AI Policies Fail to Reduce Undisclosed AI Use

Following the release of generative AI models like ChatGPT, many researchers began exploring these tools in their research processes, including preparing manuscripts. Most journals bar AI from being listed as an author, but many permit certain uses of AI in the writing and editing process, though the majority require this use to be disclosed.

Yi Bu, an information scientist at Peking University, was interested in how these policies influenced AI use in academic publishing. He and his coauthor Yongyuan He, a computer scientist at Peking University, compared the suspected AI usage in published articles from journals with and without AI policies. In a study published in the Proceedings of the National Academy of Sciences, they found that most articles with evidence of AI usage did not disclose this in the manuscript, regardless of the journal’s policy.¹

“The fact that AI usage has surged increasingly regardless of policy tells us that researchers are voting with their keyboards. They find value in these tools for overcoming language barriers, and current policies from journals are not a strong enough guide,” Bu said. “Our study is a fundamental piece that opens numerous critical avenues for future research.”

Bu and He used a large language model to categorize more than 5,000 journals in the Journal Citation Reports, a publication from Clarivate Analytics that provides metrics for leading journals, based on each journal’s guidelines regarding permitted AI use. The majority of journals indicated that AI could be used for writing and editing support, and more than 60 percent allowed its use in language and grammar assistance. Other categories that the model included were “reference and citation support,” “content creation and generation,” and “translation services," though fewer journals mentioned these in their guidelines.

The researchers then evaluated more than one million full-text papers for likely AI content using a method to estimate the probability of AI generated text. They found that a similar proportion of papers had suspected AI content in journals with and without AI policies, with this amount increasing in the last three years. They also showed that papers with authors listed from non-English speaking countries had greater amounts of probable AI content compared to papers by authors from English-speaking countries. Bu and He said that this could indicate AI’s use in helping lower translation barriers, but they acknowledged that their model was limited in its ability to distinguish AI content introduced during language polishing and outright AI text generation.

While most journals allowed specific AI use, more than 3,500 required this assistance to be disclosed. However, the researchers found that the majority of articles with suspected AI use did not disclose this use in the paper, regardless of whether the journal had a policy or not.

Alex Glynn, a data scientist focused on research literacy and communication at the University of Louisville, said that one of the challenges in studying this topic is that there isn’t an objectively correct way to perform these analyses. But with the researchers using a multipronged strategy to verify their findings, he said, “I think they’ve taken a better approach than some do.” However, he said that the method of categorizing the different uses of AI is questionable because there is so much potential overlap in the examples.

Bu and He noted that, although the disclosure rate in publications from the first quarter of 2025 remained low, the rate was higher than it was in 2023, increasing from approximately 0.1 percent to 0.43 percent. The researchers suggested that this increase shows a positive trend toward declaring AI use that the policies are helping foster overall. They added that residual hesitation toward this could be due to concerns regarding how disclosing AI use may impact the perception of scientists’ work, leading to what they called a “cautious approach.”

Bu said that, given the obvious adoption of AI amongst academics, publishers should shift to promoting responsible integration with better infrastructure to detect and validate ethical AI use, and researchers should receive more education and training on these topics as well. Bu and his team are developing models to better detect and classify AI use in research publications and hope to complement their quantitative findings with human surveys studying how scientists use AI in research.

Glynn said that the study adds a valuable, quantifiable estimate to undisclosed AI use, agreeing with the authors that the use is likely even higher than what the study calculated. However, he disagreed with their conclusion rationalizing academics’ hesitation to indicate their AI use, likening it to failing to disclose conflicts of interest. “I wouldn't describe failure to disclose as a ‘cautious approach.’ I would describe it as a lie,” he said.

Additionally, Glynn disagreed with the conclusion that AI use in research is inevitable. “I don’t think that [the findings] make a case for AI detection software. I don’t think it makes a case for a particular avenue of further research.”

Instead, Glynn said that publishers should stop allowing ambiguity in their AI guidelines, such as not requiring authors to disclose “light editing,” which doesn’t have clear boundaries. Although addressing undisclosed AI use remains a challenge, he added that publishers can be more responsive to retracting papers with clear improper AI use in a timely manner.