Research opportunity · The evaluation apparatus: Can benchmarks, leaderboards and audits carry the weight placed on them?
How is external evaluation under the EU AI Code of Practice carried out?
Question
Under the EU Code of Practice for general-purpose AI models with systemic risk, which signatories appoint independent external evaluators, how often are the exceptions used (a model judged similarly safe or safer, or no evaluator found after a public call), and what access and time do evaluators actually receive?
Institutional practice unknownA PhD projectLow costNo study found
Why it matters
The Code writes measurement validity and external evaluation into EU compliance, but its exceptions are broad and its full evidence goes to the AI Office, with public summaries required only if and insofar as necessary, so what is disclosed in practice is unknown. Whether evaluation as governance (P4) works, or becomes the ritual of verification the audit literature describes (P6), depends on practice that no one has yet documented.
The debate lists this among what we still do not know: How are the EU code's external-evaluation requirement and its exceptions applied, and what access do evaluators get?
Evidence that motivates it
- The EU’s Code of Practice for general-purpose AI models with systemic risk requires state-of-the-art model evaluations with internal validity, external validity and reproducibility, elicitation that minimises under-elicitation and sandbagging, and independent external evaluation, unless the model is “similarly safe or safer” or no suitable evaluator is found after a public call such as one open for 20 business days; the Model Report with evaluation results goes to the AI Office, signatories commit to publish summarised versions of their framework and Model Reports if and insofar as necessary to assess or mitigate systemic risks, and external evaluators keep control over publishing their own findings.
The Code requires independent external evaluation except where a model is similarly safe or safer or no evaluator is found after a public call; the Model Report goes to the AI Office, and summaries are to be published if and insofar as necessary. - In their joint pre-deployment test of OpenAI’s o1, the US and UK AI Safety Institutes stated that their findings were preliminary and partial, that they tested an early version without the full mitigations after a limited period of access and after fixing tool-use problems with their own scaffolding, and that the results should not be read as showing whether the model was safe or appropriate for release.
US and UK state evaluators described their own pre-deployment test as preliminary, partial and limited by access. - In a 2022 field scan of algorithmic auditors (survey of 152 and 10 interviews), only 7% said their audit process used a standardized framework, 80% had recommended changes to AI systems that were not implemented, and half of third-party auditors reported lacking access, against none of the first-party auditors.
In a 2022 field scan, half of third-party auditors reported lacking access. - The International AI Safety Report 2026 finds an “evaluation gap”: performance on pre-deployment tests does not reliably predict real-world utility or risk, models “now regularly identify evaluation prompts as tests”, and although 12 companies published or updated frontier safety frameworks in 2025, there are no standardised external audits of their compliance.
The International AI Safety Report found no standardised external audits of companies' safety frameworks.
A possible design
Two arms. A document arm collects signatories' public safety frameworks, model documentation and evaluator publications (for example external evaluators' published pre-deployment summaries) and codes who evaluated, when, with what access and for how long, against the Code's Appendix 3. An interview arm, with consent and confidentiality, covers external evaluators, provider staff and, if possible, AI Office officials, about appointment, exceptions and access. Expect thin public records, since access details go to the AI Office; requests under the EU access-to-documents rules are a third route.
Data needed
Public signatory and evaluator documents; 15-30 interviews; access-to-documents requests to the Commission.
Feasibility
A PhD project. Low; mainly researcher time, with travel for interviews. Access to interviewees inside providers and the AI Office is uncertain; the study should be designed so that the document arm stands alone.
Has this been done?
The search found no study that does the core design (first-pass screen of 15 October 2026; not yet checked in depth).
No study was found of how the Code's external-evaluation commitment has been implemented: which signatories appointed evaluators, use of either exception, or the access evaluators received. Interview studies of evaluation practice exist for product teams and red-teaming dataset builders, not for pre-deployment external evaluators.
How strong this evidence is
- Who searched and checked: .
- What was read: 4 of the 4 nearest works below were judged from an abstract, summary or search listing only, and 1 has details that could not be confirmed; each is marked. Not every entry says how much of the work was read.
- Where the search could not reach: Screened by web search (4 queries) and one arXiv query on the date shown; law reviews, SSRN, EU-policy think tanks and the AI Office's own outputs were not searched, and the current signatory list was not verified. Absence of a hit is not proof that the work does not exist.
- Before building a project on it, run your own search in a scholarly database such as Scopus, Web of Science, Google Scholar or SSRN.
Nearest work found
- Prandi, Suriani, Pierucci, Galisai, Nardi & Bisconti, "Bench-2-CoP: Can We Trust Benchmarking for EU AI Compliance?," arXiv:2508.05464 (2025) Judged from an abstract, summary or search listing only
Abstract read. Maps benchmark coverage onto the Code's systemic risks; about evaluation content, not implementation or access. - Future of Life Institute, "Indicator: Pre-deployment External Safety Testing," AI Safety Index (2025) Judged from an abstract, summary or search listing onlySome details not confirmed
Judged from search listings; not read directly (the PDF was blocked). Per-company tracker of external testers and access, not specific to the Code. - van der Maden, Sadek, Xiao, Mottelson, Liao & Zhu, "Results-Actionability Gap: Understanding How Practitioners Evaluate LLM Products in the Wild," arXiv:2604.16304 (2026) Judged from an abstract, summary or search listing only
Abstract read. Interviews with 19 practitioners on product evaluation; a methodological precedent, not regulatory evaluation. - Alvarado Garcia, Wan, Oguine & Badillo-Urquiola, "Red Teaming LLMs as Socio-Technical Practice," arXiv:2602.18483 (2026) Judged from an abstract, summary or search listing only
Abstract read. 22 interviews with red-teaming dataset designers; not external pre-deployment evaluators.
Search queries used
GPAI Code of Practice external evaluators signatories safety and security chapter implementation analysisAI Office Code of Practice "independent external" evaluator "20 business days" "similarly safe" model report signatoriesinterview study external AI evaluators pre-deployment testing access AISI METR Apollo practice qualitativeethnography OR "interview study" AI evaluation practitioners frontier labs evaluation teams sociotechnical how evaluations are madearXiv: abs:"Code of Practice" AND abs:"AI Act" AND abs:evaluation
Provenance
RO-LD06-04 v1. Drafted by Claude (SocioLens Project Lead), AI-drafted from the LD-06 debate on 2026-10-15. Not yet reviewed. Not yet reviewed by a domain expert. This is a research hypothesis, not a finding.