AI Peer Review Faces Scrutiny Over Bias, Cost, and Reliability

Key points

• Tyler Cowen reports that AI can audit the entire 100 million paper scientific corpus for approximately $500 million, identifying 90% of flawed work for a few hundred dollars.
• A study by Tech Policy Press found that four out of six OpenAI models accepted nearly 100% of anonymized manuscripts, contrasting sharply with the less than 25% acceptance rate at the source conference.
• Marco Rospocher’s study in Discover Artificial Intelligence shows that LLMs give higher scores to papers attributed to authors with high h-indices or prestigious institutional affiliations, even when the manuscript text is identical.
• The AAAI 2026 conference used AI to review 22,977 papers in under 24 hours, with respondents finding AI useful for technical accuracy but poor at judging novelty.
• Critics argue that AI models trained on human feedback may inherit biases, leading to 'narcissism of small differences' where minor methodological disagreements are amplified.

Background

The integration of AI into peer review has accelerated due to a surge in submissions, with AAAI receiving nearly 29,000 papers in 2026, double the previous year. This trend coincides with advancements in large language models (LLMs) capable of reasoning and writing. While publishers and editors advocate for AI to relieve the burden on overworked early-career reviewers, concerns about the reliability and ethical implications of automated evaluation have grown. Previous efforts to automate peer review, such as plagiarism checks, have existed since the 2010s, but the current wave involves more complex reasoning tasks.

How outlets are covering it

Marginal Revolution highlights the cost-efficiency of AI auditing, suggesting that a small fraction of the scientific corpus carries the majority of the signal, and that AI can identify 'nonsense' at a low cost. However, Tech Policy Press emphasizes the unreliability of AI for peer review, citing absurdly high acceptance rates and a lack of alignment with peer review values like unbiased critique. Bioengineer.org focuses on algorithmic bias, showing that LLMs are sensitive to author metadata, such as institutional prestige and bibliometric profiles, which can distort gatekeeping decisions. These perspectives highlight a tension between the efficiency and scale of AI and its potential to introduce or amplify biases, undermining the trustworthiness of scientific evaluation.

Why it matters

The reliability of peer review is foundational to scientific integrity. If AI systems are used to gatekeep research, their biases and errors could systematically exclude valid work or promote flawed studies, distorting the scientific record. The high acceptance rates and sensitivity to author metadata suggest that AI may not be a neutral arbiter, potentially reinforcing existing inequalities in academia. As AI becomes more prevalent in editorial pipelines, understanding its limitations is crucial for maintaining public trust in science and ensuring that peer review remains a fair and rigorous process.

What to watch

Researchers and institutions are likely to develop more robust guidelines and policies for AI-driven peer review, focusing on mitigating biases and ensuring transparency. There may be a push for hybrid models that combine AI efficiency with human judgment, particularly for evaluating novelty and impact. Additionally, studies may explore the long-term effects of AI on academic careers and the distribution of research funding, as well as the potential for AI to be manipulated by authors to game the system.