An analysis of tens of thousands of research-paper submissions has revealed a sharp rise in the use of artificial intelligence (AI) in scientific writing, raising questions about disclosure and accuracy in peer-reviewed research, according to reporting by Nature.
The American Association for Cancer Research (AACR) examined 46,500 abstracts, 46,021 methods sections, and 29,544 peer-review reports submitted to its journals between 2021 and 2024. Using a tool developed by Pangram Labs, the publisher found that 23% of abstracts and 5% of peer-review reports submitted in 2024 contained text likely generated by large language models (LLMs) such as ChatGPT. Despite this, less than a quarter of authors disclosed their AI use, even though disclosure is mandatory.
Daniel Evanko, AACR’s director of journal operations and systems, told Nature that the results were “shocking” and highlighted the need for systematic screening. While the use of AI in peer-review reports briefly declined after the AACR banned LLMs for reviewers in late 2023, detections doubled again by early 2024.
Pangram Labs’ tool, trained on 28 million human-written documents and AI-generated text “mirrors,” achieved an estimated 99.85% accuracy, reducing false positives dramatically through repeated retraining and active learning. The tool can also differentiate between outputs from different LLMs, including ChatGPT, Claude, LLaMa, and DeepSeek, though it cannot distinguish fully AI-generated passages from human-edited text that used AI.
Evanko expressed concern about widespread AI use in methods sections, where subtle changes could introduce errors. Authors from non-native English-speaking countries were more than twice as likely to employ AI in their submissions. Manuscripts flagged for AI-generated abstracts were twice as likely to be desk-rejected by editors before peer review.
The analysis also revealed a significant gap in disclosure: in early 2025, 36% of manuscripts were flagged for suspected AI-generated abstracts, yet only 9% of authors acknowledged using AI. Mohammad Hosseini, an expert on research ethics at Northwestern University, told Nature that these findings should be interpreted with caution, noting that the tool cannot detect AI use for data analysis or image generation. Still, he emphasized that journals need to respond to policy violations to maintain credibility.
