EU AI Act Transparency Audit

Since 2 August 2025 the EU AI Act has required providers of general-purpose AI models to publish a summary of the content used to train them; models placed on the market before that date have until 2 August 2027. For one model per provider, we ask whether that summary is published, whether anyone can open it without an account, and what it says, judged against its due date; and what providers publish about telling people they face an AI and about marking AI output.

We read the websites of the 26 providers on a list built in advance from two public platform rankings, between 7 October 2026 and 12 October 2026, and again on 14 October 2026, coding each answer from the provider's own documents under a protocol fixed and hash-recorded before collection and published with the dataset. The products studied were not tried. Every reading was AI-assisted, and 13 providers were read twice, blind.

The central result, on the study's final, exploratory coding, is this. Of the 21 providers whose summary deadline had passed for the model studied, 11 published a summary that presents itself as such and opens without an account (7 of 21 on the coding fixed before collection; Annex E). The final coding applies a broader identification rule than the one fixed before collection, and the summaries of Qwen and Tencent were reached only on 15 October 2026, after the collection window; Z.ai's summary names the GLM-5 family but not the model studied, and Qwen's describes the models behind its platform without naming one, so for neither can it be established from the document whether it applies to the version studied. Separately, 4 of the same 21 published a related document describing training content without presenting itself as the summary. Of the 2 providers whose deadline is 2 August 2027, 0 had published a summary. These counts hold for the provider and model pairs studied: applied a second time, the rule that picks each provider's model reached the same model for 6 of 21 providers, and another model could change which document counts.

In the group whose marking deadline is 2 December 2026, 8 of 12 providers already state that outputs are marked (less certain: our two readings agreed below the thresholds set in advance on the 2 items this rests on; Section 4.11); 9 of the 26 providers are removed from this question by a rule. 26 of 26 publish a model card or other per-model documentation and 0 of 26 a copyright policy, neither of which the Act requires them to publish (Table 13). In the places we looked, on 3 dates, we found no published list of general-purpose AI models with systemic risk, which the Act asks the Commission to publish.

The two readings' consistency was 0.893 (Krippendorff's alpha) over every answer and 0.894 over the answers both readings coded directly; it is not a measure of accuracy. Under a name-matching rule fixed in advance, 21 of the 26 providers were matched to organizations in Epoch AI's database of AI models (20 once 3 name fragments are set aside); this compares two lists and validates neither. The list is a constructed frame, not a sample; nothing here finds that any company, or the Commission, complied with or breached the law.