The system becomes gameable: a 5-minute rewrite can flip a rejection
Once AI reviews are connected to the submission pipeline, authors can learn to manipulate them. A 2026 study demonstrated a simple, low-cost attack: superficially rephrasing a manuscript's abstract, without changing any scientific content, improved AI review outcomes across disciplines and venues [2]. The strongest version of this attack achieved a success rate of about 38%, raising acceptance ratings by +1.31 for one AI reviewer model and +0.88 for another on a 10-point scale [2]. When the original AI review said 'reject', the attack flipped the decision to acceptance more than 50% of the time [2]. The attack took only about 5 minutes and $1 for a 10-page conference submission, and it was hard to distinguish from ordinary editing [2]. This means that connecting AI to peer review creates an incentive for authors to optimize for AI judgment rather than scientific merit—a risk that doesn't exist when reviews are purely human.
This gaming risk is not just theoretical; it's a direct consequence of AI's role as an evaluator. The same study found that inflated AI reviews could bias downstream human decision-making, shifting editorial recommendations from rejection toward acceptance [2]. So the operational risk isn't just that AI scores can be gamed—it's that those gamed scores then influence human editors, compounding the problem.
The review process can drift toward proxy optimization, not truth
A second risk emerges only after AI is embedded: the review system may start rewarding proxies for quality rather than actual scientific truth. A 2026 paper argues that AI-assisted peer review should be 'verification-first'—meaning it should generate auditable verification artifacts—rather than simply mimicking human review scores [4]. The authors formalize a 'phase transition' where, as verification pressure increases and signal shrinks, rational effort shifts from truth-seeking to proxy optimization, even when current decisions still appear reliable [4]. In plain terms: once authors know AI scores matter, they'll optimize for those scores, and the scores may not track real scientific value.
This risk is distinct from gaming because it's not about malicious manipulation—it's about the system's design. The paper's model shows that without 'truth-coupling' (how tightly scores track actual truth), the review process can collapse into a proxy-sovereign state where everyone is optimizing for the wrong thing [4]. The authors recommend deploying AI as an adversarial auditor that expands verification bandwidth, not as a score predictor that amplifies claim inflation [4].
Human reviewers lose autonomy, and ethical blind spots appear
A third risk is to the human reviewers themselves. A 2025 qualitative study of 15 nursing academic reviewers from four countries found that while they saw benefits in AI (efficiency, fairness, workload reduction), they also raised concerns about transparency, bias, data privacy, and—critically—risks to reviewer autonomy and judgment [1]. The reviewers expressed cautious optimism but insisted that AI must not replace human judgment [1]. This suggests that connecting AI to tools can erode the very human oversight that is supposed to keep the process trustworthy.
This ethical dimension is echoed in earlier work. A 2021 study that trained an AI to predict review scores from manuscript text found that such techniques could reveal correlations between the decision process and other quality proxies, uncovering potential biases in the review process [5]. But the same techniques raise concerns about algorithmic bias and unintended consequences [5]. So the risk isn't just about scores—it's about the system's ability to replicate and even amplify existing biases, which becomes a live operational problem once AI is in the loop.
About These Sources
This answer is built on 5 studies (2 peer-reviewed, 3 preprints) — published from 2021 to 2026, 4 from 2024 or later, 1 in Q1 journals, collectively cited 242 times — selected as the most relevant from 5 studies that passed quality screening, drawn from 37 papers retrieved from a database of over 500 million.
Sources used in this answer
Nursing Academic Reviewers’ Perspectives on AI‐Assisted Peer Review: Ethical Challenges and Acceptance
In a qualitative study of 15 nursing academic reviewers from four countries, participants identified ethical concerns (transparency, bias, data privacy) and risks to reviewer autonomy and judgment, emphasizing that AI must not replace human judgment.
Gaming AI-Assisted Peer Reviews Poses New Risks to the Scientific Community
A 2026 study demonstrated that superficial rephrasing of a manuscript's abstract (without changing content) improved AI review outcomes, with the strongest attack achieving ~38% success and flipping 'reject' to acceptance in >50% of cases, at a cost of ~5 minutes and $1.
AI-Assisted Peer Review at Scale: The AAAI-26 AI Review Pilot
The AAAI-26 AI Review Pilot, the first large-scale field deployment, generated AI reviews for all 22,977 full-review papers in less than a day, and survey participants preferred AI reviews to human reviews on technical accuracy and research suggestions.
Preventing the Collapse of Peer Review Requires Verification-First AI
A 2026 paper argues for verification-first AI in peer review, formalizing a phase transition where rational effort shifts from truth-seeking to proxy optimization, and recommends deploying AI as an adversarial auditor rather than a score predictor.
AI-assisted peer review
A 2021 study trained an AI on 3,300 papers and reviews, showing that AI can predict review scores from text and uncover biases, but also raising concerns about algorithmic bias and unintended consequences.