EU grant reviews need an AI contribution ledger
The European Ombudswoman welcomed tighter evaluator rules but asked for a record of whether, why and how AI contributed. Funding bodies can turn that request into a small, review-level evidence contract.
What happened
The European Ombudswoman closed case 2979/2025/MIK after recognising updated Commission rules, while asking for more detailed and proactive recording of AI use by external evaluators.
Why it matters
A binary disclosure cannot show whether AI only polished language or shaped evidence and judgment. A contribution ledger makes non-delegation, confidentiality and accountability inspectable.
A disclosure needs a contribution boundary
The European Ombudswoman recognised an improvement in the European Commission’s code for external evaluators: AI should be confined to ancillary tasks and documented. She nevertheless asked the Commission to go further by recording whether AI was used, for what purpose and what it specifically contributed, then proactively sharing that information with the relevant funding body.
The inquiry did not find grounds to reopen the disputed proposal evaluation. Its decision is therefore about governance, not proof that AI changed a funding outcome. It also does not say that every language aid is unacceptable.
European Research Council guidance provides a useful boundary. Reviewers remain responsible for assessment and may not use AI to summarise proposals, judge scientific merit or draft evaluations. Limited language improvement and general-information searches are allowed if no proposal or personal data is disclosed.
Build a review-level ledger
For each review, record the tool and version, the permitted purpose, whether proposal content entered the system, the affected passage or task, the human check performed and the evaluator who retained responsibility. Use controlled categories such as language editing, general search and prohibited assessment support; leave room for a short explanation.
Do not collect prompts or proposal text by default. The ledger should prove the boundary without creating a second confidential-data store. Sample records for consistency and escalate missing, ambiguous or assessment-shaped contributions before the review is accepted.
This extends the delegated-authority logic in the personal-agent control analysis: the important question is not merely whether AI was present, but what authority it exercised and what evidence remains.
Decision rule
Do not accept a yes-or-no AI disclosure for consequential peer review. Require a contribution ledger that distinguishes ancillary assistance from delegated assessment while preserving proposal confidentiality.