Stanford researchers found that GPT-4's review comments for Nature and ICLR papers matched human reviews over 50% of the time. Researchers inputted the PDF of a paper into GPT-4 to obtain review comments, which were then compared to human feedback. The similarity of GPT-4's reviews was over 82%, with authors believing the comments were helpful. Compared to humans, GPT-4 paid more attention to the paper's impact factor and showed less concern for ablation studies. Users generally believe that GPT-4's feedback could improve accuracy and reduce the workload of manual reviews.