⚡ Quick Summary
Published by the Future of Life Institute, this April 2026 feedback submission responds to the Commission's draft implementing regulation on evaluations of general-purpose AI (GPAI) models and enforcement proceedings under Articles 92 and 101 of the AI Act (Ref. Ares(2026)2709234, 12 March 2026). The authors welcome the initiative as providing legal certainty and operationalising the Act in several respects.
The document has two parts. It first endorses provisions on broad access powers for model evaluations under Article 2(1), the prohibition on tracking of Commission evaluations, interim measures under Article 5(3), downstream provider complainant rights under Article 6(1), and the five-year limitation periods and interruption provisions in Articles 10 and 11.
It then sets out seven recommendations for improvement: expert independence and competence criteria, the scope of interim measures, protection of complainant identity, the legal basis for periodic penalty payments, transparency of evaluation outcomes and civil society participation, evaluation duration, and deeper access to model configurations.
🧩 What’s Covered
The document follows a two-part structure and reproduces the draft provisions it discusses.
- Provisions supported: endorsement of Article 2(1) access powers covering model weights, hosting infrastructure, system state inspection and all levels of access granted to a provider's own employees; the power to require providers to disable logging measures; interim measures under Article 5(3); complainant participation under Article 6(1); and the five-year limitation period and interruption provisions in Articles 10 and 11.
- Recommendation 1 — expert independence and competence: comparison with EMA's three-year cooling-off rules and EFSA's five-year interest declarations, five sub-recommendations on current ties, a transitional period running until 2 August 2028, mandatory public declarations of interests over the past three years, a 12-month post-appointment restriction, and revocation of compromised appointments.
- Recommendation 2 — interim measures: confirmation that interim measures apply to models already deployed on the market and are not confined to the pre-proceedings phase.
- Recommendation 3 — complainant identity: treatment of the complainant's identity as confidential information under Article 9, given the commercial dependence of downstream providers on GPAI providers.
- Recommendation 4 — periodic penalty payments: gaps in the grounds, procedure and maximum levels for such payments, compared with the DSA and DMA.
- Recommendation 5 — transparency and participation: publication of summary evaluation findings and a mechanism for civil society, researchers and downstream actors to submit observations.
- Recommendation 6 — evaluation duration: periods commensurate with model complexity and risk profile, with extensions where systemic risk is present.
- Recommendation 7 — deeper access: access to model configurations that reveal underlying capabilities before safety mitigations, alongside the guardrailed version.
💡 Why it matters?
For regulators, evaluators and compliance teams working on GPAI obligations, the submission sets out where the procedural architecture of Article 92 and Article 101 enforcement is workable and where it is thin. It explains why access to weights and infrastructure, rather than API-only testing, is needed to detect systemic risk, and it maps independence expectations against EMA and EFSA practice, which practitioners can use to anticipate the Commission's expert-selection criteria. It also flags where the AI Act enforcement regime lacks tools that the DSA and DMA provide, such as express provisions on periodic penalty payments.
❓ What’s Missing
The document is a set of comments on a draft regulation rather than an assessment of the AI Act itself, and it does not reproduce the draft provisions in full, so readers cannot verify all the wording it responds to. It leaves open how evaluation configurations would be specified technically, deferring that to future technical guidance. Costs, timelines and staffing of the evaluation regime are not addressed, nor is the role of national market surveillance authorities. The recommendations rely on analogies to EMA, EFSA, DSA, DMA and competition law without assessing how far those regimes transfer to GPAI models.
👥 Best For
Policy and regulatory affairs teams tracking the AI Act's GPAI enforcement architecture, EU officials and AI Office staff handling the implementing regulation, and compliance leads at GPAI providers and downstream deployers who need to understand how evaluations and proceedings are likely to work in practice, including what the Commission may demand from model providers.
📄 Source Details
The document is Feedback on the Draft Implementing Regulation on Evaluations and Enforcement Proceedings under the AI Act, published by the Future of Life Institute in April 2026 and authored by Risto Uuk, Head of European Policy and Research. It runs to nine pages; the transparency register number printed is 787064543128-10. No URL for the document itself is printed. The input was a full text extraction of all nine pages, including the cover, the seven recommendations, the conclusion and the contact details.
