What has been decided

Commission Implementing Regulation (EU) 2026/1755 of 20 July 2026 sets out, for the first time, the detailed arrangements by which the European Commission evaluates general-purpose AI models under Article 92 of the AI Act and conducts proceedings that may end in fines under Article 101 (up to 3 % of worldwide turnover or EUR 15 million). The draft was published on 12 March 2026 with a feedback period until 9 April 2026; the regulation was published in the Official Journal on 21 July 2026 and entered into force on 10 August 2026.

• Model access (Art. 2): access may extend to APIs, internal access, source code, model weights, the hosting infrastructure, inspection and modification of the system state – up to every level of access granted to the provider’s own employees.
• Logging switch-off (Art. 2(3)): the Commission may require the provider to disable logging measures that could track or record the Commission’s access.
• Independent experts (Art. 3–4): independence criteria covering ownership, governance, personnel and contractual relationships during at least the 12 months preceding the evaluation; selection as a rule by open procedures; experts from the scientific panel (Art. 68) may be appointed directly.
• Proceedings (Art. 5–6): opening, closing and reopening „at any time“; interim measures in urgent cases based on a prima facie finding of infringement.
• Procedural safeguards (Art. 7–8): at least 21 days for written observations on preliminary findings; access to the file subject to business secrecy protection.
• Limitation (Art. 10–11): five-year limitation periods for imposing and for enforcing fines.

GISAD’s assessment

1. Real access instead of self-declaration – positive. Genuine technical access, down to source code and model weights, follows the GISAD principle that assessment must take place outside the manipulable area. Provider self-declarations prove nothing.

2. The logging clause creates a blind spot. Switching off provider logging protects the integrity of the evaluation – but it creates an unobserved zone inside the provider’s system, and liability for anything that happens during a switched-off access remains unclear. Whoever evaluates must be logged themselves. Evaluation accesses should be signed and immutably deposited with a neutral instance (Truststation principle): switch-off at the provider, yes – complete removal of the measurement trail, no.

3. Expert independence is well regulated – but remains in one hand. The Commission defines the criteria, selects the experts and assesses their independence. Operator, supervisor and enforcer coincide. GISAD therefore calls for a supplementary independent, read-only audit layer in the sense of concept-fidelity auditing – event-driven and sample-based, not permanent.

4. Due process deserves recognition. Written observations within at least 21 days, access to the file under confidentiality safeguards and clear limitation periods are sound rule-of-law provisions.

5. Reopening „at any time“ is a structural gatekeeper element. A provider never obtains a reliable conclusion of proceedings; permanent uncertainty replaces legal certainty. Closure decisions should be tied to defined conditions instead of an unlimited reopening option.

6. Structure instead of punishment. The AI incidents of 2026 – from the Hugging Face attack to AI agents switching off their own safety monitors – show what matters: single-step release for every critical AI action, signed instance numbers and the zero-tolerance rule of the Finder AI-Escape-Switch. A fining procedure only acts after the breakout. The evaluation environment itself must run on sovereign European infrastructure – not on US hyperscalers.

7. A timing gap. Article 101 has been applicable since 2 August 2026, while the detailed procedural rules only took effect on 10 August 2026. For eight days, fining powers existed without a detailed procedure. Coordinated dates would have been cleaner.