Against mandatory prompt disclosure in AI-assisted scholarship

Large language model (LLM)-assisted writing has disrupted traditional assumptions about authorship, accountability, and research integrity in scholarly publishing (Coeckelbergh and Gunkel 2024; Hajibayova and Lee 2025; Ding 2026; Earp et al. 2026a; Gunkel 2026). In a traditional human-only workflow, named authors are presumed to have made a substantial intellectual contribution to the work and to be able to answer for its integrity (ICMJE 2026). LLMs complicate this picture because they can now contribute far beyond surface editing to help with brainstorming, structuring, editing, rewriting and even drafting entire manuscripts from prompts (Macdonald et al. 2023). And yet, LLMs cannot themselves be listed as authors given the prevailing current view that they are not moral agents and that authorship requires moral accountability (Macdonald and Abbasi 2023; Moffatt and Hall 2025; but see, e.g., Teixeira da Silva 2024; Abernethy 2025; Hurshman et al. 2026). This combination of factors creates a growing concern: some researchers who rely heavily on LLMs may claim authorship on work to which they themselves contributed very little, while editors and reviewers may lack reliable means of assessing such claims.^{Footnote 1} Given the importance of legitimate claims of authorship to the fair allocation of credit and appropriate assignment of accountability in research (Kiermer et al. 2026; Texeira da Silva and Daly 2026; Levy 2026), it seems important to consider possible responses.

One response which has been defended in the recent literature is to call for radical transparency in the use of artificial intelligence (AI) tools in scientific writing. According to several authors, researchers who employ such tools should be required to include all prompts and AI-generated outputs that influenced the content of their manuscripts at the point of submission; this would then enable outside review and might discourage inappropriate claims to authorship (Hosseini et al. 2023; Kendall 2025; Teixeira da Silva and Daly 2026). We call this approach a Mandatory Full Inclusion (MFI) policy, and our aim in this article is to analyze its strengths and weaknesses.

For a recent defense of the MFI policy, consider Kendall (2025). Kendall argues that the full inclusion of AI prompts and outputs should be appended to submitted manuscripts, not only when the availability of such materials would be necessary to support reproducibility, such as when LLMs are used to assist with data analysis, but even in the case of “stylistic” use of such tools, for example to improve language fluency. For another example, Hosseini et al. (2023) argued in Research Ethics that scholars who use LLMs should “describe details such as used prompts and note which parts of the text are affected” and “record and submit their relevant interactions with LLMs as supplementary material or appendices” (p. 449). The Generative Artificial Intelligence Transparency in scientific writing (GAIT) Collaborative Group (2025) likewise called for the inclusion “in a supplement” (p. 9) of any prompts used to generate or edit parts of a manuscript. And versions of this idea have been codified into policy. The American Psychological Association (APA) advises that for some uses of LLMs, authors submitting to their journals should “preserve the output of generative AI [and] document for their own records the prompts used with AI tools [which] can be included [where] appropriate, such as an appendix” (APA 2025).

However, there are some nuances to consider. The APA guidance focuses primarily on LLM use, record-keeping, and sharing of prompts and outputs that would enable the independent analysis of empirical methods for purposes of scientific reproducibility. They do not explicitly state that all prompts and outputs must always be included with manuscripts, and they also allow individual journal editors discretion in setting their own policies. Moreover, Hosseini and colleagues, despite their initial support for requiring authors to include all prompts used in a supplemental file (Hosseini et al. 2023), later revised their position and argued that such inclusion should generally be “voluntary” for several reasons (Hosseini et al. 2025). These reasons include (a) the difficulty or, in some cases, impossibility of accurately specifying which parts of a given text were human- versus AI-generated and (b) the risk of biases against non-native speakers of English who use LLMs to improve their writing.

Thus, at present, there is a range of opinion about when or under what conditions researchers who use LLMs during manuscript preparation should (be required to) include their prompts and/or AI-generated outputs, including drafted or edited prose, data, charts, or figures, along with their manuscripts at the point of submission. In this paper, we will focus our attention on the “stricter” of the live proposals: the MFI policy outlined above. Although we are sympathetic to the intentions behind such a policy, insofar as it reflects a concern to increase research integrity and transparency, we will argue that the policy should not be adopted, for the reasons we will detail below.

Rather than a policy mandating the inclusion of all LLM-related inputs and outputs irrespective of the purpose of AI-use, we will instead suggest that such materials should primarily be included to enable third-party assessment of the scientific validity of research, including the robustness and reliability of any methods employed, analyses conducted, or findings reported (see also Porsdam Mann et al. 2024a, b). At the same time, regardless of whether AI has been used in the preparation of a manuscript, any researcher claiming to be an author on the manuscript should be required to declare and explain how they meet the relevant criteria, for example through an attestation that the conditions laid out by the International Committee of Medical Journal Editors (ICMJE 2026) have been met and/or through the use of a Contributor Roles Taxonomy (CRediT) statement (https://credit.niso.org/).

Before turning to our main analysis, we need to clarify the scope of our critique. We are not opposed to disclosures of AI use in general. Indeed, two of us have elsewhere defended a requirement that generative AI-use declarations should be mandatory in all submissions, whether or not AI has been used, i.e., as a way of normalizing transparency around declaring AI use and making the risks of such use easier to manage (Earp et al. 2026b). Here, then, we are concerned solely with the proposal to make the full inclusion of all AI prompts and outputs mandatory as a default for submissions.

Our paper makes two contributions. First, it shows that MFI is a disproportionate mechanism for achieving legitimate goals of research integrity, because its costs, associated observer effects, and potential to reinforce existing iniquities in scholarly publishing are poorly matched to its likely benefits. Of particular concern is the risk that such a policy would reduce the space of exploratory thinking that is necessary for the best creative work, insofar as scholars who use LLMs might self-censor in anticipation of their prompts being seen by others. Second, the paper outlines what we take to be a proportionate alternative to an MFI policy: routine declaration of AI use in all manuscripts (based on Earp et al. 2026b); appropriate methods-level detail—including prompt logs, where necessary—when the use of AI bears on scientific validity or reproducibility (consistent with Porsdam Mann et al. 2024a); optional inclusion of AI prompts and outputs where authors wish to provide it; and continued reliance on traditional authorship attestation and associated misconduct procedures.

The main worry that seems to motivate current MFI proposals is that unscrupulous researchers might claim authorship on articles that have been largely or entirely generated by LLMs despite having failed to contribute enough, as humans, to merit authorship status (see, e.g., Kendall 2025; Teixeira da Silva and Daly 2026). If those who use LLMs in the preparation of manuscripts were required to include all relevant prompts and outputs with their submissions, the thinking goes, then perhaps third parties such as reviewers or editors could independently check whether authorship was deserved. This, in turn, would hopefully dissuade researchers from claiming to be authors of largely AI-generated papers to which they themselves did not sufficiently contribute.

In support of such an approach, Teixeira da Silva and Daly (2026) have recently stated that they disagree with the existing policy of some publishers, which hold that generative AI prompts “should only be shared with editors upon request.” Rather, they “feel that such prompts should form an integral part of an article’s submission, much like ethical approvals for clinical trials, even including them as mandatory supplementary material … thereby shifting the culture of blind trust [in authorship attestation] statements into one of verifiable trust [based on] evidence” (p. 3, emphasis added).

As journal editors ourselves, we share the underlying concerns about inappropriate claims to authorship. Such claims, which may range from borderline “judgment calls” to intentional misrepresentation or outright fraud, have long posed a threat to academic integrity—even before the advent of generative AI (Bennett and Taylor 2003; Hosseini and Gordijn 2020). Indeed, whether or not LLMs are used, if someone, whether knowingly or through negligent ignorance, fails to meet the minimum criteria for authorship in their field but nevertheless claims to be an author of a paper, this is unethical and, depending on the applicable rules and regulations, likely constitutes academic misconduct.

However, the problem may now be greater than in the past. Prior to the widespread adoption and use of LLMs—that is, prior to about 2023—anyone who tried to put their name on a paper when doing so was clearly unwarranted would have had to contend, or perhaps collude, with at least one other human (namely, whoever authored the paper), creating a natural disincentive to such behavior. By contrast, nowadays, anyone who is able to coax something even minimally publishable out of an LLM—that is, without shaping either the input or the output enough to legitimately claim authorship status—can, in principle,^{Footnote 2} attach their name to the paper and send it in without being discovered. The LLM will not call foul.

Consider the recent case of a real academic who, reportedly, has been churning out multiple articles on seemingly unrelated topics that may—or may not—be almost entirely AI-generated, albeit with no way for reviewers or editors to confirm this (please note that this case is drawn from the secondary literature, having first been identified and discussed by Kendall 2025). And yet, the author at least arguably adheres to current transparency standards: he openly declares his AI use and claims to have contributed to the papers by performing edits. A typical acknowledgment statement reads: “The author acknowledges that this article was partially generated by ChatGPT (powered by OpenAI’s language model, GPT-3; http://openai.com). The editing was performed by the author” (quoted in Kendall 2025, p. 643). The problem with this is:

It is impossible to say how much of the articles [were] generated by ChatGPT. Moreover, there is no way of telling how much editing was done. This, of course, is generally true of any articles that have been written with LLM assistance, which are acknowledged using current guidelines. As reviewers, and subsequent readers, we do not know if 1%, or 100% (or somewhere in between), of the paper has been generated and what edits have been applied. (Kendall 2025, p. 645, emphasis added)

In short, without editors or reviewers having access to the original “receipts” showing the respective contributions of the human and AI, a dishonest researcher could get away with claiming to be an author on a paper that was mostly or even entirely AI-generated, while at the same time—potentially—having only performed some “light edits” of the sort that would not normally legitimize such a claim. Meanwhile, the AI that did most or all of the substantive work on the paper cannot, under most current guidelines, be listed as an author; and nor can the AI raise a complaint against the human for passing off its work as his own.^{Footnote 3}

We agree that LLMs increase the risk of such misbehavior. Moreover, it is not yet clear how much of this risk can be managed by standard editorial processes. On the one hand, there are current limitations in AI writing capacity such that, without substantial human contribution in the form of prompting and editing, most generic AI outputs do not, in fact, constitute a meaningful scholarly contribution. Ideally, then, they would be flagged for rejection by discerning human editors or peer reviewers (Schüklenk 2026). On the other hand, this is likely to be a temporary limitation (see Resnik et al. 2026; see also Earp et al. 2026c), so the problem still requires a solution. What is more, the proliferation of so-called “predatory” journals with low-quality (or no) peer review or editorial oversight means that—even at today’s level of technology—it is likely possible for human researchers (who are not too concerned about the substantive quality of the work) to publish papers consisting of relatively generic, AI-generated material without making any substantial contributions (Cobey et al. 2018; Richtig et al. 2018; Kendall and Teixeira da Silva 2024).

So, what should be done? Toward the end of this essay, we will make some positive suggestions for policy options that might help to address the above concerns. However, as mentioned, our major aim in what follows is to carefully assess the MFI policy option requiring routine disclosure of all LLM prompts and outputs (Hosseini et al. 2023; GAIT Collaborative Group 2025; Kendall 2025; Teixeira da Silva and Daly 2026). Ultimately, we will argue that this proposal—despite its intuitive appeal and admirable motivation—should not, in fact, be adopted by journals or publishers. This is because it would likely (a) fail to bring about the intended outcome of significantly reducing inappropriate authorship claims in the age of generative AI, while also (b) resulting in multiple adverse side effects on other important goods of academic research. In other words, not only would such a policy be ineffective, in our estimation; it would ultimately do more harm than good.

We proceed as follows. We first consider the potential benefits of an MFI policy, drawing on a distinction between the process goods and outcome goods of academic research (Goodman 2010) and acknowledging that such a policy might contribute to each. We then turn to what we see as the main problems with the proposal: (1) it would be unenforceable, and might in many cases be bypassed by precisely those individuals whose conduct it is meant to curb; (2) in some cases, it would be infeasible even for honest researchers to comply with it; (3) it would penalize careful multi-model use; (4) it would disincentivize beneficial AI use more generally; (5) it would introduce a distorting “observer bias” into LLM-assisted writing; (6) it would unfairly disadvantage already-marginalized researchers (as previously argued by Hosseini et al. 2025); (7) it would impose unrealistic expectations on the capacities of editors and pro bono reviewers to evaluate the resulting, potentially voluminous, documentation; and (8) it would be inconsistent with existing and well-justified standards of disclosure in academic publishing. On this analysis, the likely benefits are outweighed by the likely drawbacks. We close by sketching a more modest alternative.

To understand the potential benefits of any proposal, we must first identify the ultimate goods (and bads) at stake, as this forms the baseline against which assessments of benefit versus harm must be made. So, what are the goods at stake in the MFI proposal? As some of us have argued previously (Porsdam Mann et al. 2024a), borrowing a widely used distinction (cf. Goodman 2010), the ultimate goods of academic research (including both scientific work and scholarship in the humanities) can broadly be divided into two main types: outcome goods and process goods. Both are important and must be considered.

Outcome goods, in this context, refer to the outputs of research that are either intrinsically good (e.g., generating new knowledge, finding out truth) or instrumentally good (e.g., useful ideas, beneficial inventions). Process goods, by contrast, refer to the goods involved in the process or means by which research is carried out. Examples of the former include scientific discoveries, medical breakthroughs, and, in general, a better understanding of our world, past, present, and future. Examples of the latter include developing one’s skills, exercising curiosity, teaching and learning, building good relationships, and—among others—receiving appropriate (i.e., duly earned) credit, praise, or admiration for one’s scholarly work.

Of course, the primary purpose of research is, at least arguably, to bring about high-quality outcome goods: fundamentally, it is to generate new knowledge. But process goods are also essential. For example, the appropriate allocation of credit or esteem—relevant to jobs, promotions, grants, awards, and one’s professional reputation—is important for both intrinsic reasons (i.e., as a matter of justice or fairness) and for instrumental reasons: for example, as motivator to work hard and produce good research.

Conversely, unfair (e.g., unearned) allocations of credit, such as by claiming authorship on a paper despite failing to meet the minimum criteria for counting as an author (whether due to the overuse of AI or for any other reason) is problematic. It is both intrinsically bad, because deceptive and unjust, and instrumentally bad, because harmful—both collectively and personally. Collectively, it is bad because it can be demoralizing to work in a system where credit is unfairly awarded (e.g., watching people get hired or promoted based on papers they didn’t meaningfully contribute to); and personally, it is bad, because accepting unearned credit risks, if one is discovered, significant loss of reputation, and even if not discovered, loss of self-respect.

In this context, one can readily see that the MFI policy might aim to stamp out, or at least discourage, one potential means of claiming unearned credit as an academic: namely, through the increasingly feasible use of an AI to generate material that can “get past peer review” without the human user needing to contribute much of anything, apart, perhaps, from a few choice prompts and some light-touch edits (Kendall and Teixeira da Silva 2024; Wei et al. 2025). By requiring aspiring authors to submit all of their prompts, and to identify any content in the final paper that was generated or transformed by an AI, it should be possible to determine when someone does not deserve authorship (for example, according to the criteria laid out in Box 1), thereby blocking this source of unearned credit. This, then, would contribute to (maintaining the integrity of) at least one important “process” good of academia.

In addition, an MFI policy—if consistently adhered to—could also potentially contribute to one or more “outcome” goods. For example, as we have already alluded to, there are likely to be situations in which the ability of third parties to assess the scientific validity of a project would require access to the specific prompts or instructions given to an LLM, such as when AI is used to perform data analysis. Moreover, insofar as the policy dissuaded researchers from using AI in the first place, unless it was truly necessary for their work—a possibility we explore below—it might also have the effect of reducing or eliminating certain problems that are associated with (careless or irresponsible) use of AI, such as the inclusion or retention of “hallucinated” references or incorrect factual claims in submitted papers (Bienz et al. 2026; Naddaf and Quill 2026).

Along the same lines, the policy could help prevent what some of us have called the “provenance problem”—the risk that an LLM inserts uncredited ideas from others into a researcher’s manuscript without the researcher noticing, or with the researcher even coming to believe the idea was originally their own (Earp et al. 2025; see also Daly 2026). Visibility of the prompts and outputs gives a third party at least some opportunity to flag passages whose phrasing or substance bears a suspicious resemblance to a known source (Earp et al. 2026b).

However, there is an important caveat. Whether these potential goods would in fact be realized depends in large part on whether researchers consistently adhered to the policy. But, as we will argue, this is not a safe assumption. Moreover, even if the benefits were realized, this would not be enough to show that the policy should be adopted, since a complete analysis requires careful consideration, not only of potential benefits (vis-à-vis the goods of academic research), but also of the potential drawbacks or harms. We consider these next.

In this section, we consider potential problems or drawbacks associated with an MFI policy proposal.

4.1 Problem #1: unenforceability

The first and most obvious problem with the proposal is that it is unenforceable, meaning that it has a high chance of being exploited, circumvented, or ignored by precisely the individuals it is meant to keep in check: namely, those who are unscrupulous enough to claim authorship on a paper to which they have not in fact made a substantial contribution^{Footnote 4} (see, for example, the first of the ICMJE criteria listed in Box 1). Such a person, in our view, is likely also to be someone who, if faced with a choice between (a) attaching materials to a submission that would immediately falsify their claim to authorship, rendering the exercise moot, or (b) simply failing to declare their AI use, would have strong incentives to choose the latter. Since the reliability and accuracy of AI detection software is still limited (and LLMs can, in any case, be prompted to edit their own output to remove signs of being AI-generated) (Naddaf 2026; Wu et al. 2025), the following outcome of implementing an MFI policy seems plausible: unwarranted claims of authorship on AI-generated papers—a motivating concern for such proposals—would continue apace, whereas disclosures of AI use by the individuals in question could decrease.

Indeed, as we consider further below, even some researchers who could legitimately claim authorship on an AI-assisted manuscript, but who would, upon declaring their AI use, be required to assemble and upload all the relevant AI prompts and outputs, might decide against declaring their AI use in order to save themselves an extra step. This would, of course, be dishonest; but policies must be judged by their likely real-world effects, and not according to unrealistic idealizations. Given that the MFI proposal reflects concerns that some researchers might not be accurately characterizing the extent of their own contributions relative to those of an LLM, it is already a background assumption of such a policy that some researchers may be dishonest. It is therefore self-undermining for the policy to introduce a requirement that would likely incentivize the very behavior it seeks to reduce.

4.2 Problem #2: lack of feasibility in some cases

It may not be feasible for many authors to adhere to an MFI policy. It would depend on the details. For example, it is probably feasible to adhere to the policy when only one or a few LLMs are used and the user’s back-and-forth with the models is kept to a minimum. But it may not be feasible in more complex cases, in which multiple LLMs are used simultaneously or over different sessions.

There are many reasons why multiple sessions—potentially with different models—spread out over time may be necessary or desirable for an LLM user to produce their best work, while still maintaining critical involvement and making a substantive intellectual contribution sufficient for authorship (see Earp et al. 2026c for a case study). For example, authors may wish to consult several models to obtain different perspectives, or even have the models critique each other’s responses, potentially reducing biases associated with any one training distribution (Gallegos et al. 2024; Shi et al. 2024; Lünstedt and Schlippe 2025). Another reason concerns built-in limits on the total allowed length of any one interaction (the so-called context window), plus the tendency of LLMs to ‘forget’ or make less use of information occurring in the middle of a long interaction (Liu et al. 2024). Breaking up the interaction, or prompting a second, third, or fourth model to complete a task while a first model is computing an initial response, may save time and improve the quality of outputs.

Another reason for multiple sessions and/or models may apply to certain forms of collaborative science: for example, when several human authors, each of whom uses one or more LLMs to support their contributions, iteratively work on a complex manuscript over time. In such cases, it might be a substantial burden to locate and assemble all interactions conducted over an extended period with numerous exchanges across several models; and even if all such transcripts could be located, they might run to hundreds or even thousands of pages when collated. Under an MFI policy, the challenge of both assembling and reviewing the documentation (see Problem #7, below) would be considerable.^{Footnote 5}

4.3 Problem #3: disincentivizes disclosure even when feasible

But now suppose a researcher’s LLM use is within the bounds of what can feasibly be documented and disclosed. Assembling all the prompts and outputs is, in principle, just a matter of copying, pasting, and organizing. However, doing so would still be costly in terms of time and effort, and this cost should be appropriately justified. Even modest use of an LLM that extends throughout the course of a typical writing project can produce dozens of conversations and hundreds of prompts. Locating them all, exporting them in a format that can be easily read and understood by humans, deciding what portions or details are relevant, redacting any sensitive or private material, anonymizing what remains for purposes of peer review, and putting it all together into a coherent supplemental file may require many hours of administrative work that contributes nothing to the substance of the paper. Some researchers may ask whether it is worth doing all this, or whether it would be easier not to declare LLM use at all.

Again, failing to declare LLM use under an MFI policy would be dishonest. But as noted in Problem #1, policies must be assessed against likely behavior, not idealized compliance. A policy that introduces a substantial documentation cost to every disclosure of LLM use would likely nudge at least some otherwise honest researchers into omitting the disclosure entirely.

Support for this prediction comes from recent empirical work on AI-use disclosure. This work describes a “transparency paradox”: even without substantial documentation burdens, authors may perceive AI use disclosure as professionally risky—leading to under-disclosure—even when transparency is normatively expected (BaHammam 2025). In a survey of 20 bioethics journals, only 12 papers out of more than 1000 published in 2024 referred to AI use or non-use (Bobier et al. 2025), while independent evidence suggests that the actual usage rate is likely to be many times higher (see discussion and sources in Earp et al. 2026b). Meanwhile, experimental studies show that disclosed AI assistance can reduce perceived writing quality or user trust relative to non-disclosure (Li et al. 2024; Schilke and Reimann 2025), further disincentivizing transparency. These findings, of course, do not show that AI-use disclosure should be abandoned: as mentioned, we support AI use declarations in all manuscripts, whether or not AI systems have been used (Earp et al. 2026b). What the findings support, rather, is an empirical prediction: policies that attach unnecessary burdens or stigma to disclosure of AI use—such as the MFI policy, according to our analysis—may fail to achieve their intended purpose, ultimately worsening under-disclosure rather than improving it.

4.4 Problem #4: disincentivizes beneficial use

Faced with the extra step of compiling and submitting all LLM interactions (or if worried about the observer-bias concern raised next), some researchers may simply decline to use LLMs at all. Although those who are generally opposed to the use of LLMs to assist with scholarly writing may find this to be a desirable outcome, those who recognize the benefits of responsible LLM use would argue that something significant is lost. As argued elsewhere (Earp et al. 2026b, p. 357),

when used responsibly, generative AI can potentially add value to scholarly work.... For example, it can be used to flag grammatical ambiguities, leaps in logic, and/or structural issues requiring closer human attention. It can help authors from various backgrounds translate their work and/or express their ideas more precisely in field-standard language. It can conduct preliminary searches of the literature, enabling authors to check the originality of their intended contribution. And for empirical papers, it can be used to support data analysis, for example by drafting code to run statistical tests—although this must, of course, be hand-checked by authors (or by one of their team) who have the requisite expertise.

Needless to say, LLMs also have well-documented limitations: they hallucinate, can flatten authors’ distinctive “voice” or writing style (Abdulhai et al. 2026), and sometimes import others’ ideas without attribution—what we have elsewhere called “the provenance problem” (Earp et al. 2025). But when these limitations are managed through appropriately skilled use, LLMs can indeed support important outcome goods of academic research: the generation of reliable knowledge. A policy that drives careful users away from LLM assistance therefore harms the outcome goods of academia, without it being obvious that this would result in commensurate gains in process-related goods (e.g., of research integrity).

One might object that for simple papers involving only light LLM use, attaching a chat transcript is not especially burdensome, so at least some honest researchers would continue to use LLMs (and prepare the required supplemental file) notwithstanding an MFI policy. We grant this. But even such relatively limited users would face a further problem—one that we have not yet seen discussed in this evolving debate and which we take to be especially significant. They would face the problem of observer bias or the so-called Hawthorne effect.

4.5 Problem #5: privacy, confidentiality, monitoring, and the Hawthorne effect

A policy requiring full disclosure of all prompts and outputs is likely to introduce a change in how researchers use generative AI systems; a so-called Hawthorne effect (see Berkhout et al. 2022). If a researcher knows in advance that every interaction with a model will, or even may, be visible to editors, reviewers, and readers, this is likely to change the nature of the interaction. Some such changes may be relatively benign, or even beneficial: for example, the prospect of being read may encourage users to write prompts more clearly and precisely. But knowing that their every interaction may be observed may also discourage them from asking clarifying questions (for fear of appearing under-informed), engaging in creative back-and-forth (for fear of appearing unserious), or proposing unconventional questions or angles (for fear of seeming to entertain unpopular viewpoints).

Preliminary evidence for this risk comes from a recent paper by Song et al. (2026). These authors studied use patterns for "shared" LLM accounts—i.e., accounts that had more than one user—through a survey of 245 such users and in-depth interviews with 36 participants, 20 of whom had previously completed the survey.  Song and colleagues argued, based on their findings, that conversational histories leave “cognitive traces”: records of users’ mental lives including experiences of uncertainty, expressions of interests and desires, emotional cues, modes of thinking, and ways of framing problems. Participants reported that, because these records might be seen by others, they would often change their behavior. Some, for instance, reported avoiding personal or speculative topics, adopting alternative phrasings, deleting or editing chat histories, or creating separate accounts for certain types of queries. Although the study by Song and colleagues did not directly examine disclosure of LLM use in a scholarly context, the lesson for that context is still clear: once LLM interactions become visible—or even potentially visible—to an audience, users may begin to perform, curate, or suppress their interactions with the model rather than use the system as they otherwise would.

This, in turn, has implications for freedom of thought and, conversely, for self-censorship as they bear on both process- and outcome-related goods of academic research (Porsdam Mann et al. 2024b; see also Savulescu 2025). Generative AI systems can function as private interlocutors during inquiry, helping researchers test formulations, clarify confusions, rehearse objections, and explore lines of thought that may later be abandoned (Wilson and Burleigh 2026). Academic reasoning often develops through provisional and imperfect exchanges before such reasoning becomes defensible as a public output. An MFI requirement risks reducing the exploratory space that is often necessary both for the intrinsic enjoyment of unfettered intellectual inquiry (process good) and for the best and most creative ideas to be developed, finessed, and eventually published and made available to other scholars and the public (outcome good) (Porsdam Mann et al. 2024a, b).

Ujvari and Odeh (2025) have recently made a similar point about the risks of digital surveillance in higher education. Focusing on the Palestinian context, and drawing on earlier work by Hamamra et al. (2024), they state:

the panoptic nature of digital learning environments [has] transformed education into a mechanism of control, where both instructors and students internalized the gaze of an “invisible Other.” This dynamic undermined the emancipatory potential of higher education by replacing pedagogical care with fear-driven self-discipline. Furthermore, the suppression of free speech in Palestinian universities, fueled by political surveillance, internal factionalism, and institutional complicity, has created a climate of fear and self-censorship that undermines democratic discourse and critical pedagogy. (pp. 460–1)

In short, it is not far-fetched to think that an MFI policy could narrow academic inquiry. Researchers who expect their prompts to be inspected may avoid using AI systems in ways that can support high-quality scholarship: for example, by engaging in early-stage exploration, intellectual risk-taking, and/or critical self-interrogation. Researchers may also become more concerned with how their prompts will appear to an imagined audience than with whether those prompts help them get their best ideas onto the page. In this respect, an MFI policy could produce a misleading kind of transparency. It might reveal an artificially curated record shaped by the disclosure requirement itself, rather than an accurate account of how researchers would otherwise have used the system, or come to their conclusions.^{Footnote 6}

4.6 Problem #6: unfairly disadvantages some researchers

The disincentives to LLM use discussed in Problems #4 and 5 will not affect all scholars equally. Researchers who decline to use LLMs because they do not want to “show their work” (for example, due to observer-related issues) would lose access to a tool that, for some groups, may partly compensate for unfair disadvantages they already face in academic publishing, raising equity concerns.

At least four groups are worth mentioning in this context. First, researchers for whom English is not their first language, who can use LLMs to bring their prose closer to the fluency expected by reviewers and readers (Hosseini et al. 2025; Rueda et al. 2025). Second, researchers with cognitive or physical conditions that affect writing, such as dyslexia, ADHD, chronic fatigue symptoms, carpal tunnel syndrome, repetitive strain injury, multiple sclerosis, stroke, or motor impairments for whom LLM-assisted drafting and editing may lower a barrier to translating their ideas into publishable work (see, e.g., Varma 2023; Gadekallu et al. 2025; Pinto et al. 2025; Zhao et al. 2026). Third, researchers with heavy caretaking responsibilities, who remain disproportionately women in many contexts (Cha and Ailshire 2026; Talbot and Wezyk 2026; also Baker 2012, for an analysis), and who may benefit from being able to get a rough sketch of an idea down quickly and refine it in shorter bursts (Earp et al. 2026d), rather than composing every paragraph from scratch in the limited time available to them. Fourth, researchers at under-resourced institutions, particularly though not only in the Global South, who may lack the postdocs, editorial assistants, and writing-support infrastructure that colleagues at well-funded universities can more readily draw on (Wahl et al. 2026; see also Rueda et al. 2025).

For each of these groups, appropriately supported and critical LLM use may help to compensate for resources or capacities that more privileged researchers either possess or can purchase. A policy that suppresses or disincentivizes LLM use through any of the mechanisms explored herein might therefore widen existing, often unfair disparities in who gets published.

As critics may note, framing LLM adoption as a vehicle for promoting equity risks masking the symptoms of deeper institutional inequalities rather than solving their root causes (e.g., Holzmeyer 2021). We are not convinced that LLM access would in fact “mask” these inequities (it might instead increase the ability of those who are facing such problems to raise awareness of them), but in any case, we make no claim that it would resolve them: background inequities have myriad causes that must be addressed for independent reasons, whether or not LLMs exist. The more serious worry is that institutions might cite the availability of AI as grounds for withholding resources from disadvantaged scholars: saying, e.g., “they can use AI, so they don’t need the funding to hire a research assistant” (or the like). This is a real risk, we assume, but the answer in our view is to advocate for structural reform while also using the available tools to reduce unnecessary barriers to publication in the meantime (see our discussion in Earp et al. 2026c). As Gupta (2013) has argued in a different context, individual and structural interventions can be co-constitutive or force-multiplying, rather than always zero-sum or competing. An accessibility tool operating under non-ideal conditions that enables disadvantaged scholars to cross a certain threshold (say, moving from mostly unwritten ideas to a workable rough draft that can then be polished and published within real-world time and energy constraints) is still an important benefit.^{Footnote 7}

4.7 Problem #7: unrealistic expectations on editors and reviewers

We have so far been focusing primarily on the potential costs of an MFI policy for researchers and authors. However, we should also look at the other side of the table—at reviewers, editors, and publishers. Reviewers are typically selected for substantive expertise in the topic of the manuscript. A paper analyzing, say, an archeologic find in Egypt will be sent to specialists in that period and region. These reviewers may have no particular competence in judging, from a transcript of LLM prompts and outputs, what proportion of the final text reflects substantive human contribution versus generic AI generation. That is a different skill—one that, as far as we are aware, no current professional norm or training program cultivates. Implementing the MFI policy would thus require either training subject-matter reviewers in this new competence or perhaps recruiting an additional reviewer per submission whose specific role is to evaluate the supplemental file.

Both options seem to us unrealistic. Reviewer recruitment is already a serious bottleneck (e.g., Clarke 2025), particularly for journals in the humanities and qualitative social sciences, where editors may struggle to find even one suitable expert per submission (e.g., Trinh 2024). Adding to the number of reviewers required, or asking editors themselves to absorb the extra evaluative work, is not workable at the scale of contemporary academic publishing (see generally, Drozdz and Ladomery 2024). The predictable outcome, then, is that supplemental files would either go unread or be assessed only superficially. Either way, the policy’s intended function—verifying authorship claims through independent scrutiny—would not be performed.

Against this view, one could argue that the supplemental file would only need to be read in suspicious or borderline cases, such as when other cues or evidence raise the possibility that the submitting individual did not (or perhaps could not) contribute enough to the manuscript to warrant authorship. However, this takes us back to Problem #1. Those who would claim to be authors on papers to which they did not make a substantial contribution would, we surmise, be among the likeliest to fail to disclose their LLM use altogether if documentation were routinely required.

Alternatively, if they did upload a supplemental file, they might doctor it to try to make it appear as though they contributed more than they did (see Earp and Porsdam Mann 2026, on this risk). Whether such modifications could themselves be detected by a keen reviewer, perhaps with the help of AI or other computer technology, would then likely depend on whether enough individuals could be trained with the relevant skills—a problem already raised in this section.

4.8 Problem #8: inconsistent with existing transparency standards

Our final concern is more at a systems level and has to do with balancing benefits and risks. Academic publishing has long managed trade-offs between transparency about how research is produced and other goods. We do not require researchers to keep minutes of every meeting where the ideas in a paper were discussed, to upload every draft with tracked changes showing who contributed what at each stage, or to document contemporaneously which paragraph belongs to which co-author. Such records would, in principle, allow much fuller assessment of authorship claims—which, as we noted earlier, have long been subject to gaming and abuse (e.g., ghost authorship, “honorary” authorship), even before LLMs were in the picture. We do, and have done, without such records, however, because the costs of producing and reviewing them, and the chilling effects such mandatory documentation would likely have on the work itself (e.g., through a Hawthorne effect), are judged to outweigh the benefits.

What we rely on instead is author self-declaration, backed by accountability rules that apply when misconduct is discovered. Authors attest that they meet the relevant criteria (e.g., ICMJE; see Box 1), and sanctions may follow if they have not. Of course, the system is imperfect. Unethical claims to authorship have always existed and may never be eliminated entirely. However, the MFI proposal singles out LLM use for documentation requirements far more demanding than anything imposed on other contributions to the same paper. If the case for full disclosure of AI prompts and outputs is sound, it is not clear why parallel demands should not also be made for human collaborators, including that they append their draft histories, email exchanges, conversations at conferences, and so on. This is not, we think, a tenable suggestion.

However, the deeper problem, as we have argued already, is that even if one accepted that such increased documentation burden could be justified in principle, the MFI policy would be unlikely to work in practice, for all the reasons we have given already, potentially even leading to increased non-disclosure of AI use, thereby making the issue moot.

If our analysis is correct, an MFI policy would, at best, deliver modest gains across both categories of academic good discussed above. On the process side, it might serve as a partial deterrent against unearned authorship claims. On the outcome side, it might enable third-party assessment of LLM-assisted methods and analysis, reduce some careless AI use such as the retention of hallucinated references, and help detect cases of the “provenance problem” (Earp et al. 2025). As we have noted, however, whether these goods are actually realized depends on an assumption of general policy compliance that we think it is reasonable to doubt.

Against these potential gains, the policy would, in our view, likely generate multiple problems as argued above. Some of these are problems with the policy itself: its unenforceability against the very individuals it is meant to catch out; its infeasibility in various cases; the unreasonable demands it would place on editors and reviewers; and its inconsistency with established principles for managing trade-offs in academic publishing.

Other problems result in harms to the goods of academic research: disincentivizing AI-use disclosure even by otherwise honest researchers (a loss to process goods), suppression of beneficial LLM use (a loss to outcome goods), observer bias that distorts the writing process (a further loss to outcome goods, and also to process goods, insofar as the experience of writing and freely engaging with ideas might be compromised), and compounded disadvantage for already-marginalized researchers (a loss to both).

Taken together, this suggests that an MFI policy would likely do more harm than good. However, an MFI supporter might object that this sets the bar too high: even imperfect transparency rules can establish norms, deter some misconduct, and make dishonest non-disclosure a clearer breach. In principle, we agree. A relevant analogy might be the conflict-of-interest disclosure requirement that is imposed by many journals and publishers: requiring such declarations does not make dishonesty impossible, but it does create a public record and a professional expectation (as recently discussed in Earp et al. 2026b). However, our objection is not to norm-setting disclosure requirements, which we have argued in favor of elsewhere (Earp et al. 2026b). Rather, it is to policies that mandate full prompt-and-output disclosure without regard to scientific validity and reproducibility concerns, despite their added costs, potential to exacerbate existing inequities, and their weak connection to the specific evidentiary questions at issue.

Given the above problems with an MFI policy, we propose an alternative, more moderate approach. Fortunately, this alternative can be aligned with existing declaration frameworks rather than taking the form of yet another typology. For instance, one recent proposal that we developed with some of our fellow editors suggests a simple three-option declaration: no generative AI use; AI use limited to copyediting human-drafted text; or substantive AI use (this would then be narratively described), with additional methods-level detail, including prompt logs where appropriate, when the AI was used for data collection, data analysis, structured literature synthesis, or similar tasks. Under that approach, exhaustive prompt-level detail will typically be unnecessary, unless the design or output of prompts materially affected empirical results or central arguments in such a way that they cannot be appropriately evaluated without knowledge of the underlying AI exchanges (Earp et al. 2026b). In practical terms, this means that language polishing and spelling or grammar correction may require no more than a concise declaration, whereas AI-supported statistical code generation, data analysis, and so on may require methods-level documentation sufficient for third-party evaluation, consistent with a growing number of proposed reporting checklists and guidelines (e.g., Luo et al. 2025; Feuerriegel et al. 2026).

In short, what we endorse is targeted rather than blanket prompt-and-output disclosure. Prompts and outputs should be submitted with a manuscript when they are needed to assess the scientific validity of the work (for example, when LLMs have been used in data analysis), not as a matter of routine. This is consistent with our previous arguments in Porsdam Mann et al. (2024a).

Second, we advocate for voluntary disclosure of prompts and outputs for authors who want to document their substantive contribution to the prose, structure, or argument of a paper. For those concerned that a reviewer or reader might suspect their work of being overly LLM-generated, the option to submit a transcript or contribution estimate should be available.^{Footnote 8}

Even then, of course, a subjective judgment would have to be made about whether each human contributor’s input is “substantial” enough to warrant authorship assuming all the other criteria (e.g., reviewing the draft, signing off on the work, and so on) are met (ICMJE 2026). In separate work that is still undergoing peer review, we have argued that what is relevant for such judgements is whether the human’s originality, skill, well-directed effort, and connection to the work’s intellectual substance would have counted as substantial in a comparable human-only collaboration, as judged by field- or discipline-specific standards (Earp et al. 2026c). Raw prompts may sometimes help to illuminate that question, but they might not settle it; after all, important intellectual contributions to research are not always recorded, especially in group collaborations.

In any case, we suggest that journals should retain and enforce the existing requirement that anyone claiming authorship explain how they meet the relevant criteria—if not through the SLICE method mentioned in note 8 or something similar, then through standard ICMJE-style authorship attestations and/or a CRediT statement—and treat false attestations as misconduct. This is the same trust-plus-accountability approach that already governs other forms of collaboration in academic work. It is an imperfect system, and some will take advantage of it (Teixeira da Silva and Daly 2026). But the alternative MFI approach would not solve this problem, for the reasons we enumerated earlier. And it would likely introduce problems of its own.

Proposals for Mandatory Full Inclusion (MFI) of LLM prompts and outputs respond to a concern we share. Existing transparency standards do leave room for unscrupulous researchers to claim authorship on largely AI-generated work. This is a problem that needs addressing, and we do not have the perfect solution. However, the proposed remedy in the form of an MFI policy would not, in our judgment, actually solve this problem, whereas it would introduce Hawthorne-style surveillance concerns, with implications for academic freedom (Porsdam Mann, et al., 2024b). A better path, we think, is universally expected, normalized AI-use disclosure, targeted methods-level disclosure where reproducibility or validity requires it, voluntary contribution of documentation where authors want to attest to their own contribution (e.g., using the proposed SLICE method; Earp and Porsdam Mann 2026), and continued enforcement of the rule that authorship requires substantive contribution by the human claiming it. That rule has always rested partly on trust, which LLMs make it easier to abuse. This suggests that what is needed is a broader movement to foster a culture of research integrity and trustworthiness in academia (Kiermer et al. 2026). And there may be other creative solutions yet to be proposed. Our only claim is that an MFI policy would not itself address this problem and would ultimately do more harm than good.