The rise of generative models changes what peer reviewers must evaluate: not only the written prose but the underlying models, training data, and code that produced scientific claims. John P. A. Ioannidis Stanford University documented systemic weaknesses in reproducibility that predate AI and that AI can amplify if unchecked. Review processes must therefore center transparency and reproducibility to preserve trust.
Verification and transparency
Peer review should require authors to disclose the role of AI, the model version, prompts, and preprocessing steps, and to deposit code and data in accessible repositories. Victoria Stodden University of Illinois at Urbana-Champaign has argued for standards that treat computational artifacts as first-class research outputs; reviewers should have access to runnable code or containers to replicate claims. Human expertise remains essential: reviewers need domain knowledge to assess whether AI outputs reflect valid reasoning or superficial pattern matches, and computational expertise to validate pipelines. Editorial policies should require an audit trail for automated analyses and declare limitations of probabilistic outputs that can hallucinate plausible but false statements.
Ethical, cultural, and environmental checks
AI introduces biases tied to training corpora and uneven geographic representation. Emma Strubell University of Massachusetts Amherst highlighted the environmental costs of large-model training; journals should weigh the environmental impact and encourage efficient methods or shared, audited models to reduce duplicate training. Peer review must also respect cultural and territorial data sensitivities by validating consent and provenance for datasets drawn from Indigenous, marginalized, or local communities. Ethical review boards and editors should coordinate to ensure AI-assisted findings do not compound historical inequities.
Practical changes include recruiting computational reviewers, adopting standardized AI disclosure forms, and enabling post-publication replication badges that incentivize verification. Publishers and institutions can share trusted verification infrastructures to reduce barriers for researchers in resource-limited settings, addressing territorial disparities in capacity. Magdalena Skipper Editor-in-Chief Nature advised that AI cannot be treated as an author and that human accountability must be explicit; peer review should therefore document responsible human oversight.
Incorporating these measures aligns peer review with modern scientific practice: it preserves human judgment, enforces verifiable computational standards, mitigates bias and environmental harm, and protects communities whose data underpin scientific advancement.