AI is advancing at a rapid pace: the reasoning capabilities of Large Language Models (LLMs) are improving; multimodality enables them to simultaneously analyze text, images, and other sources of data, and access to web resources is increasing their ability to retrieve information, access recent knowledge, and self-correct with autonomous fact-checking. Vigorous efforts are also underway to overcome fundamental limitations of auto-regressive LLMs, such as new ways to learn causal “world models”, to improve understanding of scientific concepts and AI model robustness (Richens and Everitt, 2024).

Along with these expanded capabilities of AI systems, applications dedicated to critically analyze and review scientific papers have become available as part of research projects (Bougie and Watanabe, 2024; D’Arcy et al, 2024), commercial platforms (Nature Research Assistant, https://natureresearchassistant.com; AlchemistReview, https://www.hum.works/review), specialized startups (qed, https://www.qedscience.com, Reviewer3, https://reviewer3.com) or even as self-described ‘weekend projects’ by prominent AI engineers (Andrew Ng’s agentic reviewing AI, https://paperreview.ai).

A major incentive of these efforts is to address several problems and bottlenecks of human-based peer-review (Mann et al, 2025). It is a time-consuming activity that is carried out by already overcommitted researchers and authors often have to wait for weeks or months for a decision on their manuscript. Finding reviewers with suitable expertise who are willing to evaluate manuscripts is increasingly challenging for journal editors. Moreover, human evaluation can be variable, affected by individual emotions, politics, and biases that are difficult to measure.