Summary

  • EISMEA found that some external evaluators used AI to edit review text and research background material; the Ombudsman accepted EISMEA’s explanation of the individual decision.
  • The system question is who receives a usable record of AI’s role, and when that record travels with the human review.

A review has more than one authoring layer

A grant decision can carry a human verdict and a machine-assisted edit in the same file. The European Commission’s current expert code keeps the reviewer responsible, limits AI to strictly ancillary tasks and requires documentation to be available to the contracting authority on request. The rule assigns responsibility. The record may still wait for someone to ask for it.

That distinction surfaced in a complaint from a Polish company that applied to the European Innovation Council Accelerator under Horizon Europe. The application was refused in 2024. The company argued that outside experts had used third-party AI in preparing their reviews and raised fairness and confidentiality concerns.

The file contains a narrower fact than the allegation. EISMEA had found that some evaluators used AI to edit their reviews and research background information. The European Ombudsman found EISMEA’s explanation of the individual evaluation clear and reasonable, and did not open an inquiry into the merits of that result. The public case record identifies neither the tools nor their precise contribution to this proposal’s review; it makes no finding that AI set the score.

The inquiry widened because one answer about one application could not settle how the wider system records machine assistance. Horizon Europe normally sends proposals to a panel of at least three independent external experts. The Granting Authority produces a ranked list and the list of applications proposed for funding. Those reviews carry institutional weight. A trace of AI use belongs beside the work the authority relies on, not in a file that appears only when an applicant knows to challenge it.

The rule arrived after the review

When the proposal in this case was evaluated, the expert contract did not yet contain a formal AI-use clause. The Commission told the inquiry that it updated its Code of Conduct on 15 December 2025 after applicants had complained in 2024. Today, section 1.3 permits AI only for strictly ancillary purposes, leaves responsibility with the expert and requires use to be documented for the contracting authority on request. Confidentiality and data-protection duties apply when AI is used.

The code leaves the shape of that record open. Commission representatives suggested keeping a version of the review before AI editing and another after it. They did not make that method mandatory; different forms of documentation are allowed. EISMEA said additional guidance is also given orally during required evaluator briefings.

EISMEA told the Ombudsman team that it particularly requests the evidence when a complaint about AI use is lodged. The Ombudswoman called a complaint-triggered record request potentially unreliable and asked the Commission to explore logging whether AI was used, for what purpose and what it contributed, with that information provided alongside each evaluator’s review to the funding authority.

A receipt lets the funder audit its own process

The Commission pointed to scale: its programmes and agencies receive tens of thousands of proposals, with each assessed by at least three independent experts. A universal record creates work for reviewers and administrators. But a blank field also has a cost: the authority cannot tell whether AI polished a sentence, supplied background material or played another role before the reviewer signed.

The inquiry now asks whether the Commission can see the process it commissioned. It has until the end of April 2027 to report what it will change.

A short receipt could answer four questions beside the review: whether AI was used; which tool; for what ancillary task; and what the expert kept, changed or rejected. The expert remains accountable for the recommendation, and the contracting authority gains a way to check that the rule and the work match. That is the difference between responsibility written into a contract and responsibility a funder can verify.

Sources