Research opportunity · Regulation-in-Practice: Rights, Compliance & Developmental Steering
What do evaluators, researchers and regulators actually get to see under EU law?
In short: has this been done?
Related work: Partly addressed. Related studies cover parts of this question; what they leave open is set out below. Only a first-pass search has been done, so it is not yet known whether a fuller search would find a study that does the whole design.
How far checked: First-pass screen only. A first-pass AI web search on 16 October 2026. No deeper check yet. Every search and check so far was done by AI; no person or domain expert has checked them yet.
The searches, the nearest work found and what the search could not reach
Across debates: it also draws on The evaluation apparatus: Can benchmarks, leaderboards and audits carry the weight placed on them? and Who can know about AI? Epistemic dependence and jurisdiction over the field, and is listed on each of their pages.
This page shows RO-LD12-07 version v1, whose record never changes. It is the current version; if the opportunity is revised, the newer version appears at its main page.
Question
For general-purpose AI providers that also run platforms or search engines designated under the Digital Services Act, what does each legal route for outside scrutiny deliver in practice (external evaluation under the GPAI Code of Practice, which applies only to models with systemic risk, vetted-researcher data access under the Digital Services Act, and the AI Office's information requests and model evaluations), recorded for each route as what object is reached (model, platform data or documents), who can start it, how long it takes and how deep the access goes?
Gap type: Institutional practice unknownOne PhD chapter or paperLow costRelated work: Partly addressed
Why it matters
Three debates each find a legal route to outside scrutiny whose use is not yet known. LD-06 records that the Code of Practice requires independent external evaluation, with exceptions, for general-purpose AI models with systemic risk (signing the Code does not by itself bring a model under that commitment), and that third-party auditors reported lacking access in a 2022 field scan. LD-08 records that the Digital Services Act route gives vetted researchers internal data of designated platforms and search engines (it does not, on its face, give access to models or to AI chat logs) and had produced applications but no reported access by May 2026. LD-12 records that the AI Office has begun information requests, and that adhering to the Code counts towards demonstrating compliance. The routes grant different things to different people (evaluators contracted by the provider, researchers who apply, a regulator that can compel), so studied one at a time each looks like a new regime starting up. Recorded side by side for the same firms, they show what outsiders can and cannot see about the same AI features, which is the question behind all three debates.
Evidence that motivates it
- The EU AI Office began using information-request powers after its GPAI enforcement powers became applicable; the exact report that more than 30 providers received initial RFIs remains press-relayed. (LD12-HQ-C04 v1.0)
How this study uses it: The AI Office began using its compulsory information-request powers once its general-purpose AI enforcement powers applied; the count of recipients is only press-relayed. - The EU GPAI Code of Practice is voluntary, but Commission guidance recognizes adherence as a route to demonstrate compliance; non-signatories must demonstrate compliance by other adequate means. (LD12-HQ-C03 v1.0)
How this study uses it: Adherence to the voluntary Code counts towards demonstrating compliance, so the Code's evaluation commitments matter for the regulator's route too. - 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. (LD06-SL-C24 v1.0)
How this study uses it: For general-purpose AI models with systemic risk, the Code requires independent external evaluation unless the model is similarly safe or safer or no evaluator is found after a public call; the provider engages the evaluator, and evaluators keep control of publishing their own findings. - 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. (LD06-SL-C26 v1.0)
How this study uses it: In a 2022 field scan, half of the 11 third-party auditors who answered reported lacking access, against none of the first-party auditors; a small, pre-generative-AI sample. - As of 19 May 2026, national Digital Services Coordinators had received 49 applications for vetted-researcher access under the Digital Services Act, mostly for social-media data and on risks including illegal content, advertising transparency and “AI features”; the coordinators were assessing the applications, and their update reported no access granted. (LD08-SL-C25 v1.0)
How this study uses it: By May 2026 national coordinators had received 49 vetted-researcher applications, some on AI features, and reported no access granted. - On 2 July 2025 the European Commission adopted a delegated act on data access under the Digital Services Act that lets vetted researchers obtain internal data from very large online platforms and search engines to study systemic risks; applicants must show affiliation to a research organisation, “independence from commercial interests” and the ability to keep the data secure. (LD08-SL-C24 v1.0)
How this study uses it: The delegated act gives vetted researchers a route to internal data of designated platforms and search engines; the announcement does not address the conversation logs of AI chat services.
Each claim is shown as it read when this version was published.
A possible design
Take the providers that both signed the Code and run designated platforms or search engines with generative-AI features (from the Commission's signatory and designation lists, checked on the date of the study), and record for each of their models whether it is classified as a general-purpose AI model with systemic risk, since only those fall under the Code's external-evaluation commitment and the AI Office's systemic-risk powers. For each route, assemble a register from official sources (AI Office publications and enforcement documents, including any evaluations under Article 92 of the AI Act; Digital Services Coordinators' reports and the Commission's data-access portal; providers' published summaries of Model Reports), access-to-documents requests and structured interviews with external evaluators, applicant researchers and officials. Because the routes differ, code each instance by what it reached (model, platform data or documents), who started it (provider, researcher or regulator), its stages and their duration, and the depth of access per object (for models, black-box, grey-box or white-box, following published access taxonomies). Compare what is known about the same AI feature through each route over 2026 to 2028.
Data needed
Official AI Office and Digital Services Act implementation documents; the Commission's signatory and designation lists; access-to-documents requests; interviews with evaluators, researchers and officials; providers' published Model Report summaries where they exist.
Feasibility
One PhD chapter or paper. Low; documentary research, access-to-documents requests and interviews. The set of firms that are both signatories and designated platforms is small, so this is a comparison of routes within a few firms, not a sample. Much of the record is confidential and evaluators may be bound by agreements, so what could not be seen is a finding. The routes start at different times (the AI Office's evaluation powers applied from August 2026), so early comparisons are partial.
Has this been done?
Partly done already: related work covers some of this question (first-pass screen of 16 October 2026; not yet checked in depth).
Taxonomies of evaluator access, proposals for third-party auditing and expert workshops on secure access exist, and commentary contrasts the Digital Services Act and AI Act routes in law. No study was found that records what the evaluator, researcher and regulator routes deliver in practice for the same firms. The search was web-based and partial.
How strong this evidence is
- Who searched and checked: no checker is recorded beyond the first-pass screen, which was part of the AI drafting (Claude (SocioLens Project Lead), AI-drafted from the LD-12, LD-06 and LD-08 debates). Every check so far was done by AI; no person or domain expert has checked it yet.
- What was read: 3 of the 3 nearest works below were judged from an abstract, summary or search listing only; each is marked. Not every entry says how much of the work was read.
- Where the search could not reach: Screened by three web searches and arXiv records on the date shown; legal databases and EU-law journals were not searched. 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
- Charnock, Tlaie, O'Brien, Casper and Homewood, "Expanding External Access To Frontier AI Models For Dangerous Capability Evaluations", arXiv:2601.11916 (2026) Judged from an abstract, summary or search listing only
Abstract read. Proposes a taxonomy of evaluator access (model access, information, timeframe) and notes that the Code's "appropriate access" is not defined in practice; does not measure access given. - Brundage, Dreksler, Homewood, McGregor, Paskov and others, "Frontier AI Auditing: Toward Rigorous Third-Party Assessment of Safety and Security Practices at Leading AI Companies", arXiv:2601.11699 (2026) Judged from an abstract, summary or search listing only
Abstract read. Sets out assurance levels for third-party audits; a proposal, not a measurement of the EU routes. - RUSI, "Secure Third-Party Access to Frontier AI" (2026) Judged from an abstract, summary or search listing only
Seen in search results only. Draws on expert workshops on third-party access to closed models and excludes regulators from scope.
Search queries used
comparison researcher access DSA Article 40 AI Act external evaluators GPAI code of practice access to models studythird-party access frontier AI models external evaluation access levels empirical survey evaluators 2026AI Office model reports external evaluators appointed signatories GPAI code 2026 implementation tracker
Cite this page
SocioLens project. (2026, October 16). What do evaluators, researchers and regulators actually get to see under EU law? [Research opportunity RO-LD12-07, version v1]. SocioLens. https://sociolens-web.pages.dev/opportunities/ro-ld12-07/v1/
@misc{sociolens_ro_ld12_07_v1,
author = {{SocioLens project}},
title = {What do evaluators, researchers and regulators actually get to see under EU law?},
year = {2026},
date = {2026-10-16},
howpublished = {\url{https://sociolens-web.pages.dev/opportunities/ro-ld12-07/v1/}},
note = {Research opportunity RO-LD12-07, version v1}
}The link points to this version, whose record never changes. For a finding, cite the original study as well: SocioLens summarizes research and does not replace it.
Provenance
RO-LD12-07 v1. Drafted by Claude (SocioLens Project Lead), AI-drafted from the LD-12, LD-06 and LD-08 debates on 2026-10-16. Not yet reviewed. Not yet reviewed by a domain expert. This is a research hypothesis, not a finding.