Scholarly peer review, the bedrock of scientific quality control, is facing a crisis. The sheer volume of research being produced is overwhelming human reviewers. Could AI language models (LLMs) be the solution? New research suggests that while LLMs can generate text that looks remarkably like human reviews, relying on them for peer review poses significant risks. Researchers have found that LLMs are vulnerable to manipulation. By subtly inserting hidden text into their papers, authors can trick LLMs into giving glowing reviews, potentially boosting the acceptance of subpar work. This manipulation can even influence rankings, pushing deserving papers down the list. Another issue is that LLMs tend to echo the limitations authors themselves disclose, potentially missing deeper flaws. While human reviewers critically assess a paper's weaknesses, LLMs might overemphasize minor, self-reported issues. Beyond manipulation, the research uncovered inherent flaws in LLMs. For example, some LLMs hallucinate, producing positive feedback even for empty papers! They also show biases towards longer papers and those from well-known authors or institutions. These findings raise serious questions about the fairness and reliability of LLM-driven peer review. While the idea of automated review is appealing, the research suggests we're not there yet. LLMs need more robust safeguards against manipulation and bias before they can be trusted with such a critical role. For now, human judgment remains essential for maintaining the integrity of scientific publishing. LLMs can be helpful tools, offering supplementary feedback, but they shouldn’t replace the nuanced evaluation of expert reviewers.