Artificial intelligence is now firmly established in peer review practices, at the very heart of scientific research. An international survey conducted by the publisher Frontiers among approximately 1,600 researchers in 111 countries reveals that a majority of reviewers have already used AI tools to examine manuscripts submitted for publication. This rapid development, often at odds with official publisher recommendations, highlights a growing gap between actual practices and existing regulatory frameworks. According to the results published on December 11, more than half of researchers acknowledge having used AI in peer review. Nearly a quarter even indicate having increased this use over the past year. This growth confirms the entrenchment of tools based on large language models in academic practice, particularly to cope with the increasing workload and complexity of reviewed manuscripts. Research integrity officers at Frontiers have observed that this practice has become widespread, far exceeding the initial guidelines. Many publishers still advise against using external tools to process unpublished manuscripts, primarily due to confidentiality and intellectual property concerns. Yet, in practice, researchers have already integrated these technologies into their workflows, sometimes without formally disclosing it.
Between saving time and ethical grey areas
The survey provides a better understanding of the nature of this usage. A majority of the researchers involved use AI to help them formulate their evaluation reports, by structuring comments or reformulating their analyses. Others use it to quickly summarize an article, identify methodological weaknesses, verify references, or detect signals that may suggest a problem of scientific integrity, such as textual similarities or inconsistencies in visual data. These practices, however, raise numerous questions. Research ethics specialists emphasize that AI can facilitate certain technical tasks, but that it cannot replace human scientific judgment. Current tools excel at reformulation and synthesis, but still struggle to assess the true novelty of a work, the conceptual soundness of a hypothesis, or the relevance of an interpretation.
Several researchers have also tested the capabilities of these models in practice.
Recent experiments show that AI-generated evaluations often replicate the expected form and tone of an academic report, while remaining superficial. Factual errors, methodological approximations, and a lack of nuanced criticism are regularly observed, even when the tools are used with detailed instructions and a provided scientific context. Faced with this situation, publishers are under pressure. Some, like Frontiers, allow controlled use of AI, provided it is declared and limited to assistance functions. Other, more cautious publishers continue to express measured confidence in the actual contribution of these technologies to peer review. Parallel surveys conducted by other industry players also suggest that many researchers remain skeptical about AI's ability to substantially improve the quality of evaluations. For observers, the debate is no longer about whether AI is being used, but how it should be used. Calls are mounting to adapt editorial policies to this new reality, by defining clear rules, increasing transparency, and ensuring full accountability for human reviewers. Without this adjustment, there is a risk of opaque practices becoming entrenched, potentially undermining trust in a crucial pillar of scientific production.
