AI reviewers are here — and the scientific community is not ready. Preprint platforms have long served as the agile alternative to the slow, rigid process of traditional academic publishing. Now, openRxiv, the nonprofit behind bioRxiv and medRxiv, is pushing the envelope with its latest innovation: integrating an artificial intelligence-powered reviewing tool into its preprint system. Developed by q.e.d Science, a startup based in Tel Aviv, the AI tool delivers feedback on biomedical manuscripts in as little as 30 minutes. It evaluates originality, detects logical inconsistencies, and suggests improvements to methodology and writing. The appeal is clear. For scientists who have endured months of waiting for peer review, or deciphered cryptic and often unhelpful critiques from anonymous reviewers, AI-generated feedback promises speed, consistency, and impartiality. Large language models can analyze manuscripts instantly, flag plagiarism, verify citations, and assess statistical rigor—tasks that could free human reviewers to focus on the most complex and creative aspects of science. But efficiency is not the same as validity. Peer review serves two essential roles. First, it validates the vast majority of scientific work—careful, incremental studies that test hypotheses and fill knowledge gaps by ensuring methodological rigor, sound statistics, and logical coherence. Second, it evaluates the rare, groundbreaking discovery that challenges existing paradigms. This second function requires more than rule-checking; it demands judgment, context, and the ability to recognize when established rules no longer apply. While humans are not perfect, they can, in theory, perform both tasks. AI, however, is fundamentally limited. LLMs excel at pattern recognition and synthesis based on existing data, but they lack true understanding, creativity, and the capacity to question assumptions. They are trained on past research, meaning they are more likely to reinforce the status quo than to recognize revolutionary ideas that break from it. A 2024 study using GPT-4 confirmed this concern: the model was highly effective at predicting the average response of human reviewers, but it failed to capture the diversity of expert opinion. In other words, it produces a consensus—but not necessarily a correct or insightful one. This “regression to the mean” effect risks homogenizing feedback, discouraging novel approaches, and favoring conventional thinking. Worse, the system is already being exploited. Scientists have begun embedding hidden messages and subtle cues in their manuscripts to manipulate AI-generated reviews, a practice that undermines the integrity of the process. If AI can be gamed, it may not just be inefficient—it could be misleading. As AI becomes more embedded in scientific publishing, the community must ask not just how fast peer review can be, but how well it serves science. The goal should not be to replace human judgment with machine output, but to use AI as a tool that enhances human expertise—flagging errors, saving time, and allowing reviewers to focus on what only humans can do: assess significance, interpret novelty, and challenge the boundaries of knowledge. Without careful oversight, the promise of AI review may deliver not progress, but a new kind of scientific inertia.