Why 99.9% of Scientists Who Use AI to Write Don't Disclose It

Quick answer

Seventy percent of academic journals now have some kind of policy on AI-assisted writing. Almost none of their authors are following the disclosure part of it. A March 2026 study in the Proceedings of the National Academy of Sciences analyzed 5,114 journals and more than 5.2 million papers published between 2021 and 2025, then checked full text on a 75,172-paper subset published since 2023. Only about 76 papers, roughly 0.1%, contained an actual AI-use disclosure statement. Meanwhile, the same researchers' own text-pattern analysis of the full 5.2-million-paper corpus found AI-assisted writing rising sharply across nearly every discipline, with journals that have a disclosure policy showing no measurably slower growth than journals that don't. A policy that nobody follows and nobody enforces isn't a transparency mechanism. It's a sentence in the instructions-for-authors page that everyone quietly ignores.

What the study actually measured, and what it didn't

Yongyuan He and Yi Bu, both at Peking University's Department of Information Management, built this analysis in two separate pieces. First, they classified the AI policies of 5,114 journals into four buckets: strict prohibition, open/permissive, disclosure-required, or no policy at all. Roughly 70% fell into one of the first three, and the overwhelming majority of those, 3,556 out of that group, were disclosure-required rather than an outright ban. Only 27 journals banned AI writing outright, and just 2 had a fully open policy with no disclosure ask at all.

Second, they measured actual disclosure behavior. Out of 75,172 full-text papers published since 2023 that they could access, only about 76 explicitly said, in the methods or acknowledgments section, that AI was used to help write the piece. That's the number that makes headlines: 0.1%. But the more important number sits alongside it. The same team ran a separate statistical text-pattern method (a maximum-likelihood estimate built on how AI writing tools systematically over-use certain word choices, the same family of technique behind the Kobak et al. "excess vocabulary" method we've covered before in our biomedical-writing piece) across the full 5.2-million-paper set. That analysis found AI-assisted writing rising steadily since 2023, fastest in physical sciences, fastest among authors from non-English-speaking countries, and fastest at high-open-access publishers like MDPI and Frontiers. The growth curve for journals with a disclosure policy looked statistically the same as the growth curve for journals without one.

Put those two findings together and the paper's headline claim follows directly: AI-assisted writing is surging, and the policies that are supposed to make that surge visible aren't catching it. The authors are careful about what they're not claiming. Their detection method can't tell the difference between a sentence lightly polished by AI, which almost every policy explicitly permits without disclosure, and a paragraph substantially generated by AI, which most policies require an author to flag. That's a real limitation, and it's central to how a formal scientific rebuttal challenged the paper months later.

The pushback: a real scientific disagreement, not a debunking

In June 2026, Hong Kong University of Science and Technology physicist Yibo Wang published a formal PNAS "Letter" challenging the paper's causal framing. Wang's core argument: the study used a single policy snapshot from 2025, not each journal's actual policy-adoption date, so a paper published in 2023 under no policy at all could get counted as evidence against a policy that didn't exist yet when the paper was written. Wang also pointed out that lumping "strict prohibition," "open," and "disclosure required" into one binary "has a policy" category obscures the fact that 96.8% of policy-having journals actually permit AI-assisted writing and editing, they just want it flagged. If most policies don't ban AI use at all, rising AI-associated writing isn't automatically evidence the policy failed at its actual goal.

He and Bu's published reply doesn't back down from the core finding, but it does sharpen what the paper is and isn't saying. Their answer: they were never claiming a causal, journal-by-journal test of "did adopting a policy change this specific journal's behavior." They were asking a narrower, aggregate question: has the current policy regime, taken as a whole, produced visible transparency or measurable restraint? Their answer stays no. Policy-having journals didn't show slower AI-writing growth than policy-free journals, and public disclosure stayed rare regardless. They also conceded a real point: their method measures public, full-text disclosure only, so an author who disclosed AI use through a submission-system checkbox or a private cover letter, rather than a sentence in the printed paper, wouldn't show up in their count at all. That's a genuine measurement gap, not a refutation of the transparency shortfall itself.

This is exactly the kind of scientific back-and-forth that should make you trust the underlying finding more, not less. Two peer-reviewed exchanges, both published in the same journal, arguing over methodology while agreeing on the basic shape of the problem: a huge and growing amount of AI-assisted academic writing, and a disclosure rate close to zero.

Why the gap is this wide

Nobody in the PNAS debate feels puzzled about the mechanics. Instead, two forces explain most of the confusion. The first is reputational risk. Even a tiny admission of AI help, like drafting or pulling a literature recap, doesn't sit well with reviewers who might start second-guessing the originality of your ideas. Editors, too, might read the rest of your work more skeptically. In practice, there's almost no independent check in the standard editorial workflow to flag a missed disclosure. Only if the paper triggers another flag do people even look. The second force is the fuzzy line between what needs to be disclosed and what doesn't. Rules from Nature Portfolio and elsewhere say routine AI-assisted copyediting or syntax fixes don't require disclosure, but they do for AI that draws conclusions, generates data, or writes out major sections. For a real manuscript, it's hard to spot whether a particular sentence falls in or out of the range. That fuzziness invites people to assume it doesn't count. As a result, weak detection lowers the cost of skipping disclosure, and weak compliance forces the detection mechanism to perform enforcement, a role no editorial policy was designed to handle alone.

What this means if you're an author, editor, or reviewer right now

Few people openly admit using AI tools. Reviewers and editors normally spot these hidden AI sections. The research indicates that editors are becoming increasingly aware of this overall trend. Checking a draft with a detector before it goes out is comparable to performing a citation audit. The detector flags any potential AI voice, and the feedback helps authors recognize parts that might violate their journal's disclosure rule. Importantly, the detector treats both polished text and fully AI-generated content the same way, something we've covered elsewhere. A verification-free policy offers more protection to the institution than to the principles of openness. Clear, numbered disclosure items give writers a definitive guide, a far better option than vague wording. Institutions still working out their own approach can look at how NIH applies detection tools to grant proposals.

FAQ

Does a 0.1% disclosure rate mean only 0.1% of papers actually used AI, and is this the same debate as whether AI detectors are accurate?

Not really, on either count. The 0.1% number just tracks how many papers said out loud, in the text, that AI helped write them. A separate check, the text-pattern analysis across all 5.2 million papers, found AI-assisted writing far more common and climbing steadily since 2023. Usage and compliance have split apart almost entirely. And no, this study isn't asking whether tools like TheChecker.AI catch AI-generated text after the fact, that's a different question covered elsewhere on this site. This one sits upstream of detection: are authors following the disclosure step their own journal already requires, regardless of how good any detector gets at catching an undisclosed passage later. Weak detection and weak disclosure feed each other, but they're not the same problem.

Did the follow-up critique disprove the original study?

Yibo Wang's June 2026 correspondence challenged the authors' causal approach, arguing that a policy snapshot masks true adoption timing and that different policy types shouldn't be merged into one binary indicator. He and Bu's reply admitted certain measurement constraints yet upheld their core finding: at the broad level, journals with policies show no lower AI-writing growth than journals without, and public disclosure stays uncommon across the board either way. The consensus is clear: AI-assisted writing is accelerating and disclosure practices aren't keeping up.

Check your own manuscript before a reviewer does

If you used AI to help compose, translate, or refine a manuscript and can't tell whether you've crossed your journal's disclosure line, get the segment-level breakdown before you submit, not after a reviewer asks. Run a paragraph or a full draft through our free demo and see exactly which sections read as AI-assisted, instead of guessing at a policy line from memory.