ERC AI guidelines
The ERC AI guidelines are the European Research Council’s rules, issued on 24 March 2026, on the use of AI in the evaluation of ERC grant proposals. They draw two red lines for reviewers: the assessment of merit may not be delegated to an AI system, and the content of applications may not be uploaded to external AI tools. The guidelines restrict evaluation. They do not prohibit administration.
Most coverage repeats the two prohibitions. The useful question is the one almost nobody answers: what does that leave?
Administration is allowed. Evaluation is not.
Read as a boundary rather than a ban, the guidelines split the work of a review process into two categories.
Off limits under the guidelines
- Letting an AI system produce or determine the merit judgment: the score, the ranking, the funding recommendation.
- Placing applicants’ proposal content into external AI tools outside the controlled evaluation environment, confidentiality being the stated reason.
Still open: the administration side
- Eligibility and completeness checks before review starts.
- Deadline, document and panel logistics.
- Structuring and preparing the material reviewers read.
- Consistency checks on the process itself: was every proposal handled under the same procedure.
The operative sentence: a system that prepares the material and the process, while humans perform the assessment of merit, operates on the administration side of the line. The rule applies to ERC evaluations; other funders increasingly cite it, and your own program’s counsel has the final word. The primary source is the ERC itself at erc.europa.eu.
Built on the administration side of the line.
EvalLens was architected on this boundary before the guidelines named it. The AI Total Score produced by six independent judges is advisory and has no path into a ranking: leaderboards are built from human Jury Scores only, so merit assessment is never delegated. Reports prepare the material with cited evidence, and submissions are processed inside a closed perimeter, never used to train models. We do not certify anyone’s ERC compliance, your counsel does; what we offer is an architecture where the human side of the line is structural. AI prepares the analysis, people decide.