The scientific publishing apparatus is under unprecedented strain. Driven by the relentless pressure to “publish or perish” and further accelerated by generative AI tools that make churning out manuscripts easier than ever, journal submissions are surging. Meanwhile, the supply of qualified human researchers willing to spend unpaid hours evaluating papers has reached a breaking point.
Faced with this bottleneck, journals, conference organizers, and researchers are turning toward an obvious, if controversial, candidate for relief: artificial intelligence itself.
Recent experiments and pilot implementations, however, reveal a complex reality: while AI offers undeniable utility for targeted editorial checks, it remains ill-equipped to replace human critical judgment—and introduces entirely new vulnerabilities to the scientific record.
The Promises: Where AI Excels
AI tools are proving genuinely useful when assigned narrow, labor-intensive, and standardized tasks that human reviewers often miss or lack the time to conduct thoroughly:
- Methodological and Statistical Screening: Large language models (LLMs) and specialized algorithms can rapidly scan manuscripts to check whether standard reporting guidelines were followed, verify basic statistical consistency, flag missing controls, or identify non-replicable code structures.
- Literature Mapping and Citation Verification: AI excels at scouring vast databases to flag missing foundational references, detect hallucinations or fabricated citations, and ensure that cited works actually support the claims being made.
- Detecting Red Flags and Malpractice: Machine-learning detectors are increasingly employed at the desk-review stage to catch image duplication, textual plagiarism, and signs of paper-mill manufacturing.
- Structuring Feedback: For non-native English speakers or overworked reviewers, LLMs can help synthesize notes, organize clear section-by-section critiques, and ensure tone consistency.
When deployed as an “editorial assistant,” AI has the potential to handle technical triage, allowing human reviewers to focus their scarce time on substance.
The Pitfalls: Critical Vulnerabilities
Despite these capabilities, researchers studying automated peer review warn that today’s systems carry profound flaws when tasked with evaluating scholarly merit:
- Superficial Critiques and Hallucinations: Tests of AI reviewers demonstrate that while they produce articulate feedback, roughly a quarter or more of their substantive criticisms can be factually wrong, irrelevant, or based on fabricated details.
- The “Paper Laundering” Problem: AI review systems can be easily “gamed.” Studies show that running a weak manuscript through an LLM to align its vocabulary and phrasing with what the evaluating model favors—a practice researchers dub paper laundering—can artificially boost scores and acceptance rates without improving the underlying science.
- Inability to Assess True Novelty: High-impact science often challenges prevailing consensus. Because generative models predict patterns based on existing training distributions, they tend to reward conformist, conventional work and penalize genuinely disruptive or unconventional methodologies.
- Confidentiality and Data Ownership: Uploading unpublished, proprietary manuscripts to commercial AI platforms raises significant ethical and legal concerns regarding data privacy, intellectual property theft, and non-consensual model training.
- Amplifying Bias: Automated evaluators risk institutionalizing historical citation and geographic biases, systematically favoring well-known institutions and prevalent paradigms over work emerging from underrepresented regions or niche fields.
The Verdict: An Assistant, Not an Arbiter
The emerging consensus among publishers and computer scientists is clear: AI cannot solve the peer-review crisis on its own.
Automating the entire review pipeline risks flooding the scientific literature with homogenized, superficially polished, but methodologically compromised work. Instead, the most sustainable path forward positions AI strictly as a triage tool—relieving reviewers of administrative and baseline compliance burdens while preserving human expertise as the ultimate judge of novelty, significance, and scientific integrity.