Article 53 of the EU AI Act lays out the obligations for general-purpose AI model providers. They sit separate from, and in addition to, the transparency obligations that Article 50 places on AI system providers and deployers. Early implementation effort routinely conflates these two layers, and the result is compliance work that satisfies neither.
This is a working checklist, not legal advice. Providers should validate against their own counsel and the most recent EDPB and Commission guidance. If you are not yet sure whether Article 53 even reaches your product, the interactive GDPR + EU AI Act Applicability Wizard walks you through a few questions and tells you which obligations attach to your deployment.
Scope: who Article 53 applies to
The obligations attach to providers of “general-purpose AI models.” The definition centers on three properties: the model is trained on a large amount of data using self-supervision at scale; it displays significant generality; and it is capable of performing a wide range of distinct tasks. Recital 98 adds that models with at least a billion parameters, trained with a large amount of data using self-supervision at scale, should be considered to display significant generality, so in practice the definition captures most foundation models, plus smaller models if they exhibit the generality criterion. The Commission’s guidelines of 18 July 2025 add an indicative compute criterion: a model trained with more than 10^23 FLOP that can generate language (as text or audio), text-to-image or text-to-video is considered a general-purpose AI model. It supplements the generality test rather than replacing it. Fine-tuned variants typically inherit the obligation when they materially alter capabilities.
The obligation does not attach to deployers: a company that builds a chatbot on top of a foundation model without modifying the model has different obligations, chiefly the Article 50 transparency duties and, if the use is high-risk, Chapter III. Foundation model providers cannot transfer their Article 53 obligations downstream by contract.
The four core obligations
1. Technical documentation
A provider must maintain technical documentation describing:
- The model’s intended tasks, modalities, and capabilities
- The acceptable use policies the provider applies
- Training methodology at a level of generality that allows downstream evaluation
- Architecture and parameter counts (or comparable indicators of scale)
- Compute used for training (FLOPs)
- Energy consumption estimate
- Known limitations and biases identified during evaluation
The documentation must be sufficient for downstream deployers to assess whether the model fits their use case and to satisfy their own AI Act obligations. “Sufficient” is the litigated word here — it does not mean “complete” but it does mean more than a marketing one-pager.
2. Training data summary
This is the disclosure that has generated the most pushback from US providers. Article 53(1)(d) requires providers to “draw up and make publicly available a sufficiently detailed summary about the content used for training of the general-purpose AI model, according to a template provided by the AI Office.” The AI Office’s template, not the provider’s own format, sets the structure of the summary. The Commission published it, with an explanatory notice, on 24 July 2025. It has three parts:
- General information: model identification, the modalities covered (text, images, video, audio), estimates of the training data’s size, and language and demographic coverage.
- List of data sources: large datasets named individually and smaller ones described collectively; content licensed from rights holders; web-crawled data, naming the top 10% of domain names used (the top 5% or 1,000 domains for SMEs); user data from interactions with the model or other services; and synthetic data generated by other AI models, identifying the source.
- Data processing aspects: how the provider respected text-and-data-mining opt-outs under the EU Copyright Directive, and the measures used to detect and remove illegal content.
The summary must be updated every six months, or sooner when further training is a material change.
Recital 107 sets the level of detail. While taking due account of the need to protect trade secrets and confidential business information, the summary should be “generally comprehensive in its scope instead of technically detailed”, so that parties with legitimate interests, including copyright holders, can exercise and enforce their rights. The recital’s examples are listing the main data collections or sets that went into training, such as large private or public databases or data archives, and giving a narrative explanation of the other data sources used.
This is operationally heavy. Providers should treat it as a data engineering project, not a legal-team deliverable.
3. Copyright compliance documentation
Providers must implement a policy to comply with EU copyright law, including respecting machine-readable opt-outs (robots.txt, IETF AI Preferences when standardized, ai.txt). They must maintain records sufficient to demonstrate compliance — which in practice means logs of which sources were excluded by which mechanism at which point in the training pipeline.
Note: this is independent of any pending litigation in the US or other jurisdictions. EU compliance is required regardless of whether the use would be fair use under US law.
4. Systemic-risk obligations (Article 55)
Models classified as general-purpose AI models with systemic risk carry a further set of duties under Article 55. The classification test sits in Article 51: under Article 51(2), high-impact capabilities are presumed where the cumulative compute used for training exceeds 10^25 FLOPs, and the Commission is empowered to adjust that threshold. Once a model is in that class, Article 55 adds:
- Documented evaluation against state-of-the-art protocols, including red-teaming
- Tracking and reporting of serious incidents
- Cybersecurity protections at the model and infrastructure level
- An incident response process with notification to the AI Office
The Article 51(2) threshold captures GPT-4-class models and above today. Smaller specialized models can also be designated by the Commission if they pose systemic risk despite lower compute — for example, a biology-specialized model that meaningfully advances dangerous-capability frontiers.
What disclosure does NOT require
Common over-disclosure errors in early drafts:
- Specific training data contents. Recital 107 asks for a summary that is “generally comprehensive in its scope instead of technically detailed”, not file-by-file disclosure.
- Model weights or architecture in full. Architecture descriptions can be at the level of “transformer with X layers, Y attention heads.”
- Customer or deployer names. Article 53 is about the model itself, not who uses it.
- Trade-secret training methodology. Procedural descriptions are required; algorithmic implementations are not.
Disclosing more than required has both competitive and adversarial costs. Adversaries use detailed model documentation to design jailbreaks. Disclose at the granularity the Act requires; not more.
Operationalization: a 90-day plan
Providers without an existing compliance posture can use this sequence:
- Days 1-15: Inventory existing model documentation. Map gaps against Article 53 requirements. Identify whether you exceed the Article 51(2) 10^25 FLOP threshold.
- Days 16-45: Build the training data summary as a data-engineering deliverable. Most of the friction is upstream: reconstructing which data collections and sources fed training from pipelines that weren’t designed to record it.
- Days 46-75: Document the copyright compliance policy. Wire up checks for
robots.txtand other machine-readable rights reservations, and instrument logs. - Days 76-90: For systemic-risk models, stand up the red-team and incident-response processes. The AI Office expects ongoing artifacts, not a one-time document.
Most providers will find step 2 takes the entire 90 days. Plan accordingly.
Timeline and enforcement
- Since 2 August 2025: the Article 53 obligations, and Article 55 for models with systemic risk, have applied to models placed on the market from that date (Article 113(b)).
- From 2 August 2026: the Commission can fine GPAI providers under Article 101, which Article 113(b) carved out of the 2025 date. Fines reach up to €15 million or 3% of worldwide annual turnover, whichever is higher.
- By 2 August 2027: models placed on the market before 2 August 2025 must comply (Article 111(3)).
Enforcement sits with the Commission, which has exclusive powers over the GPAI chapter and entrusts them to the AI Office (Article 88). Until a harmonised standard is published, Article 53(4) lets providers rely on a code of practice under Article 56 to demonstrate compliance. The General-Purpose AI Code of Practice, published on 10 July 2025, is that code: a voluntary tool with three chapters, Transparency and Copyright for every GPAI provider and Safety and Security for the small group of providers of the most advanced models. Signatories demonstrate compliance by adhering to it; a provider that adheres to no approved code and meets no harmonised standard must demonstrate alternative adequate means of compliance for the Commission to assess.
