AI tool detects LLM-generated text in research papers and peer reviews (www.nature.com)

The American Association for Cancer Research (AACR) has revealed a significant rise in the use of AI-generated text in research paper submissions and peer reviews, identifying that 23% of abstracts and 5% of peer-review reports submitted in 2024 likely contained text produced by large language models (LLMs). This finding, based on an extensive analysis of over 46,000 abstracts and 29,000 peer-review comments using an AI detection tool developed by Pangram Labs, highlights both the growing reliance on AI in academic writing and the ongoing challenges in enforcing disclosure policies. Despite mandating authors to report AI usage, less than a quarter complied, raising concerns about transparency in scientific publishing. The detection tool, Pangram, stands out for its technical sophistication—trained on 28 million human-written documents combined with AI-generated "mirrors," it achieves an impressive 99.85% accuracy and a false-positive rate as low as one in 10,000, far surpassing existing detectors. Notably, Pangram can distinguish among texts produced by various LLMs like ChatGPT, DeepSeek, LLaMa, and Claude, owing to its uniquely self-generated training dataset. However, it cannot yet differentiate between fully AI-written passages and those merely edited using AI. The AACR’s analysis also revealed geographical disparities, with authors from non-native English-speaking countries more than twice as likely to employ LLMs, particularly in methods sections—raising concerns about potential accuracy risks when AI rephrases crucial experimental details. This development is significant for the AI/ML community as it underscores both the proliferation of AI tools in research workflows and the urgent need for robust, transparent detection mechanisms to maintain scientific integrity. Pangram’s success offers a promising approach for publishers aiming to balance AI’s benefits in language assistance with rigorous safeguards against misuse.