AI Writing, Journal Policies and Disclosure: Evidence and Limits
— Gatsbi
Journal policies and the spread of AI-assisted writing raise related questions: how writing practices are changing, what authors disclose, and whether particular rules achieve their intended purpose. A large observational study offers evidence about the first two, but cannot settle the third.
This report discusses Yongyuan He and Yi Bu’s preprint on academic journals’ AI policies. Originally posted on 7 December 2025, it was revised on 20 January 2026. We use the v2 full text, including its expanded discussion of measurement and policy limitations. It remains a preprint in the record checked for this update.
Editorial update, 18 October 2026: the original version of this blog overstated the findings as proof that journal policies were powerless. This revision distinguishes observed trends from causal policy effects and identifies the limits of text-based estimates. The original study illustrations are retained with explanatory captions.
What the Study Examined
The authors analyzed policies from 5,114 JCR Q1 journals and 5,235,012 papers published between January 2021 and June 2025. About 70% of the journals had an AI policy in the January 2025 policy snapshot, mostly requiring disclosure. A second policy collection in October 2025 documented further changes. These figures describe the sampled journals and collection dates, not all journals today.

Rising Text Signals Are Not Individual Usage Records
The study found rising AI-associated language signals after 2023. Its main measure uses word-frequency distributions to estimate AI content at group level, with additional methods used as robustness checks. This is different from observing the tools, prompts or edits used by individual authors.
The revised paper acknowledges possible false positives and cannot reliably distinguish language polishing from substantive text generation. A signal associated with AI therefore does not identify a particular author’s conduct or establish that a specific manuscript violated a journal rule. Reading the chart as a direct percentage of machine-written sentences would also misstate the measure.

Policy Comparisons Do Not Establish Causal Effects
The authors did not find a statistically significant difference in the growth rates of their estimated measure between journals grouped as having or not having an AI policy. That comparison does not establish that policies are effective or ineffective. It also does not show that every kind of disclosure requirement or prohibition has the same effect.
The v2 limitations explain why. Exact policy introduction dates were unavailable, publication follows writing and submission with a delay, and the analysis reduced different rules to a broad policy/no-policy distinction. A rule that permits editing while requiring disclosure has a different intended outcome from a ban on generating substantive text. An increase in permitted use would not by itself demonstrate failure.

Group Differences Need the Same Caution
The paper also compares countries, disciplines and journal open-access categories. These descriptive comparisons can suggest questions for further research; they do not establish why authors used a tool or whether the resulting work was sound. Institutional affiliations are not direct measurements of an author’s language proficiency, and an open-access category is not a quality assessment.


A Disclosure Gap, With Important Denominators
In a retrieved full-text sample, the authors identified 76 writing-assistance disclosures among 75,172 papers published since 2023, approximately 0.1%. That denominator is the sampled papers, not a verified count of AI users. The study also reports an increase in detected disclosures over time.
A low recorded disclosure rate does not tell us which individual papers required a disclosure or why a statement was absent. The text signals cannot separate uses governed by different rules, and the full-text retrieval process may favor more accessible publications. The v2 discussion recognizes both sampling limitations and the possibility that policies may encourage some authors to disclose. Neither observation resolves policy effectiveness causally.

What the Study Cannot Decide for Your Manuscript
Do not use this report to decide that a detector has identified misconduct, that a journal’s rules can be ignored, or that disclosure is unnecessary. It also does not compare writing products, measure their factual reliability or show that AI assistance improves acceptance chances.

Practical Implications for Authors and Editors
The following are editorial suggestions, not interventions tested by this study. Authors can keep a short record of the tool and version used, the writing task, the affected sections and the review they performed. Preserve the source manuscript and accepted revisions so coauthors can distinguish language changes from changes to methods, results or interpretation.
Check the current instructions for the intended journal and the applicable institutional requirements. Describe what the tool actually did and what the authors checked. If a requirement is unclear, resolve it with the journal rather than inferring an exemption from a population-level estimate.
For editors, useful questions include whether the policy distinguishes writing assistance from research activity, where disclosure should appear, and what evidence authors need to retain. Evaluating a policy would require an outcome matched to its purpose and a design that addresses when the rule took effect and how authors were exposed to it.
The study raises questions about changing writing practices and transparency. Its revised limitations matter as much as its headline results. Treat it as evidence for further investigation, while making manuscript decisions from the actual research record and the rules that apply to the submission.