Researchers are using AI against the rules, because being left behind is worse
At a major 2026 conference, more than one in five reviewers who were told not to use an LLM used one anyway. The fear driving that is rational.
What the evidence shows
ICML 2026 ran an experiment on its own reviewers. The conference handled over 24,000 papers with 17,000 reviewers. Some reviewers were assigned a policy that prohibited all LLM use and others a policy that allowed limited help. Afterward, 1,486 of them answered an anonymous survey.
| Rule | What people reported doing | Source |
|---|---|---|
| No LLM use at all in reviewing | 22.5% used one anyway | Kim and coauthors, ICML 2026 study |
| Limited LLM help allowed | 36.5% did at least one thing the policy explicitly disallowed | Same study |
| Disclose AI use when you submit | One third have never disclosed it | Springer Nature and TBI survey of over 1,000 researchers |
| Review without AI, by default at many journals | 53% of reviewers have tried AI during peer review | Frontiers survey of 1,645 researchers |
These are self-reports from people who chose to answer. The true figures are unlikely to be lower.
Why the rules lost
The ban changed nothing anyone could measure. The ICML authors found that policy assignment had near-zero effects on final paper decisions, paper scores and reviewer confidence. A rule that costs the people who follow it and changes no outcome does not hold.
Only bad use gets caught. Careless AI output gives itself away, the way early image generators drew hands with six fingers. Stock phrases, invented citations and hidden prompts are the six fingers of a manuscript, and those are what journals catch. Careful use leaves nothing to find: a review, a paragraph or a coefficient does not show how it was made.
The young are not waiting. In the Frontiers survey, 87% of early-career researchers already use AI tools. Among reviewers with five years of experience or less, 61% use AI regularly. Among those with fifteen years or more, 45% do. The people entering the field treat this as the normal toolkit.
Where this goes next: your data
Peer review rules carry a reputational penalty. Data licenses carry a harder one. University libraries say plainly that their contracts forbid putting licensed content into AI systems, and the University of Waterloo warns that a breach “can result in loss of access for individual users or the entire campus.”
The same forces apply. The rule is hard to enforce, the tool is on every laptop, and the researcher who complies watches others move faster. It would be naive to think every one of them stops at the license boundary.
What we are saying, and what we are not
We are not telling you to break a rule. A data license is a contract, and the penalty for breaching it falls on everyone at your institution.
We are telling you three things.
- Your competitors have a head start. Some earned it and some took it. Either way it is real, and it grows every month you wait.
- There is a way to get most of the speed inside the rules. Keep licensed rows out of the model’s context and let the agent write code that you run. Our guide shows how.
- Rules nobody follows get rewritten by whoever shows up. If your library, your department and your journals hear nothing from you, the next version will be written for the vendor’s convenience.
The comfortable position, compliant and unhurried, no longer exists. You can be fast and careful, or you can be behind.
Sources
- arXiv arxiv.org
- Springer Nature, researchers embrace AI and the disclosure gap springernature.com
- Frontiers, AI in peer review frontiersin.org
- University of Waterloo Libraries, use of library resources with AI uwaterloo.ca