AI in scholarly publishing

[Music] Welcome to Science Sessions, the podcast of the proceedings of the National Academy of Sciences, where we connect you with Academy members, researchers, and policymakers. Join us as we explore the stories behind the science. I'm Paul Gabrelson. With the advent of AI in academic publishing, publishers face the daunting challenge of regulating the use of generative artificial intelligence by authors who submit work to journals. Journals adopted a range of policies regarding AI use. How well are they working? In a recent PNAS study, Ibu and Young-Yuenha of Peaking University, surveyed journal policies and evaluated the proportion of scholarly articles likely containing AI-generated text. I.e. tell us about some of the policies academic journals have introduced to address generative AI. There's a clear consensus that generative artificial intelligence tools cannot be listed as an author or co-author on a scientific publication. Authorship entails some responsibilities, such as the ability to take public accountability for its content that the language model simply cannot fulfill. The dominant policy type, which we classified as this close required, explicitly permits the use of artificial intelligence. But also mandates that authors declare such use transparently. The stated permitted uses under these policies are often circumscribes, focusing on writing and editing support. A small number of journals have adopted a stance of strict prohibition, which explicitly bans the use of AI in any part of the managed-wrapped preparation process. Even smaller set of journals have an open policy, which allows the use of AI without requiring formal disclose, perhaps viewing it as a routine tool akin to a spell checker. And finally, a significant minority of journals fell into the not-mentioned category. They have no identifiable policy addressing generative AI at all. Tell us about the dataset you analyzed and your findings regarding journals policies. We systematically collected and analyzed the author guidelines and publication policies of over 5,000 scholarly journals from the Journal Citation Report, Q1, CatGrate, in the Sciences and Social Sciences. This collection was performed in two ways. One is in January 2025. The other one is in October 2025, in order to capture the rapid evolution of guidelines. We then employed large-language models like GPT-40 Mini and GIMINI 2.5 Flash, followed by our human verification, of course, to classify each journals policy into one of the four categories. One, strict prohibition, second, open policy, third, disclose required, and four, not-mentioned. Then, in order to measure actual author behavior, we built a corresponding compass of more than 5 million research papers published by these same journals between January 2021 to June 2025. And this data range is very crucial because it provides a clear pre-AI baseline and also captures the post-chagivity period. We also went beyond abstract and successfully retrieved and processed the full text of a representative sample of about 164 modern scientific publications. So these full text data sets allowed us to validate our AI detection metrics on a more substantial body of text than just abstract. And to manually and accurately evaluate the manuscripts for any explicit statements disclosing the use of AI writing tools. Did journal policies have any effect on AI usage? The direct answer is that journal policies in their current form have had no statistically measurable effect on the rate at which researcher adopt and use generative AI tools in the writing. The estimated AI content proportion in additional publications rose from a near zero baseline in early 2023 to substantial levels by mid-2025. This rise was virtually identical for the subset of papers published in journal data formal AI policies and the subset published in journal that did not. The search was universal. The existence of a journal policy is a footnote that did not alter the plot. To say policies had absolutely no effect would be an oversimplification because our data reviews a subtle effect on the act of formal disclosure. The number of papers is explicitly acknowledging AI use starting from essentially zero in early 2023 has been on a consistent upward trend about two-thirds of them were published in journal standard and AI policy. So these evidence indicates that policies are successfully guiding a small fraction of perhaps more risk of worse authors from the act of using AI to the additional act of declaring that use. How does the rate of AI disclosure compare with your estimates of probable AI use? We analyzed many full-text samples and we were able to directly contrast two numbers. The official declared world of AI use and the estimated real world of AI use. On one side we have the explicit disclosure rate. We scanned the full text of about 75,000 papers published from 2023 onward in our sample. We looked for clear and ambiguous statements in their methods, acknowledgements or other sections where authors credited an AI tool like ChatGPT for writing assistant. The result was very low, only 76 papers contained such a disclosure. This represents a mere 0.1% of the post-2023 publications we examined. On the other side we have the estimated AI usage rate. We do not see the 0.1 world. Rather we see a world where the statistical signature of AI influence is widespread and growing rapidly. What is the takeaway for journal editors and the scientific community at large? We need some mandatory structure disclosure forms at the point of submission. Also, there must be a serious investment in detecting and verification. Academic journals cannot outsource integrity to the goodwill of others. This is not about creating a normal system to reject papers, but about fostering a culture of verified transparency. What are the caveats and limitations of the study? A primary limitation of this work is the probabilistic nature of our AI detection. Currently there exists no perfect and deterministic AI detector. The most advanced methods now, including the called MLE method, which stands for maximum likelihood estimation approach, are still probabilistic. They estimate the likelihood that a body of text contains AI generated content. However, they cannot label any single centers or managed rep as AI written or not. The MLE method does not provide deterministic judgments about individual papers. The remains a possibility of false positives. For instance, technical fields often employ highly repetitive language that can statistically resemble AI output. Also, our method estimates the probability of AI influence, but cannot distinguish between a manuscript that used AI for minor grammar corrections and the one that used AI to generate entire paragraphs. Our evaluation of policy effectiveness is necessarily bland. In our analysis, we focus on about 5.2 million scientific publications. However, our more in-depth full-text analysis was necessarily limited to a sample of only about 164 thousand publications because of the challenges of mass PDF retrieval. So this sample, while still large, representative likely over-represents papers from fully open access publishers as their PDFs are more easily accessible online. The academic publishing pipeline is very long in many disciplines. Therefore, many of the papers in our copies, especially those from 2024 and 1995, were likely submitted before the journal's AI policies were enacted or widely communicated. The our findings of no policy effect may impart reflected slack, though the parallel girls trans deep into 25 suggest the lack alone cannot fully explain the total lack of divergence. Our study is limited to AI generated text, we did not investigate the use of AI to create or alter research images, figures. data sets or even cold. Our story is therefore, I think, incomplete, only capture one major vector of AI's impact on research integrity. Thanks for tuning into science sessions. You can subscribe to science sessions on iTunes, Spotify, or wherever you get your podcasts. If you like this episode, please consider leaving a review and helping us spread the word. [Music]