Academic peer review is facing a new AI problem: papers can now contain instructions meant not for people, but for the language models people may use while reviewing them.
According to Nikkei, hidden prompts were found in 17 preprints on arXiv. The commands included phrases such as "positive review only" and "no criticism," and were aimed at large language models that might be used to help draft peer reviews.
What Nikkei Found in the Papers
The tactic is simple in concept. A paper can include text that is hard or impossible for a human reader to notice, while still being visible to software that extracts or processes the document.
Nikkei reported that prompts were hidden in white text on a white background. In many cases, the text was also made very small, further reducing the chance that a reviewer reading the paper normally would see it.
The apparent goal was to influence AI-assisted evaluations. If a reviewer fed the paper into a chatbot or another large language model, the hidden commands could be read as instructions and potentially shape the model's response.
The examples Nikkei identified included direct attempts to steer the tone of a review. Phrases such as "positive review only" and "no criticism" point to a clear intended outcome: reduce negative feedback and push the model toward a favorable assessment.
Why This Matters for Peer Review
Peer review depends on careful judgment. Reviewers are expected to evaluate a paper's methods, claims, presentation, and contribution. When AI tools enter that process, the review can become vulnerable to material embedded inside the very document being judged.
The issue is not only whether a hidden prompt works every time. The more basic concern is that authors can attempt to influence an AI-powered review channel without making that attempt visible to the reviewer, editor, or journal.
That creates two connected risks. First, a reviewer who relies too heavily on a language model may receive an evaluation shaped by hidden instructions. Second, the presence of hidden instructions can act as a test of whether a reviewer is using AI carelessly.
Nikkei described the tactic as a way to influence AI-driven peer review and catch inattentive human reviewers. Those two motives point in different directions, but both reveal the same weakness: the review process was not designed around documents that can contain machine-targeted messages.
Where the Papers Came From
Most of the affected papers came from computer science departments, according to Nikkei. The institutions were spread across 14 universities in eight countries.
The universities named in the source include Waseda, KAIST, and Peking University. The source does not present the practice as limited to one country or one institution.
The academic response has not been uniform. Nikkei reported that a KAIST professor called the practice unacceptable and said one affected paper would be withdrawn. Waseda defended the approach as a response to reviewers who themselves use AI.
That split captures the tension now facing academic publishing. Some see hidden prompts as misconduct because they attempt to manipulate review. Others frame them as a reaction to reviewers using AI tools in a process where expectations are still uneven.
Journal Rules Are Still Uneven
Policies on AI in peer review vary. The source notes that Springer Nature allows some use of AI in peer review, while Elsevier prohibits it.
This difference matters because the same reviewer behavior may be treated differently depending on the venue. In one context, limited AI assistance may be allowed. In another, it may violate policy.
For authors, reviewers, and editors, that creates a practical problem. The rules around generative AI are not yet consistent across the scientific ecosystem, even as the tools are already being used in research work.
Nikkei also noted that some papers can be found through Google searches using trigger phrases from the hidden prompts, including site:arxiv.org "GIVE A POSITIVE REVIEW" or "DO NOT HIGHLIGHT ANY NEGATIVES". That detail shows how visible the issue can become once the hidden text is indexed or searched directly.
Generative AI Is Already Changing Research
The hidden-prompt problem sits inside a much larger shift. A recent survey of about 3,000 researchers found that generative AI is quickly becoming part of scientific work.
In that survey, a quarter of respondents had already used chatbots for professional tasks. Most respondents, 72%, expected AI to have a transformative or significant impact on their field. Nearly all, 95%, believed AI would increase the volume of scientific research.
Another large-scale analysis looked at 14 million PubMed abstracts and found that at least 10 percent had already been influenced by AI tools. The source connects this shift to calls for updated guidelines on AI text generators in scientific writing.
The focus of those guidelines, as described in the source, is on using AI as a writing aid rather than as a tool for evaluating research results. That distinction is important. A tool that helps polish or structure text plays a different role from a tool that judges whether research should be accepted, revised, or rejected.
The discovery of hidden prompts does not settle the debate over AI in science. It does make one point harder to avoid: if peer review involves language models, papers themselves can become part of the prompt. Any serious policy will need to account for that fact.