Large Language Models for Peer Review in Biotechnology

As researchers use Large Language Models (LLMs) for rapid manuscript feedback, a key question is whether they can function as reliable peer reviewers in biotechnology. This study tests AI peer review using 763 preprints (398 with open peer reviews) and 12 grant proposals. We find that AI reviewers (GPT-5, Qwen-Plus, and Gemini 2.5 Pro) all provide substantive and well-structured comments, with a strong emphasis on experimental design and statistical analysis; however, they tend to be more lenient overall than human reviewers. AI reviewers are less likely than humans to critique paper positioning or ask for more citations. LLMs often rate grant proposals more favorably than humans (i.e. clustering at 3.2~3.8 vs human average 2.5 in scale of 1-4) and have less variation in word choices. AI detectors failed to reliably identify AI-generated text in review comments, as simple rewording bypassed them and detectors usually lag behind fast-evolving LLMs. Our results suggest that: (1) AI can serve as a valuable and less biased ad hoc reviewer; (2) the use of public LLMs in peer review introduces privacy and copyright concerns; (3) It is important to develop a review agent capable of identifying AI-generated content and verifying that all scientific claims are rigorously evidence-based; and (4) clearer guidelines and sustained human oversight are essential, along with greater transparency through open peer review. Nevertheless, as artificial general intelligence continues to advance, future AI systems may match, or even surpass, human researchers in evaluating scientific manuscripts.