TL;DR: A March 2026 study in Proceedings of the National Academy of Sciences analyzed 5,114 journals and more than 5.2 million papers and found that roughly 70% of journals now have some form of generative-AI policy — but among 75,172 papers published since 2023 that the authors could check with full-text access, only about 76 (roughly 0.1%) actually contained an AI-use disclosure statement. The gap between policy adoption and author compliance is the real story, not whether policies exist.
What the study actually measured
The study, “Academic journals’ AI policies fail to curb the surge in AI-assisted academic writing,” was published in PNAS on March 6, 2026 (DOI: 10.1073/pnas.2526734123), by Yongyuan He and Yi Bu of Peking University’s Department of Information Management (also posted as an arXiv preprint, arXiv:2512.06705). The authors combined two separate analyses:
- Policy classification. Journal instructions-for-authors and editorial policy pages across 5,114 journals were classified (using a mix of human review and automated screening) into four categories: an outright ban on AI-generated text, a disclosure requirement, an open/permissive policy, or no mention of AI at all. Roughly 70% of the journals sampled fell into one of the first three categories — i.e., had adopted some explicit position on AI use, most commonly requiring disclosure.
- Disclosure-statement detection. Separately, the authors checked full-text access for a subset of 75,172 papers published since 2023 for the presence of an actual AI-use disclosure statement (the kind of Methods-section or acknowledgments-section sentence that policies like the generative-AI disclosure statement convention call for). Only about 76 papers — roughly 0.1% — contained one.
To estimate how much AI-assisted writing was actually happening independent of what authors disclosed, the authors also ran a statistical text-pattern analysis (a maximum-likelihood approach built on the tendency of AI writing tools to over-use certain word choices) across the broader 5.2-million-paper corpus. That analysis is the basis for the paper’s headline claim that AI-assisted writing has increased substantially across disciplines since 2023, with no statistically significant difference in that increase between journals that had adopted a disclosure-requiring policy and journals that had not.
Why the gap is this wide
The study and the coverage around it point to two recurring explanations, neither of which is unique to this one paper — both show up throughout the broader research-integrity literature on self-reported misconduct and questionable research practices generally:
- Reputational risk. Authors reasonably worry that disclosing AI assistance — even limited, permitted assistance like drafting help or literature summarization — will prompt reviewers or editors to scrutinize originality or competence more harshly than the actual policy warrants. Disclosure is voluntary and self-reported at the point of submission; there is essentially no independent verification step in a normal editorial workflow that would catch an omitted disclosure unless a paper is investigated for a separate reason.
- Genuine policy ambiguity. Many journal AI policies distinguish between AI use that must be disclosed (drafting substantial portions of text, generating or analyzing data, producing figures) and AI-assisted copyediting or grammar polishing that is typically exempted — see, for example, Nature Portfolio’s AI policy, which explicitly excludes routine AI-assisted copyediting of an author’s own text from its Methods-section disclosure requirement. Where that line falls is not always obvious to an author deciding, in the moment, whether a specific use “counts.”
Neither explanation implies the 0.1% figure means AI-assisted writing is rare — the study’s own text-pattern analysis suggests the opposite. The 0.1% is a disclosure-compliance rate, not a usage rate, and the paper’s core argument is that the two have become almost entirely decoupled.
How this differs from AI-detection accuracy debates
It’s worth distinguishing this compliance-gap finding from a separate, related conversation this site also covers: disputes over whether AI-detection tools (Turnitin’s AI writing indicator and similar) reliably identify AI-generated text at all, including well-documented false-positive controversies — see AI Detector False-Positive Controversies: What Researchers Should Know in 2026. That debate is about detection accuracy on the output side. The PNAS study is about something upstream of detection entirely: whether authors are following the disclosure step their own journal’s policy already asks for, regardless of how good or bad any detector is at catching an undisclosed instance after the fact. The two problems compound each other — weak detection reduces the practical cost of skipping disclosure, and weak disclosure compliance means detection is doing more of the enforcement work than any editorial policy was designed to rely on.
What this means for editors, research offices, and authors
For research-integrity offices and journal editorial teams, the practical implication is that a published AI-use policy is necessary but not remotely sufficient on its own — the enforcement mechanism (or lack of one) matters at least as much as the policy text. Points worth acting on:
- A policy without a verification step is close to self-certifying. If disclosure compliance is checked only when a paper is already flagged for another reason, the policy functions more as a liability shield than a transparency mechanism. Some publishers are moving disclosure questions into the formal submission workflow itself (a required checkbox or field, rather than a sentence buried in instructions-for-authors) specifically to make the decision explicit rather than easy to skip.
- Ambiguity is a fixable input, not a fixed cost. Journals that publish a short, concrete “what needs disclosure vs. what doesn’t” list (as several publisher policies already do) give authors a clearer decision rule than a general disclosure requirement alone. See AI disclosure laws for how this same disclosed/exempt distinction shows up across government, funder, and publisher requirements — it is not a single universal rule.
- Institutions should not treat “our journals have a policy” as equivalent to “our authors are compliant.” The two are measurably different things per this data, and an institution’s own compliance monitoring (where it exists) should be designed around that distinction rather than assuming policy adoption alone closes the loop.
Where the policy landscape is headed
The compliance gap this study documents is part of what’s driving a live, ongoing standard-setting effort: the Committee on Publication Ethics (COPE), together with the World Conferences on Research Integrity Foundation, the International Science Council, STM (the International Association of Scientific, Technical and Medical Publishers), and the Global Young Academy, is working toward a joint “Global Reporting Standard for AI Disclosure in Research” — informally referred to as the Vancouver Standard, after its role as a focus track at the World Conference on Research Integrity in Vancouver in May 2026. A staged public consultation is running through 2026: an initial round mapping disclosure needs and preferred formats closed in early 2026, a second round on what specifically should be disclosed is running through September 2026, with a final version targeted for later in the year. Whether a shared, cross-publisher standard narrows the gap this study measured — as opposed to simply adding another policy layer to a compliance problem that current policies already aren’t solving — is exactly the open question the next round of data on this topic will need to answer.
Frequently asked questions
Does a 0.1% disclosure rate mean only 0.1% of papers actually used AI?
No. The 0.1% figure is the rate at which papers contained an explicit AI-use disclosure statement, not a measurement of how much AI-assisted writing occurred. The same study’s separate text-pattern analysis of AI-typical word-choice patterns across its full 5.2-million-paper corpus found evidence of a substantial, ongoing increase in AI-assisted writing since 2023 — the paper’s central point is that this increase is essentially unrelated to whether the disclosure rate moved at all.
Did having an AI policy make any measurable difference?
Per the study, no significant difference in the growth of AI-assisted writing was found between journals that had adopted a disclosure-requiring policy and those that had not — the authors’ framing is that current policies have “largely failed to promote transparency or restrain AI adoption,” at least as measured by this dataset.
Is this the same issue as AI-detector false positives?
No — they’re adjacent but distinct problems. This study is about authors not disclosing AI use their own journal’s policy already requires them to disclose. AI-detector false positives are about a separate tool incorrectly flagging human-written text as AI-generated. See AI Detector False-Positive Controversies for that side of the issue.
What should a journal’s AI-use disclosure statement actually contain?
Most current publisher policies converge on the same basic shape: name and version of the AI tool used, the specific purpose it was used for, and an explicit statement that a human author reviewed and takes responsibility for the resulting content. See the generative-AI disclosure statement entry for the operational definition and worked examples.
