KI ist kein Trend mehr: warum Angst der falsche Berater ist | Umair Zaffar
Umair Zaffar November 8, 2026 0 Comments

EU AI Act Article 53: GPAI Provider Obligations, Implemented in Practice

EU AI Act Article 53 binds not the bank but its model provider: technical documentation, information for downstream providers, a copyright policy and a summary of the training data. We turn that into a provider dossier in the registry that every system on the model hangs from: version pinned, documentation hashed, deprecation notice and incident disclosure settled by contract.

Almost every language-model system in a bank or an insurer today sits on a general-purpose AI model from an external provider. Article 53 governs what that provider has to deliver. It addresses the provider of the model, not the company that uses it. I still consider it one of the most important articles for regulated firms: it is the legal basis for what you can demand from your model provider. And it becomes your own article the moment you change a model so heavily that you become a provider yourself.

The EU AI Act has been in force since 1 August 2024. Prohibitions and AI literacy have applied since February 2025, the obligations for GPAI models since August 2025, the transparency obligations under Article 50 since August 2026. After the postponement by the amending Regulation (EU) 2026/1744 (Digital Omnibus), the obligations for high-risk systems under Annex III apply from 2 December 2027, for Annex I products from 2 August 2028. Article 53 belongs to the GPAI phase (Art. 113(b)) and therefore already applies. Models placed on the market before 2 August 2025 have a transition period until 2 August 2027 (Art. 111(3)). The overview is in EU AI Act in practice.

What Article 53 requires

  • Art. 53(1)(a): The provider draws up and keeps up to date the technical documentation of the model, including training, testing and evaluation results. The minimum content is Annex XI. The documentation goes to the AI Office and national authorities on request, not to customers.
  • Art. 53(1)(b): The provider makes information and documentation available to providers of AI systems that intend to integrate the model, so they understand its capabilities and limitations and can meet their own obligations. The minimum content is Annex XII; trade secrets remain protected. This is the paragraph that gives you, as a bank, something.
  • Art. 53(1)(c): The provider maintains a policy to comply with Union copyright law, in particular to respect rights reservations under Art. 4(3) of Directive (EU) 2019/790.
  • Art. 53(1)(d): The provider publishes a sufficiently detailed summary of the content used for training, following the AI Office template.
  • Art. 53(2): Points (a) and (b) do not apply to models under a free and open-source licence where weights, architecture and usage information are public, but they do apply to models with systemic risk. The copyright policy and the training summary always apply.
  • Art. 53(3) and (4): Cooperation with the Commission and authorities. Until a harmonised standard exists, providers can demonstrate compliance through codes of practice under Article 56; whoever follows neither a code nor a standard must demonstrate alternative adequate means.
  • Art. 53(5) to (7): The Commission may set measurement methods for compute and energy consumption (Annex XI, point 2(d) and (e)) and adapt Annexes XI and XII. What authorities receive is subject to the confidentiality rules of Article 78.

Annex XII, the part you receive: intended tasks and system types, acceptable use policy, release date and distribution methods, software versions, architecture and parameter count, modalities and formats, licence, plus the technical means for integration, maximum input and output sizes such as the context window, and information on training, testing and validation data by type, provenance and curation. Annex XI, which only the authority sees, adds training process, compute, energy consumption, measures against biases and, for systemic risk, evaluation strategies and red teaming.

Systemic risk: Articles 51 and 55

Under Art. 51(1) a model has systemic risk if it has high-impact capabilities or if the Commission classifies it as such under Annex XIII; Art. 51(2) presumes this above a cumulative training compute of 10^25 floating point operations. For such models Art. 55(1) additionally requires evaluation under standardised protocols including adversarial testing, assessment and mitigation of systemic risks at Union level, reporting of serious incidents to the AI Office without undue delay, and adequate cybersecurity. The large models that banks buy usually fall into this class. So the provider has a reporting duty, and I want the relevant incidents on my desk as well, by contract.

The code of practice under Article 56

Art. 56 tasks the AI Office with facilitating codes of practice that cover at least Articles 53 and 55; if none materialises, the Commission can set common rules by implementing act (paragraph 9). The GPAI code of practice has been available since July 2025, with chapters on transparency, copyright, and safety and security; the transparency chapter contains a documentation form for Annexes XI and XII. Ask your provider whether it has signed and will deliver the form.

The technical example

An insurer builds two systems on a hosted GPAI model: a customer-service assistant that answers questions about policy terms through RAG, and an extractor that reads claims notifications, pulls out fields and routes the case. The data flow is the same for both: business application, gateway, pseudonymisation, retrieval, model call against a specific model version, signal check, answer or escalation. Both are in the registry with purpose, risk class, scope, model version, evaluation and expiry.

Article 53 bites in three places. First, purpose: without Annex XII the insurer does not know what the provider’s acceptable use policy excludes. If it excludes decisions about insurance benefits, the extractor, as decision preparation, has to be reviewed before it gets a route. Second, version: the provider replaces the model version with a new one in a few weeks. The evaluation in the registry covers the old one; nobody has tested the new one on the insurer’s own test set. Third, incidents: the provider reports a security problem to the AI Office under Art. 55(1)(c). Nothing obliges it to tell the insurer as well.

The roles are separate. The model provider owes Article 53. The insurer is the provider of the two AI systems and the deployer under Article 26; whether one is high-risk is decided by Article 6, regardless of the model. And once the claims team retrains the model extensively, the fourth question arises: has the insurer itself become a GPAI provider?

How we implement it

At SIFAMO we keep not only one record per system in the registry but also one per GPAI model, the provider dossier: provider, model name, exact version identifier, the Annex XII documentation as a file with SHA-256 hash and retrieval date, version of the acceptable use policy, licence, links to the copyright policy and the training summary, whether the code of practice was signed, the agreed deprecation notice period and the incident clause. Without the dossier, no system can be registered on that model. No registration, no route, one level further down as well. How the components fit together is in the post on the governance layer.

Every system pins a version. The system record holds the exact version identifier, and the gateway routes only there. Registration rejects an alias like “latest”, because the evaluation would otherwise cover a model that no longer exists tomorrow. At the quality gate the hash of the provider documentation is copied into the system record. If the provider changes its documentation, every system with the old hash gets a review task for its business owner: has the purpose, context window or acceptable use policy changed, and does that affect us?

We treat deprecation notices as expiry dates. When the provider announces the end of a version, the registry sets the expiry of all dependent systems to that date minus the time for an evaluation run on our own test set. Contractually we demand a notice period of at least one full evaluation cycle and an incident clause: incidents the provider reports under Art. 55 that affect our use are disclosed to us too. They land in the evidence log of the affected systems, and the business owner decides on the kill switch. The auditor thus sees the chain for every decision: evidence log, system record, dossier, documentation hash and the Annex XII file as it stood on that day.

Where the AI Act leaves room: Article 53 gives you Annex XII, not Annex XI. Evaluation results or compute figures we obtain by contract, for high-risk systems backed by Art. 25(4). And the provider’s acceptable use policy is, for us, a hard scope boundary in the gateway, not a contract annex.

When the bank itself becomes a GPAI provider

Whoever modifies a GPAI model becomes, under recital 109, the provider for that modification; the Article 53 obligations are then limited to the modification. The Commission guidelines on the GPAI obligations from July 2025 give an indicative threshold of more than one third of the original model’s training compute. We turn that into a registry rule: every fine-tuning run logs its compute against the base model’s figure from the dossier, or against a documented estimate if need be. Below the threshold the model remains a version of the system, with documentation of data, method and evaluation. Above it the registry opens a provider dossier for ourselves: Annex XI for the modification, a copyright policy for the fine-tuning data, a training summary for publication. Parameter-efficient adaptations on a few thousand examples sit far below that. But without a log you cannot prove it later.

Typical mistakes

  • The model version in the code is “latest”. The evaluation covers a version that has since been retired, and nobody noticed.
  • The Annex XII documentation was downloaded once and never looked at again. No hash, no date, no comparison with the current edition.
  • Nobody has read the provider’s acceptable use policy. The use case falls into an excluded category, and the conformity argument is built on sand.
  • The provider’s incident reports go only to the AI Office. No contractual channel to the customer, no process that reaches the business owner.
  • Fine-tuning without a compute log. Whether the firm has become the provider of the modified model cannot be reconstructed later.

EU AI Act Article 53 checklist

  • A provider dossier exists in the registry for every GPAI model; without a dossier, no system can be registered.
  • The Annex XII documentation is stored as a file with hash and retrieval date; the copyright policy and the training summary are linked.
  • Every system pins an exact model version; the gateway routes to no other.
  • The documentation hash is copied into the system record at the quality gate; a change triggers a review.
  • Deprecation notice period and incident clause are agreed in the contract and recorded in the registry; a retirement announcement sets the expiry of the dependent systems.
  • The provider’s acceptable use policy is a scope boundary in the gateway and is checked against the purpose of every system.
  • Every fine-tuning run logs compute, data, method and evaluation; the provider role is re-assessed on every run.

Open questions

  • The one-third threshold for fine-tuning sits in Commission guidelines, not in the Regulation. Whether it is the right measure for parameter-efficient methods is open.
  • Harmonised standards for Annexes XI and XII are missing. What one can demand from model providers that have not signed the code is a matter of negotiation.
  • Which level of detail in the training summary actually helps a downstream provider, for instance for bias checks on its own test set, is a research question.

Frequently asked questions

Does Article 53 also apply to us as a deployer?

No. The addressee is the provider of the GPAI model. You are the beneficiary of paragraph 1(b), and your own obligations under Article 26 or, for high-risk, Article 13 depend on you receiving and keeping that information. How it belongs in the registry is under AI governance.

What may we demand from the model provider?

By law at least Annex XII, plus the published copyright policy and training summary. Annex XI goes only to the authority. Evaluation results, compute figures, deprecation notice periods and incident reports you get only by contract, for high-risk systems backed by Art. 25(4).

Does fine-tuning make us a GPAI provider?

Only for a significant modification, with more than one third of the original training compute as the indicative threshold. Then Article 53 applies to your modification. Log every run, and the answer becomes provable instead of estimated. The other articles in the series: all articles at a glance.

Related: EU AI Act · All articles · AI Governance · AI Consultant

Leave Comment