AI agents 2026 are built on training data the public can't see. The EU AI Act requires providers to publish training-data summaries — our audit of 21 providers found 11 publish one and 6 publish nothing. This tracker names names.
The EU AI Act requires general-purpose AI providers to publish a summary of their training data. Our audit of 21 providers found 11 publish one and 6 publish nothing — the rest are partial or unclear. This tracker names names, so you can see which labs meet the transparency duty and which are still silent.
An EU AI Act training-data summary is a public document in which a general-purpose AI provider describes the data used to train its models — required by the Act's transparency duties for general-purpose models.
Since August 2025, EU law has required providers of general-purpose AI models to publish a summary of what they trained on — the EU AI Act training-data summary. It's the one document about these models the law makes public — so anyone can go and look for it. An independent audit did. Of the 21 providers whose deadline had passed, only 11 publish a summary that presents itself as such, 4 publish a model card instead, and 6 publish nothing at all. Search and sort the table.
Short answer: the publishing providers, per the audit's coding, are Anthropic, Cohere, DeepSeek, Microsoft, MiniMax, NVIDIA, OpenAI, Qwen, Tencent, Thinking Machines Lab, and Z.ai. The audit does not name which four publish only a related document or which six publish nothing — and nothing in the report is a finding of legal compliance or breach.
Per the Oct 16, 2026 preprint audit of 26 providers, 11 publish a training-data summary: Anthropic, Cohere, DeepSeek, Microsoft, MiniMax, NVIDIA, OpenAI, Qwen, Tencent, Thinking Machines Lab, and Z.ai. Four publish only a related document, six publish nothing. These are the study's statuses from websites read October 7–14, 2026 — not a legal finding.
Based on the EU AI Act Transparency Audit (AI Tools Police), preprint published 16 October 2026. The researchers read providers' public websites between 7–12 October and again on 14 October 2026. Status reflects what they found on those pages, on those dates — every row's source is the preprint dataset.
| Provider | Status | Detail | Source |
|---|
Looking for a provider not listed? It is not in the dataset — the study covered 26 providers drawn from public platform rankings, Hugging Face downloads, and OpenRouter token rankings. Statuses are the study's, not ours, and we invented none.
Sources: the audit summarized via Mücahit Kaya's write-up of the preprint (Oct 2026); study page and open dataset: aitoolspolice.com/research/eu-ai-act-transparency-audit; DOI 10.5281/zenodo.23415007. EU duty background: Lexology — EU Releases AI Training Data Template.
Since 2 August 2025, providers of general-purpose AI models must publish a public summary of the content used to train them (Article 53 GPAI duties). Models placed on the market before that date have until 2 August 2027. The European Commission published a mandatory template for the summary on 24 July 2025. It is the one document about these models that the Act makes public by law — so anyone can go and look for it.
The EU AI Act Transparency Audit (AI Tools Police, preprint published 16 October 2026) read the public websites of 26 AI model providers. Of the 21 whose summary deadline had passed for the model studied: 11 publish a summary that presents itself as such and opens without an account (Anthropic, Cohere, DeepSeek, Microsoft, MiniMax, NVIDIA, OpenAI, Qwen, Tencent, Thinking Machines Lab, Z.ai); 4 publish only a related document such as a model card; and for 6 the researchers found nothing in the places they looked.
Not by itself. The researchers state explicitly that nothing in the report is a finding that any provider complied with or breached the law. The study checked specific public pages on specific dates, and models placed on the market before 2 August 2025 have until 2 August 2027 to publish. A missing summary is a transparency gap, not a legal verdict.
Under the study's coding, a document the provider itself titles as the training-data summary, which names the model it covers and opens without an account. Of the 11 summaries found, 9 name the model studied; Z.ai's names the GLM-5 family without naming the model studied, and Qwen's describes the models behind its platform without naming one.
The Act does not require providers to publish their copyright policy, and the audit found 0 of 26 present one as such (2 of 26 state that one exists). But 10 of 26 publish a specific route by which a rightsholder can reserve their works against training use — such as the name of the crawler to block.