arbitrariness in the categorization, I would support its inclusion. I think it contains a lot of interesting information, the visual looks good and would fit well at the beginning of the "Applications" section, and there is coverage in reliable news websites like Ars Technica which suggests some notability and credibility. But consensus is needed for inclusion (rather than removal), so you would need more support than opposition to include it into the article. Alenoach (talk) 14:50, 24 April 2026 (UTC)[reply]
By the way, in case someone is interested, there is a smaller study here of how people use ChatGPT, by The Washington Post based on conversations published by OpenAI. Alenoach (talk) 23:54, 24 April 2026 (UTC)[reply]
I think this is a well written article. Solid research. Well explained methodology. Not judgmental. I would use it. WhaleFarm (talk) 03:40, 25 April 2026 (UTC)[reply]
I plan to polish this more
How People Use ChartGPT according to 2025 NY Times study.
Czarking0 (talk) 03:48, 25 April 2026 (UTC)[reply]
Why is the bar graph credited to NY Times? The SVG seems to have been made using AI. Is the credit maybe a hallucination?
The Washington Post exposé is a primary source. I cannot find a single secondary source for the graph. This is rather concerning given that the NBER paper from around the same time is widely cited both by journalists and academics. In fact, the Washington Post can be cited as a secondary source for a good chunk of NBER graph.[1] Uhoj (talk) 17:48, 25 April 2026 (UTC)[reply]
I just made a typo. I am ok with not including this either. If one of the two charts should be included I think the argument for this one is stronger. The source primary but it is independent. I'll see if secondary sources reference this Czarking0 (talk) 18:42, 25 April 2026 (UTC)[reply]
Also I will warn that you need to WP:AGF Czarking0 (talk) 18:43, 25 April 2026 (UTC)[reply]
The caption has been updated to describe sample selection.
Seems like it might be good to run an RfC about this. I'm thinking a question along the lines of "Should the chart be included in the article?" Does anyone have alternative ideas for the question? Uhoj (talk) 15:08, 24 April 2026 (UTC)[reply]
your question is to the point, but maybe you need to add a little about how we got here, and why the inclusion of uses. WhaleFarm (talk) 19:17, 24 April 2026 (UTC)[reply]

References

  1. ^ Orland, Kyle (15 September 2025). "What do people actually use ChatGPT for? OpenAI provides some numbers". Ars Technica. Retrieved 21 April 2026.
  2. ^ Das, Prabhanu Kumar (17 September 2025). "Over 70% ChatGPT Interactions Are Non-Work, Says OpenAI". MEDIANAMA. Retrieved 21 April 2026.
  3. ^ Chatterji, Aaron; Cunningham, Thomas; Deming, David; Hitzig, Zoe; Ong, Christopher; Shan, Carl Yan; Wadman, Kevin (2025). How People Use ChatGPT (Report). Cambridge, MA: National Bureau of Economic Research. doi:10.3386/w34255.

Updates

It has been some time since the last round of talk page discussion. I tried to incorporate everyone's comments as I saw fit. Are the outstanding issues on the talk that people feel like need to be addressed prior to a new GAN. Also I think a new GAN should include Grayfell as a nom if they desire. Czarking0 (talk) 00:21, 9 May 2026 (UTC)[reply]

Retrieved from "https://en.wikipedia.org/w/index.php?title=Talk:ChatGPT&oldid=1356341520"