AI Policy Wiki
Dashboard
⚲
AI Policy Wiki › Concepts & Frameworks

AI Research Authorship and Attribution

medium confidence · updated 2026-09-13

The norms and emerging rules governing who counts as an author of a research result an AI system produced, what disclosure of tool use is required, and who bears the obligation to credit the prior human work a model draws on. Converging answers across the Leiden Declaration, developer authorship statements, journal and preprint-server policy, and the copyright standard of meaningful human authorship.

Contents

As AI systems began producing results that would carry an author's name if a person had produced them, three separate questions came apart that publication norms had previously answered together: who may be named as an author, what must be disclosed about how a result was produced, and who owes credit to the prior human work a model drew on. Different institutions have answered them differently, and the answers are being set by professional declarations, developer statements and platform policy rather than by regulation.

The mathematics case is the most developed instance and is treated at Machine-produced mathematics: claims, verification and credit; this page covers the norm itself and its reach into other research fields.

Authorship: the exclusion of machines

The position that credit belongs to humans is nearly unanimous among the bodies that have stated one, and it is stated as a matter of responsibility rather than of contribution.

The Leiden Declaration on Artificial Intelligence and Mathematics, endorsed by the International Mathematical Union, puts it as an instruction to individual researchers: "Credit and responsibility continue to belong to humans within the mathematical community and should not be given to automated systems. Artificial intelligence may obscure, but does not replace, the collective human labor behind a result." It pairs the rule with a second one assigning where responsibility falls when tools are used — "the responsibility for the correctness and adequacy of the arguments and results, as well as for the completeness and accuracy of citations to relevant prior work, remains exclusively with the human authors."

OpenAI reached the same exclusion from the other direction, and drew the opposite operative conclusion. Citing the Declaration's signers in the statement accompanying its September 2026 mathematics results, the company said that "claiming human authorship for a proof generated entirely by an AI system would misrepresent both the system's contribution and the nature of genuine human intellectual work," and published a 249-page paper with no individual authors, stating that it prepared the manuscripts, formalized the proofs and takes responsibility for their correctness "while the mathematical arguments themselves were generated by our system" (Ten Advances in Mathematics and Theoretical Computer Science). The Declaration's rule is that a human must be named; OpenAI's application of the same premise is that nobody is.

The two are not reconciled, and the disagreement is about what an author's name certifies. On the Declaration's account it certifies responsibility, which a system cannot bear and which therefore must attach to a person. On OpenAI's account it certifies intellectual origination, which would be misdescribed by naming a person who did not originate the argument. The copyright standard of meaningful human authorship turns on the second reading, and Jessica Silbey and Woodrow Hartzog connect the two directly, linking the disclosure obligations courts have imposed after AI-fabricated citations to that copyright standard (AI Slop (Silbey and Hartzog, 2025)).

Disclosure of tool use

The most concrete rule is disclosure, and it has moved fastest because it requires no position on authorship. The Leiden Declaration asks researchers to "[t]ransparently disclose the use of automated tools, including large language models, machine learning systems, proof assistants, and other mathematical software," and specifically to "[i]nclude a 'Tool and computational resource disclosure' section in your papers," acknowledging that "the precise form of such a section will necessarily evolve." It asks organizations to take the lead in developing publishing and reviewing guidelines covering "tool and computational resource disclosure, attribution, rules pertaining to authorship, and codes of conduct."

A disclosure rule serves reviewing as much as credit. The Declaration's separate recommendation to "[s]upport the needs of reviewing" states that AI use "can introduce material that makes reviewing more demanding," and asks authors to compensate by disclosing tool use, giving complete references and providing formal proofs where feasible.

Attribution of prior work

The obligation to credit the work a result builds on is where the Declaration's authors locate the most direct mechanical harm. Models trained on published work "frequently return outputs that do not properly cite the human works they synthesize," and many are "built on data obtained by systematically exploiting licenses and access arrangements that were not made with artificial intelligence in mind, or indeed by simply violating copyright protections" (Leiden Declaration on Artificial Intelligence and Mathematics (2026)). The Declaration converts that failure into a duty on the researcher rather than the developer: "[t]he known limitations of automated tools in properly attributing ideas create a corresponding obligation for proactive effort to find and credit the sources that made a new result possible," and where satisfactory attribution is not possible, to say so in the publication.

The twenty-five Fields Medallists who signed "A Severe Misalignment of AI in Mathematics" locate the same failure in announcement timing rather than in model behaviour: solutions "announced in a rush" leave no time for "a proper writeup, the isolation of new methods and ideas, and citing relevant previous work of others," which they say raises "severe attribution and plagiarism questions" (A Severe Misalignment of AI in Mathematics (Fields Medallists' declaration, 2026)). Both accounts converge on the same practical consequence — that the attribution work does not disappear, it moves onto people who did not produce the result and are not credited for the labour. The Caltech open letter of September 2026 states that cost explicitly: after a result is announced, "research mathematicians are compelled to step in to properly verify, disseminate, and sometimes discredit entirely the claimed results," labour that "goes uncompensated, uncredited, and unacknowledged" (Source: businessinsider.com).

Developer practice on prior-work credit is not uniform. Anthropic's account of formalizing Fermat's Last Theorem credits the Wiles–Taylor proof as a culmination integrating work by Gerhard Frey, Jean-Pierre Serre, Ken Ribet, Barry Mazur, Robert Langlands, Jerrold Tunnell, Yutaka Taniyama, Goro Shimura and André Weil (Formalizing Fermat's Last Theorem (Anthropic, September 2026)). In the September 2026 Navier–Stokes dispute, by contrast, the cascade technique underlying the construction was traced by outside commentary to Luis Martínez-Zoroa's 2021 dissertation and his later work with Diego Córdoba, neither of which used computers, rather than by the announcement itself (Source: implicator.ai).

Enforcement through publication gatekeeping

Where declarations state norms, publication venues have imposed consequences, and the operative rule so far concerns fabrication rather than authorship. arXiv tightened its endorsement rules on January 21, 2026: new researchers must obtain a personal endorsement from a previously approved researcher, and an academic email address is "no longer a sufficient credential for determining minimum research competence." Thomas G. Dietterich, vice chair of the editorial advisory council, announced that authors submitting AI-generated fake citations would be banned for a year (Source: blog.arxiv.org). See AI-Fabricated Citations.

The Leiden Declaration asks mathematical organizations to go further, demanding "that mathematical results continue to be published in peer-reviewed venues such as journals, proceedings, and books," with press releases and blog posts confined to "a valuable supporting role" that "cannot replace peer-review or community scrutiny." That ask is addressed at a publication channel rather than at a disclosure field, and it is the recommendation that bears most directly on how developers have announced results.

Reach beyond mathematics

Both mathematics documents state that their subject is not confined to the discipline. The Leiden Declaration is written "in solidarity with other research endeavors and creative professions facing similar challenges, both within and beyond academia," and situates itself alongside the Uppsala Code of Ethics for Scientists, the San Francisco Declaration on Research Assessment and the UNESCO Recommendation on Open Science. The Fields Medallists describe "a general threat to intellectual work, with misalignment between the outcome of the use of AI and its initial purpose," and frame the question as one "that all of humanity might face."

The conditions that made mathematics the first site are specific: formalized proofs can be checked automatically, which makes the field attractive for training general-purpose systems, and results are individually attributable in a way that many empirical findings are not. The Leiden Declaration states the first condition plainly, noting that automatic checkability makes it possible "to generate and check vast numbers of problems … to produce an effectively unlimited source of feedback for training." Fields where results are not individually attributable, or where verification is empirical rather than formal, face the disclosure and prior-work questions without the authorship question presenting in the same sharp form.

Open questions

  • Whether any journal or professional body has adopted a rule naming an AI system as a non-author contributor, rather than either requiring a human author or permitting an unauthored paper.
  • What a "Tool and computational resource disclosure" section is required to contain. The Leiden Declaration proposes the section and states that its form will evolve; no venue's implementation is recorded here.
  • Whether the copyright standard of meaningful human authorship and the publication standard of responsible authorship will converge or diverge. They currently give different answers to the case of a proof generated by a system and prepared for publication by a company.
  • How the attribution obligation applies where a model's training corpus is undisclosed, since a researcher cannot make "proactive effort to find and credit the sources that made a new result possible" for sources they cannot identify.

Relationships

Sources

Categories: Concepts & Frameworks | global | ethics | academic | copyright | technical | opinion
What links here (6)