AI evidence extraction

Reflections on the impact of artificial intelligence on peer-review practices and its implications for greener scientific evaluation

Adrián Fuente-Ballesteros, Vânia G. Zuin Zeidler · Green Analytical Chemistry · 2026

AI-generated evidence extraction, verified across multiple analytical personas. Not a substitute for the peer-reviewed original.

9/10

Relevance

2/4

Quality (LMQS)

Evidence

Citations

11.60

FWCI

Top 10%

Impact

This is an AI analysis. Read the peer-reviewed original at the publisher: https://doi.org/10.1016/j.greeac.2026.100349

Methodology & findings

Study design

Narrative review synthesizing literature on AI use in peer-review, editorial policies, case studies of AI tools (Eliza, Enago Read, Veracity), discussion of empirical studies (NeurIPS experiment, Wiley survey, Liang et al., Nowak et al.), and hermeneutic analysis of structural conditions, behavioral patterns, and consequences of AI adoption in scientific publishing..

Primary method

No original statistical analysis conducted. The paper reviews and synthesizes findings from cited studies that employ survey methods, experimental comparisons, and bibliometric analysis.

Main result

The paper identifies that "approximately 19% of respondents had already experimented with large language models (LLMs) to streamline and facilitate their peer-review activities" according to a Wiley survey of nearly 5000 researchers. Key findings include that AI-assisted reviews often lack concrete references to manuscripts, contain generic language, show mechanical enumeration of comments, and may generate incorrect or incomplete conclusions. The study notes that "the contrast with the work behind the manuscript could hardly be stronger. Authors often spend months, and sometimes years, designing experiments, securing funding, repeating failed tests, analysing data, and building a coherent scientific argument" while AI reviews can be completed "in only a few minutes, in some cases around five minutes."

Reports effect sizes.

Research paradigm

Critical hermeneutics / interpretive analysis

Author conclusions

The authors conclude that "the challenge is not to control AI, but to stop the system from turning into a chain of opaque filters that reward what is predictable and penalize what breaks with established paths. The key issue is no longer whether AI tools will enter peer-review, but what kind of science will result once automation shifts from an occasional aid to the main structure of the whole process." They recommend that "AI should not be rejected but carefully integrated through transparent and fit-for-purpose review frameworks" and call for publishers to "provide enough time for proper and more sustainable peer-review processes, with realistic deadlines, and to expand the reviewer pool to distribute the workload and reduce regional bias."

Risk of bias

Publication bias: only open/public reviews accessible for analysis; Selection bias: editors' subjective interpretation of AI policy compliance; Geographic bias: reviewer selection algorithms repeat existing patterns favoring United States, Europe, Australia; Language bias: researchers whose first language is not English experience disadvantages; Status bias: well-known researchers receive different evaluations than early-career researchers; Structural inequality: AI training data reflects existing inequalities in representation; Selection bias in cited studies and examples (predominantly from well-resourced, English-speaking contexts); Confirmation bias in framing AI adoption as problematic; Geographic bias: discussion predominantly references North American and European scholarship; Potential publication bias in reviewed studies showing negative AI impacts; Funding and publisher bias: major publishing companies (Elsevier, Wiley, Springer Nature) are cited as primary policy sources; Selection bias in discussion of published case studies and available literature on AI in peer-review; Potential publication bias favoring studies showing problems with AI rather than beneficial applications; Geographic and institutional bias in reviewer selection noted by authors: reviewers "frequently include mainly researchers from the United States, Europe, or Australia, with very few reviewers from other regions"; Language bias affecting non-native English speakers and underrepresented languages in training data; Status bias in peer-review that may be amplified by AI systems trained on literature reflecting existing inequalities

Open questions raised

Lack of established syllabi and pedagogical frameworks for teaching AI in scientific disciplines
Need for published AI-assisted method assessment protocols clarifying when and how AI can improve evaluation quality
Need for mechanisms to detect and measure prevalence of AI-assisted reviews across the publishing industry
Absence of research on whether AI adoption has increased reviewer willingness to accept invitations
Limited understanding of whether LLMs correctly understand technical language and conventions in analytical chemistry
Need for development of pre-check filters or quality verification systems for reviewer comments

Extracted from: pdfAgreement 69%

Explore related topics

AI Peer Review Peer Review Position Paper