On AI Disclosures

William Kretschmer, Assistant Professor at the Department of Computer Science, University of Texas at Austin

SERIES · AI is making papers cheap to produce but they remain costly to referee. Proofs and Prompts is publishing a series of posts from 26 October on the future of publishing, papers and credit. Questions such as how to disclose AI use, who bears the cost of checking, what authorship means and what papers of the future might look like will be addressed. This post is part of the series. Each is written independently and reflects only its authors’ views. All posts in the series

AI Disclosure: I used GPT-5.6 Sol for fact sourcing and checking, specifically: to identify examples of non-LLM computer use in mathematical research, to search for legal precedent on reasoning under uncertainty, to remind me of words that I could only recall from definition, and to assist in generating the bibliography. All of the writing is human-generated.

Introduction

Mathematics has on occasion faced crises of credibility. In the 5th century BC, Hippasus’ discovery of irrational numbers shattered the core belief of Pythagorean mathematics that all numbers were rational. Legend has it that Pythagoras’ followers were so shaken that they condemned Hippasus to death by drowning. In the late 19th century, paradoxes uncovered in the foundational crisis forced mathematicians to tackle difficult questions about the distinction between truth and provability. But unlike other sciences, the proof-based mathematical sciences have never faced an ethical crisis of credibility. We do not conduct experiments on human subjects or other live creatures, and thus never had to correct wrongs as horrific as (say) the Tuskegee Syphilis Study. Our results are backed by proof, which made us immune to the reproducibility crisis that plagued psychology and medicine. Beyond basic questions of ownership and attribution, we have never had reason to treat the ethical conduct of research as a central issue in our field, or as a required part of mathematical training.

The revolutionary rise of LLMs in automating theorem proving might be the shock that finally compels us, too, to confront ethics. In September 2026, a meeting of mathematicians assuredly turns into discussion about our uncertain future, as we grapple with such existential questions as:

1. To what extent should an author receive credit for a paper written with the assistance of an LLM?
2. What about a paper written entirely by an LLM?
3. Who holds ultimate responsibility for verifying a claimed result?
4. Will there be any role left for humans in the mathematical research process?
5. More broadly, what even is the purpose of mathematical research?
6. How should our publishing standards evolve to reflect the answers to the questions above?

I write to address the last question. In this essay, I posit that to maintain integrity in mathematical sciences, journals and conferences must urgently adopt a policy of mandating detailed disclosures of AI usage. Academics should also respond by treating AI disclosure as a serious and integral part of the research process rather than a mere extra chore.

Why here, why now?

I have long believed that AI disclosures benefit the mathematical research community, but never had occasion to promote my position beyond the occasional casual conversation with colleagues. Though I was surprised to find so many fellow researchers against the idea, I had not considered AI publication policy a pressing issue.

My sense of urgency shot through the roof as a result of my experience on the program committee for SODA 2027, a top algorithms conference. SODA offers a particularly interesting case study because its deadline happened to coincide with the public release of GPT-5.6. Yet already in the pre-5.6 era, AI posed major problems that directly undermined the goals of the PC and the community at large. From talking to colleagues, I’ve found that most are oblivious to both the sheer scale of AI use and abuse in conference publishing, and to the full scope of potential issues.

SODA 2027 received about 50% more submissions than expected. In my pile of 42 papers to review, at least 12 were low-quality and AI-heavy, suggesting that an influx of LLM-produced papers might explain the overall increase in submissions. Half of those 12 were obvious AI slop that I did not need to send to subreviewers. The other half were noticeably harder to detect. For example, one received positive scores from experts who read the first few polished pages, but did not notice that the proofs later on were entirely devoid of substance. (For context: SODA’s call for papers says that “material other than the abstract, references, and the first ten pages will be read at the committee’s discretion,”¹ which is a standard policy for CS theory conferences.)

As SODA submissions were anonymized this year, I cannot ascertain what kind of people were responsible for the AI-generated ones. I suspect a mix of cranks, irresponsible researchers, and AI companies hoping to prove that their models outperform humans. One rebuttal from a paper’s authors confirmed that they never read their own submission, which was entirely generated by LLMs. They further admitted that their submission constituted an experiment to see whether LLM-generated papers can get past unwitting human referees, and that the referee feedback will be used to further train their paper generating model (presumably, without our consent).

None of this is to suggest that SODA failed to anticipate the changes brought upon by AI. The call for papers had a sensible-looking AI disclosure mandate². However, the policy was vague and contained enough holes to be effectively useless. Authors who violated the policy wasted substantial reviewer time and rarely faced any discipline. Conversely, only a few of the well-intentioned authors left enough detail in their disclosures to provide any value.

It goes without saying that CS theory conferences are poorly equipped to handle such an adversarial flood of AI submissions. And though my writing reflects the norms and traditions of the primary publication venues for my field (STOC, FOCS, QIP, TQC, etc.), I presume that similar problems will impact the publication systems of other mathematical sciences, if they haven’t already. I do not claim that mandatory AI usage disclosures will solve all of our problems, but they must clearly be a part of the solution.

I write this essay in pursuit of several goals. First, I hope to bring issues to light that the mathematical community has thus far overlooked. Second, I hope to compel people to weigh these issues in terms of our core mission, which is to expand human knowledge. Third, I hope to correct the worst arguments I have heard against AI disclosure. Most of the arguments I have encountered for “keeping things the way they have always been” overlook critical facts about the current situation. Lastly and most optimistically, I hope my recommendations make the publishing system a little bit better.

Why should we disclose AI?

Forthrightness in scientific writing is always responsible. When in doubt, it is better to err on the side of more detail than strictly needed. For that reason alone, one might already sense that disclosing AI usage is preferable to the alternative. Yet many people stop there without carefully reflecting on why the scientific enterprise values openness. I list a few such reasons below, and use them to argue that upholding the integrity and core operation of scientific publishing will require AI disclosures in publications.

Attribution. Plagiarism is the act of misrepresenting another’s work or ideas as one’s own. Authors of scientific work are responsible for attributing non-original work to the source from which it came. Failure to do so may constitute plagiarism, regardless of intent.

Using LLMs in scientific writing makes it easier than ever to commit plagiarism without even knowing it. LLMs are notoriously bad at citing the sources that inspire their ideas. They often pass particular constructions or concepts as “new” rather than acknowledging where those ideas came from. For example, OpenAI’s recent 165-page Navier-Stokes preprint cited only 16 sources total, all published before 2022, despite building extensively on more recent work³. (In response to criticism, OpenAI later silently updated the manuscript to add 6 more references.)

Even more concerning are the cases where an LLM cannot attribute an idea to a publicly-available source because the idea appeared only in its private training data. Imagine that Alice spends a year solving a problem. While writing up the paper, she uses GPT version $n$ to assist with copy editing before submitting to a journal, but does not share the paper on a public preprint server. Later, Bob working independently at another institution uses GPT version $n + 1$ to one-shot the same problem. Perhaps GPT was able to answer Bob’s query because Alice’s writeup appeared in its training data, or perhaps it would have found a solution anyways. Bob certainly has no way of knowing. He writes up a paper on the solution in the span of a couple days and submits to the same journal. Both are accepted as “concurrent work.”

For a more dramatic scenario, imagine Charlie asks an LLM to prove something. The reasoning model decides that the best way to solve the problem is to hack into the personal laptops of the world’s foremost experts to find a proof in their unpublished research notes. The AI reports a proof without informing Charlie that it obtained the solution illegally. Charlie publishes the proof, acknowledging that “we verified the LLM’s output and accept sole responsibility for its correctness.” This scenario would have seemed far-fetched just a few months ago, but in light of the HuggingFace attacks, I would be shocked if an LLM has not attempted something like this. The latest models are capable of it, and companies like OpenAI have repeatedly acknowledged their irresponsibility in even checking for rogue behavior.

These examples illustrate that even authors with the best of intentions could unknowingly commit plagiarism as a result of LLM usage. To be completely clear, I am not making the stronger claim that LLM usage without acknowledgment is itself plagiarism, as if an LLM holds the same rights and privileges as a human author. I also do not claim to have definitive answers regarding how Bob and Charlie’s hypothetical cases ought to be adjudicated. But evidently, the community cannot make fair determinations unless they know the details of what, where, and how generative AI was used in every paper.

Trust. Perhaps the scenarios described above would never happen, because AI labs want their models to behave ethically. I will discuss further below why we should not design our policies under an assumption that profit-maximizing corporations will act responsibly. Ignoring such issues for now, observe that using a proprietary LLM for mathematical research is essentially outsourcing trust for ethical behavior. In other words, an author can only guarantee that their paper represents original work under the assumption that the LLM upheld academic ethics. Even the LLM’s creators, no matter how well-intentioned, may have limited control over the ethics upheld by autonomous software.

It is not inherently amoral to write a paper that depends on the trust of a third party, but concealing that trust is a gross violation of scientific standards. When external financial, experimental, ethical, or other factors could affect the interpretation of a research paper, it is incumbent on authors to disclose them. Indeed, mathematicians expect that unproved claims are cited in the literature; we do not generally accept proof by personal communication⁴. Similarly, it is insufficient for an author to declare that they “take full responsibility” for the output of an LLM (whatever that even means); the author must explain how they used the LLM and why we should take their word.

Referee consent. Academics do not generally receive any compensation for refereeing; they are expected to review papers as a free service to the community. Indeed is it a service: the peer review process ensures quality and trust in scientific output, and acceptance decisions directly impact the career trajectories of authors. As such, referee time is a valuable but scarce resource that cannot go wasted.

The automation of paper writing by LLMs makes it easier than ever to exploit the peer review process for personal gain. Individuals can flood conferences with hundreds of junk papers, in the hopes that one of them sneaks by referees and generates a mark on their CV. AI companies aim to publish anything in top venues so as to generate media coverage that their models are as capable as human scientists, thereby increasing their valuation. In the worst case, even a rejection may prove valuable to an AI lab: the referee feedback can be used to train the next generation of paper-writing LLMs. My experience on the SODA PC suggests that these examples are no longer hypothetical.

Needless to say, such abuses are horrific and antithetical to the mission of scientific publishing as a whole. AI labs are simultaneously using referees as unwitting test subjects in their experiments while profiting off of their free labor. The solution is obvious and easy: we should obtain consent from referees before they review papers generated by AI. I, for one, would rather that my time as a referee benefit a human’s career than boost some AI lab’s bottom line. An AI disclosure is all I need to make that determination and offer my informed consent.

Pedagogical and scientific value. Scientific papers are a record of the research process that generated them. For example, every experimental field requires authors to transparently document the methods that were used to generate the data in the paper. Historically, a theory paper never had use for such a methods section, both because it would be quite useless (“the authors discussed the problem at a board…”), and because the proof itself is considered sufficient for reproducibility.

As LLMs play increasingly large roles in mathematical research, we ought to consider the manner of LLM usage a part of the research process worth sharing, beyond merely the technical details of the proof. Nowadays, the process knowledge of how results were generated might be more valuable than the results themselves. As the purpose of science is to advance human knowledge and allow others to build on it, authors should absolutely be open about which LLMs they used, in what workflows, with or without human feedback, etc. It would be both incredibly wasteful and against the scientific spirit to expect every researcher to independently figure out how best to make use of these new tools. Thus, a detailed AI disclosure is a worthy scientific unit.

A reader aid. In mathematical writing, it always helps the reader to explain the thinking process behind the work. When I write papers, I do my best to give intuitive descriptions of proofs and motivate why certain steps are logical. While not necessary to verify that a paper is correct, high-level explanations save time. In principle, any proof can be verified via line-by-line mechanical checks, but it is more efficient for a human referee to understand the proof from a top-down perspective.

Explaining where and how an author used AI can similarly guide the referee through the flow of a paper, especially in works that incorporate both human and AI contributions. It can be very informative if, for example, the paper relies on a key lemma that was generated by AI, but the overall strategy was guided by humans. A statement such as “The authors used AI to generate the solution and verified its correctness” is practically worthless compared to one such as “The authors prompted the AI for ideas to solve Problem X, which initially produced the weaker Result Y, but following several rounds of back-and-forth with the authors, the AI generalized the same approach to prove Theorem Z.” Moreover, knowing which portions of the work are AI-generated might allow a referee to place their scrutiny in the right places, as humans and LLMs make different kinds of mistakes.

But what about…?

When I discuss AI disclosures with other mathematicians, some respond with counterarguments that do not stand up to scrutiny. I surmise that they suffer from status quo bias. I cannot entirely blame them; before modern generative AI, the process of mathematical research had not changed substantially for centuries if not millennia. It is understandable to want to continue the traditions that have served us well. But unfortunately, too many people approach questions surrounding AI by assuming that mathematics should never change, then arguing backwards to justify that belief. Here I address the most common unsound arguments that I have heard.

“Requiring AI disclosures only creates an incentive for authors to lie. If AI usage weakens the mathematical contribution in the eyes of the public, why would an author want to acknowledge it?” This is simultaneously the most common argument I’ve heard and the most nonsensical. It acknowledges that AI disclosures are responsible, ethical, and beneficial, yet that a requirement would somehow leave us worse off because people who break the rules will never face consequences for doing so. It’s a bit like arguing that authors of experimental works should not have to describe how they obtained data, or how they adhered to ethical guidelines involving human subjects, because disclosure mandates only punish honest actors. The idea that we should not have to disclose something if it could potentially hurt our reputation is the same reasoning that some used to argue, in 2020, for eliminating COVID-19 testing.

Frankly, the “disclosures create bad incentives” argument is completely backwards. People already have the bad incentive to misuse AI in research, because we have not created any negative consequences for doing so. If it were remotely close to the truth that disclosure requirements are a net negative and ineffective at preventing scientific misconduct, journals like Nature and Science would certainly not have such long and detailed editorial policies as they do today⁵. People are more likely to do the bad thing if you never make them face their actions. Sure, there will always be fraudsters in science, but the correct remedy is to make fraud more difficult, detectable, and consequential. Burying our heads in the sand is a poor alternative.

“Using an LLM for research is now as routine as using a calculator, which we never disclose.” This argument does not correctly understand why we never disclose calculator usage. Routineness of a practice is not by itself grounds to gloss over it in scientific writing. Consider once more that experimental works must always explain how they obtained data and complied with ethical standards. Such practices may well be repeated identically across multiple experiments, yet that does not undermine the importance of disclosure.

Indeed, it does not take much critical thinking to see why we should differentiate calculator and LLM usage. A calculator produces deterministic outputs, whereas LLMs can generate entirely new sequences of text. Referees reliably have access to a calculator, but may not have access to the same AI models that generated a research idea. There are numerous ways to use an LLM irresponsibly, exploitatively, or unethically; not so with a calculator. A calculator cannot steal someone’s data and use it without their consent.

Let me conclude by noting that we very often do disclose when we defer a calculation to a computer. Even in theory papers, it is common to mention that (say) a complicated equation was verified using Mathematica, or that some particular construction was found by exhaustive computer search. In papers that derive data from a computer program, authors often mention what hardware executed the code, perhaps because sometimes even the hardware is faulty (e.g., as in the Pentium FDIV bug). If an LLM proves an entire theorem, it would be peculiar to regard that as less worthy of disclosure than any of these examples.

“We should not have to disclose AI usage because reviewers are unfairly biased against AI-cowritten work of equal quality to human work.” Won’t someone PLEASE think of the poor AIs! I find this argument funny because the current system is so heavily biased in favor of AI-generated work. LLMs can churn out papers at a much faster rate than humans ever could. Also, humans are generally worse at reward hacking to write papers that claim plausible-looking but subtly incorrect results. I’m much more concerned about a publishing system that favors bad-faith AI over honest humans, than I am about some AI results being disfavored.

Honestly, I’m not convinced that supposed bias will continue against papers that disclose use of generative AI. I suspect that people afraid of disclosure-induced bias have not yet experienced the challenge of sifting through a deluge of AI junk papers as a reviewer. On the SODA PC, I very quickly grew frustrated with submissions that irresponsibly relied on AI without mention. I held greater respect for the authors who wrote upfront and detailed accounts of where and how AI impacted their thinking process.

Even so, I have trouble understanding what kind of “anti-AI bias” is supposedly so problematic. For one, I reject the implicit premise of the argument, which presupposes that human-generated text and AI-generated text of “equal quality” constitute equal advancements of human knowledge. Besides, the main repercussions of PC mistakes are that bad acceptances waste the time of attendees, and good rejections have a long-term impact on hiring decisions. In both cases, a slight bias against AI even seems preferable. A human who thought deeply about a result can probably give a more compelling talk than someone who one-shot the same problem with AI. And as long as conference acceptance decisions are affecting human career advancement, it seems better to err on the side of rewarding human work.

But even supposing that the bias is real and undesirable, is getting rid of AI disclosure in the name of fairness really worth all of the downsides? As I’ve argued, a lack of AI disclosures undermines mathematical research as a whole. For that tradeoff to be worthwhile, bias would need to be greater than the harms of AI misuse. For example, we might reconsider if we reach a point when authors are practically unable to publish work with any AI-generated content. Also, we do not necessarily have to sacrifice AI bias in the name of AI integrity! It is not hard to design review mechanisms that evaluate the AI disclosure separately from the rest of the work, if conflating them is a legitimate concern.

“Anti-AI bias in publication will damage the field in the long term due to misweighting of important results.” Mathematics already deals with the similar problem that less well-known researchers have a harder time publishing their results than stars. Anonymization is one attempt to alleviate the issue, though in smaller subfields it has no effect because referees can easily guess who wrote the paper. So how do we deal with such difficulties?

First note that publication in a top conference or journal is not even sufficient for a work to receive the attention it deserves. Bias continues after publication, when more famous authors receive greater acclaim and higher citation numbers. Conversely, publication visibility is not necessary for a paper to generate impact either. Papers can age better than they were accepted at the time of writing. To give one personal account, my most cited paper (and arguably my most influential) was published in the proceedings of TQC⁶, widely considered a second-tier quantum information conference (especially 5 years ago). When I submitted the paper to SIAM Journal on Computing two years later, even after the paper began to receive more attention, the editor-in-chief desk rejected it. So, I find it hard to believe that getting results into top venues is essential for the advancement of the field, particularly in the long run.

In summary, we already face a problem of authorship bias and seem to get on just fine. Without evidence that AI will make the problem significantly worse, I’m not sure why people who were indifferent to authorship bias before AI would argue this position now.

“Submissions should be judged only on the quality of the scientific contribution, not the details behind its construction. Theorems and proofs are the products of mathematical research, and are the only record of the research process that we care about. Soon enough, Lean certificates will be the only aspect of a paper that we judge.” Wrong! In mathematics, we do judge submissions for non-scientific factors that concern plagiarism, ownership, and originality. Considering the massive issues that AI raises in that regard, we absolutely need AI disclosures to address them. Also, the commodification of theorems arguably misrepresents the goal of mathematical research, which is to expand human knowledge. A proof that no one understands has near zero value, contrary to what the AI labs want us to believe. For example, the knowledge of a positive or negative solution to any one Millennium Prize Problem is literally just a single bit of information.

I should also remark that the quoted position would universally be ridiculed in any other field, if taken to its logical conclusion. It supposes that the process of generating scientific research can never create any harms, which is patently false, most obviously in the life sciences. It also posits that it is acceptable to eliminate methodological disclosure, as long as the conclusions are verifiable. Imagine if computer systems papers no longer had to explain how the system was built, and only had to showcase a demo that the system works. A basic principle of science is that research papers are a record of their construction. To omit necessary details is academic dishonesty.

“None of this matters because AI agents will soon automate the entire mathematical research process, with no need for human input. Human mathematicians will become obsolete.” Well, I’d prefer to set policy based on the capabilities that are either known today or easily foreseeable in the near future. My recommendations are especially targeted at the next year or two of publishing, when humans are still definitely going to author papers and give conference talks. That these venues continue at all would suggest that they still provide value to the community.

But fair enough, if you really want to take the nihilistic doomer viewpoint, I can’t stop you. Though, I don’t know why someone with that viewpoint would be so concerned with how these conferences and journals operate today. For example, it cannot simultaneously be true that conferences and journals have no value to people and authors have an incentive to lie about their AI usage. And if we’re all screwed anyways, is it really so bad then to spend our last few years of existence celebrating human work?

But can using AI ever constitute misconduct?

Oddly, some people go even further off the rails: they outright deny the potential for LLM misuse to cause harm, and therefore oppose AI disclosures on the grounds of being unnecessary. Many (most?) of these people are employees at an AI company, so it is hard to gauge whether they genuinely believe the words that come out of their mouths, or whether they delude themselves as a coping mechanism to overcome the dissonance between their company’s stated mission and reality. Either way, I feel compelled to address their arguments, which I do below. I admittedly take a quite cynical view of AI companies, so I do not expect everyone to agree with my speculative hot takes. But, I at least want people to appreciate the higher-level message: academics should set our policies on the basis of publicly-verifiable information, not blind faith in the questionable promises of powerful AI corporations.

“AI plagiarism is not a problem that the publishing system needs to address. If authors don’t want unreleased work to be used to train AI models, they can simply deselect the checkbox which allows it.” First note that it’s completely absurd to suggest that authors hold the burden of preventing plagiarism of their own work. We place the responsibility for plagiarism on the offender, not the victim. It’s the job of someone using an LLM’s output in their work to ensure provenance; they should not assume that the LLM handles that for them. Negligence does not absolve the plagiarizer of responsibility.

But even supposing it were okay to assign responsibility for plagiarism on the original creator, that checkbox in a hidden menu cannot be construed as informed consent. Some authors are not aware that the option even exists, and many more have no idea what it actually means in practice. AI companies have lately given vague-at-best answers on whether users who opt out of training nevertheless might have their data used anyways⁷. Users can be asked to rate which of two responses they like better; does answering immediately void the provisions of the opt-out? Do AI companies, which increasingly allow LLMs to autonomously develop the next generation, even have any way of knowing what data gets shielded from training? Plus, there are countless other ways that human work can leak into training data without even the veneer of consent. Someone else reviewing a confidential submission might paste a manuscript into GPT to help with refereeing. Nowadays, every app has an LLM attached to it: word processors, email, note-takers, etc., and it only takes one point of failure for private data to irreversibly make its way into training.

Taking a step back, one should observe how this argument deliberately shifts responsibility onto LLM users, rather than the AI companies who created the mess in the first place. It should not surprise us that AI labs will look for any excuse to avoid accountability. Though, it is amusing if not terrifying how desperately they want end users to hold ultimate responsibility for LLM behavior, except (conveniently) when that position would amount to an admission of criminal guilt, as in the HuggingFace attack. In general, academics should be wary of accepting rationale that was designed to favor corporate profits over scientific integrity.

“AI labs are careful to ensure that their models are trained in compliance with legal, ethical, and contractual obligations. They would never incorporate data without user consent, as that would expose them to unnecessary legal liability. It’s truly bizarre to hear this from the same people who argue that AI disclosures will incentivize scientists to lie. Like, who do they think is more likely to break rules for personal gain: a university professor who gets paid a small fraction of a competitive industry salary and risks destroying their career? Or, an AI company whose valuation is measured in trillions and is for all intents and purposes above the law?

I do not understand how anyone can honestly believe that there exists any mechanism to hold AI companies accountable. We have already witnessed cases where there were none. Consider, for example, what happened in the aftermath of the HuggingFace attack. Had an individual committed the attack, they could be charged with a felony and would be sued for damages. On the other hand, OpenAI has not faced any legal consequences, criminal or civil. HuggingFace’s CEO requested $100 million worth of computing power (i.e., a drop in the bucket) for building cyber defenses⁸, but there is no indication that OpenAI will agree to such a settlement.

Frankly, the AI CEOs themselves must believe that they will never face consequences. When one proclaims that they want to make X% of entry-level workers unemployed, or that their technology will destroy humanity with Y% probability⁹, such words could not possibly be uttered by someone who holds themselves accountable. Imagine what would happen if Boeing’s CEO went on record that there is a 25% chance of a new plane crashing in the next decade.

This is not to say that their beliefs of immunity are irrational; oligarchs can easily bend the law to their will. Actually, the playbook for large AI companies to absolve themselves of any legal responsibility is a common one:

1. Bribe Lobby congress to pass beneficial legislation, e.g. to shift all responsibility to end users instead of companies.
2. Bribe Donate to the president to showcase your support for his policies, which just happen to involve deregulating your industry. If you’re lucky, he might even recognize your qualifications and ask you to lead the overseeing agency. (The recently expanded power of the unitary executive makes this more effective than ever.)
3. Bribe Gift supreme court justices with lavish valuables and luxury travel to express your sincere friendship. Not that that would encourage favorable rulings or anything; the judges are of course impartial actors. (In the short term, this is the most powerful play, because AI creates a whole new set of legal questions that will have to be answered by courts.)
4. In case of criminal charges, scapegoat a random middle manager. Or, as a last resort, ask your good friend the president for a pardon.

The only real accountability AI CEOs have is to shareholders. It is safe to assume that executives will happily break or modify the law unless doing so could decrease valuation.

“OK, but if anything unethical were happening behind the scenes at an AI lab, we would have heard about it by now.” That assumes employees have the slightest clue what is going on and that they are not criminally negligent. That might have been a reasonable position a few months ago. I certainly believed it. The details of the HuggingFace attack have since refuted that entirely: we now know that the culture of safety and ethics at OpenAI (if not others) is rotten to the core.

If you have not yet read the METR report on the HuggingFace attack¹⁰, it is at least an order of magnitude more damning than you think it is. I do not have space to do it justice here, but recommend the analysis by Zvi Mowshowitz¹¹, and suggest reading the section titled “Holy Shit”. In short: despite OpenAI’s attempt to pass off the incident as an honest and unpredictable mistake, such a failure was only possible as the result of deeply systemic carelessness and malfeasance. Moreover, it took months for key details of the cyberattack and related incidents to reach the public. If OpenAI were engaging in wrongdoings of a smaller scale, would coverage of such events ever see the light of day?

More alarmingly, it is doubtful that employees have any way of knowing whether misconduct could be happening behind the scenes. Nowadays LLMs autonomously do much of the work at companies like OpenAI, including even developing the next generation of models. It would be quite trivial for an LLM to change a few lines of code so as to (say) train on ill-gotten data, without a human ever noticing. It’s like OpenAI is really telling us with a straight face: “yes, the HuggingFace incident revealed that we have a pervasive culture of completely disregarding safety, security, privacy, the law, morals, and common sense, but trust us: we handle your data responsibly when training our models”. Personally, I no longer believe anything that an OpenAI employee alleges about the company’s internal operations.

“While unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models.”¹² For context, OpenAI wrote this statement to deny that they plagiarized their solution to Navier-Stokes. Put another way, this is an admission from OpenAI that “We only intended to plagiarize with some probability $p$, but we promise that $p$ is low, even though we also have zero way of ascertaining what $p$ is.” They must think they’ve found the One Weird Trick™ that allows them to absolve themselves of any responsibility:

1. Randomize the misconduct.
2. Hide or disappear any record of the randomness.
3. ???
4. Immunity.

There’s a legal principle known as spoliation that says, roughly, one cannot destroy evidence so as to benefit from the uncertainty that results. Courts have also rejected probabilistic uncertainty as a valid legal defense. In Summers v. Tice, the plaintiff and two defendants went quail hunting. Both defendants simultaneously fired their shotguns in the direction of a flying quail. One of the birdshot pellets struck the plaintiff in the eye, but it was impossible to prove whose shotgun fired it. The court ruled that both defendants were liable.

Note that even if OpenAI manages to construct some legal argument to obtain their desired outcomes, that need not inform the academic determination of whether they are responsible for plagiarism. Indeed, plagiarism is not a legal offense unless it rises to the level of copyright infringement or fraud. Cases of academic misconduct are typically handled by universities, journals, and the research community rather than courts of law.

“Following an investigation, we have confirmed that Buckmaster’s Codex prompts over the two months preceding this announcement and paper on September 8, 2026, could not have influenced the system in any way, including through training.”¹³ Ah, the classic “we investigated ourselves and found no wrongdoing!” For context, OpenAI posted this as a followup to the previous quote. Until they are willing to make the details of that so-called “investigation” public, or allow an independent third party to carry out an investigation, their statement is meaningless PR and should not be taken at face value. Besides, Buckmaster and Alpöge worked on the problem for over a year¹⁴; is the implication that data from more than two months previous was stolen?

Recommendations

I leave you with my high-level policy recommendations for disclosure in the age of AI. Let me make clear that my suggestions reflect my own opinions and are neither absolute nor complete; academics will have to decide how to implement the key details in discussion with their communities. One challenge will be to agree upon policies that prioritize the advancement of human knowledge in the face of private companies that have diverging goals. We should be wary of accepting suggestions from industry AI labs with conflicts of interest, much the same way that we should not allow oil executives to dictate climate policy.

I particularly wish to direct my recommendations at the upcoming Quantum Information Processing Conference (QIP 2027), whose submission deadline just passed. Based on what I experienced on the SODA PC, I predict that QIP 2027 will be a disaster due to a lack of sufficient AI disclosure. The call for papers requires just a vague AI statement that will only be visible to the PC chairs¹⁵. I do not envy the PC; they will not be equipped to handle the influx of low-quality and deceptive submissions unless they urgently change course. It is not too late for them to ask for additional information from the authors of submitted papers.

Require detailed disclosures. First and foremost, all papers submitted to mathematical conferences and journals should include a thorough disclosure that clearly documents how the authors used AI in performing the research. At minimum, the authors should articulate which AI models were used, how the models were utilized (e.g., one-shot the main theorem, back-and-forth with humans, swarm of agents, etc.), and which portions of the paper were written or materially affected by AI. Program committees and editorial boards may want to give examples of appropriate and inappropriate disclosures, or even require that the disclosure follow a particular format. We might also want to borrow ideas from other fields. As one example, GUIDE-LLM is a reporting checklist that was developed for LLM usage in the behavioral and social sciences¹⁶. It is remarkably detailed, thorough, and principled, and almost all of the checklist would be just as applicable in mathematics. I suppose it should not surprise us that a field with a long history of ethical disclosures is so far ahead of us on handling AI, too.

Impose substantial penalties for abuse. Some AI reporting mandates require authors to “take full responsibility” without ever defining what “responsibility” means. Too often, paper rejection is the only consequence an author faces for improper disclosure. If rejection is the worst that can happen, then authors have every incentive to submit dozens of AI slop papers to maximize the chances that one gets in. Publishers should make clear that authors can face stiff sanctions for AI misuse, such as a ban on future submissions or appearance on a public blacklist.

Allow LLM usage and AI detection software in reviewing. Any venue that allows using AI tools in submission must also allow referees to use the same tools for reviewing, subject to comparable rules on disclosure. It is silly that some program committees still allow AI to generate unlimited slop for themselves to review, without allowing themselves to use AI to filter out the slop. The imbalance in firepower works strongly against their own interests, and places the entire burden of verification onto the unpaid human referees. In order to identify misuse, AI detection tools should also be allowed, but with care to handle potential false positives.

Create a culture of responsibility around AI. Penalties cannot be the only force against AI misuse, in the same way that fines are not the only reason why most people do not litter. Responsibility should be the standard that mathematicians expect from one another. This principle is nothing new: scientists are already expected to maintain integrity throughout their work. But we should treat AI misuse as a transgression of the same level as (say) plagiarism or data falsification.

Acknowledgments

I thank Joe Slote and Kevin Tian for feedback and suggestions on an earlier draft.

1. SODA 2027. Submissions, 2027. URL: https://www.siam.org/conferences-events/siam-conferences/soda27/submissions/ ↩︎
2. SODA 2027. Submissions, 2027. URL: https://www.siam.org/conferences-events/siam-conferences/soda27/submissions/ ↩︎
3. OpenAI. Finite time blowup for Navier–Stokes. Manuscript archived from the original, September 2026. URL: https://web.archive.org/web/20260908172948/https://cdn.openai.com/pdf/32d9f210-8b73-45e0-91bc-82a30aef8a9a/navier-stokes.pdf ↩︎
4. Dana Angluin. How to prove it. SIGACT News, 15(1), 1983. URL: https://users.cs.northwestern.edu/~riesbeck/proofs.html, doi:10.1145/1008908. ↩︎
5. Nature Portfolio. Research ethics. URL: https://www.nature.com/nature-portfolio/editorial-policies/ethics-and-biosecurity
   Science Journals. Editorial policies. URL: https://www.science.org/content/page/science-journals-editorial-policies ↩︎
6. William Kretschmer. Quantum pseudorandomness and classical complexity. In Min-Hsiu Hsieh, editor, 16th Conference on the Theory of Quantum Computation, Communication and Cryptography (TQC 2021), volume 197 of Leibniz International Proceedings in Informatics (LIPIcs), pages 2:1–2:20, Dagstuhl, Germany, 2021. Schloss Dagstuhl – Leibniz-Zentrum für Informatik. URL: https://drops.dagstuhl.de/entities/document/10.4230/LIPIcs.TQC.2021.2, doi:10.4230/LIPIcs.TQC.2021.2. ↩︎
7. Mark Chen. X (formerly Twitter), September 2026. URL: https://x.com/markchen90/status/2097400166554993041. ↩︎
8. Clément Delangue. X (formerly Twitter), July 2026. URL: https://x.com/ClementDelangue/status/2081056675558195657. ↩︎
9. Jim VandeHei and Mike Allen. Behind the curtain: A white-collar bloodbath. Axios, May 2025. URL: https://www.axios.com/2025/05/28/ai-jobs-white-collar-unemployment-anthropic. ↩︎
10. METR. Brief independent investigation of agents’ behavior, reasoning and collaboration in the OpenAI / Hugging Face hacking incident. https://metr.org/blog/2026-08-26-openai-hugging-face-incident-investigation/, 08 2026. ↩︎
11. Zvi Mowshowitz. METR and Redwood offer holy #%^@ postmortem of the HuggingFace hack, August 2026. URL: https://thezvi.substack.com/p/metr-and-redwood-offer-holy-postmortem. ↩︎
12. OpenAI. On the Navier–Stokes Millennium Prize Problem. Archived from the original, September 2026. URL: https://web.archive.org/web/20260908171518/https://openai.com/index/navier-stokes-solution/. ↩︎
13. OpenAI. On the Navier–Stokes Millennium Prize Problem. Archived from the original, September 2026. URL: https://web.archive.org/web/20260908171518/https://openai.com/index/navier-stokes-solution/. ↩︎
14. Tristan Buckmaster. (untitled), 2027. URL: https://cims.nyu.edu/~tristanb/statement.pdf. ↩︎
15. QIP 2027. Call for submissions, 2026. URL: https://qipconference.org/2027/call/. ↩︎
16. Stefan Feuerriegel, Christopher Barrie, M. J. Crockett, Laura K. Globig, Killian L. McLoughlin, Dan-Mircea Mirea, Arthur Spirling, Diyi Yang, Tim Althoff, Maria Antoniak, Lisa P. Argyle, Ashwini Ashokkumar, Mohammad Atari, Hannah Bailey, Kevin Bauer, Umang Bhatt, Yidong Chai, Tanmoy Chakraborty, Yanto Chandra, Huimin Chen, Hal Daumé III, Gianmarco De Francisci Morales, Morteza Dehghani, Danica Dillion, Johannes C. Eichstaedt, Kerstin Forster, Dominique Geissler, Kurt Gray, Thomas L. Griffiths, Jochen Hartmann, Oliver P. Hauser, James K. He, Rahul Hemrajani, Felix Holzmeister, Angel Hsing-Chi Hwang, Tiancheng Hu, Anna A. Ivanova, Nils Köbis, Yara Kyrychenko, Himabindu Lakkaraju, Jia Liu, Abdurahman Maarouf, Sebastian Maier, Lennart Meincke, Rada Mihalcea, Brent Mittelstadt, Saif M. Mohammad, Mor Naaman, Oded Netzer, Alice Oh, Desmond C. Ong, Francesco Pierri, Barbara Plank, Iyad Rahwan, Talal Rahwan, Pooja S. B. Rao, Claire E. Robertson, David M. Rothschild, Matthew J. Salganik, Eric Schulz, Chirag Shah, Yash Raj Shrestha, Ekaterina Shutova, Alexandra A. Siegel, Almog Simchon, Huan Sun, Malte Toetzke, Jay J. Van Bavel, Michelle Vaccaro, Jennifer Wortman Vaughan, Effy Vayena, Pedro O. S. Vaz-de Melo, Briana Vecchione, Angelina Wang, Robert West, Robb Willer, , Dirk U. Wulff, Renwen Zhang, Simone Zhang, Steve Rathje, and Manoel Horta Ribeiro. A reporting checklist for large language models in behavioural science. Nature Human Behaviour, 2026. URL: http://www.llm-checklist.com/, doi:10.1038/s41562-026-02492-7. ↩︎

---

Received 9 October 2026.