Peer Review With AI Assistance: Confidentiality Comes First
Using AI for academic peer review raises serious confidentiality concerns as pasting unpublished manuscripts into third-party tools violates disclosure agreements.
Stock photo for illustration only, not from the actual event
- Under review manuscripts and grant proposals are strictly confidential and unpublished documents.
- Transmitting documents to external services constitutes unauthorized disclosure regardless of training policies.
- AI-generated reviews often look polished yet lack genuine expert analysis and specific critique.
- Major funding bodies like the US NIH have already prohibited generative AI tools in grant reviews.
Most discussions surrounding artificial intelligence in peer review focus heavily on whether the resulting reviews are actually any good. However, that is secondary. The primary concern is that a manuscript under review is someone else's confidential and unpublished work, and pasting it into an external service represents a disclosure you were never entitled to make.
When you accept a review invitation, you formally agree to a confidentiality undertaking. Manuscripts typically contain experimental results where authors have yet to establish academic priority, and in grant reviews, they contain unfunded research plans that stand as some of the most commercially and academically sensitive documents in the system. You explicitly agreed not to share them with anyone.
Submitting sensitive documents to a third-party service means sharing them directly with an outside entity. Data retention and training policies only dictate the severity of the breach, not whether a breach actually took place, since the core obligation is absolute non-disclosure.
Stock photo for illustration only, not from the actual event
This fundamental argument does not rely on model quality, hallucinations, or algorithmic bias. The exact same issue would apply to a flawless AI system, which explains why this specific concern drives actual institutional policy and cannot simply be solved by deploying better models. Resolution requires shifting where the computation occurs, utilizing infrastructure already covered by existing confidentiality agreements.
"Generated text can simulate the prose and cannot supply the judgement."
Dev.to
A proper review represents a named expert's professional judgment. Its true value to an editor is not merely clean prose, but the assurance that a knowledgeable specialist read the paper and formed a defensible perspective. Generated text can easily simulate fluent prose, but it cannot supply genuine analytical judgment, resulting in review-shaped objects that sound reasonably thorough while failing to engage with specific data points.
Major research funders have moved fastest and hardest because grant applications represent the most sensitive documents in the entire academic ecosystem. The United States National Institutes of Health prohibited peer reviewers from utilizing generative AI tools to analyze and formulate critiques of grant applications specifically on confidentiality grounds. Meanwhile, journal and conference policies remain fragmented, requiring reviewers to carefully read venue-specific guidelines before accepting any review assignments.
Source: Dev.to
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment