Commission publishes guidelines on general-purpose AI obligations under the AI Act

What the guidelines cover

The Commission's new guidelines run to over 50 pages and address the obligations for providers of general-purpose AI models under the AI Act ^[1]. These are the systems that can be put to many uses, from writing text to analysing images, often through an API or as a service.

The guidance sets out when a model is considered to present systemic risk, what documentation providers must keep, what information they must give to downstream providers who integrate their models, and how to meet transparency obligations around copyright and generated content ^[1]. It also explains the adversarial testing regime and the process for model evaluation ^[1].

The guidelines come from the AI Office, which supervises general-purpose AI under the regulation, and were published on the AI Act Service Desk in December 2024 ^[1]. They are not legally binding, but they show what the Office will look for when it enforces the rules.

Who the rules apply to

The obligations fall on providers of general-purpose AI models: the organisations that develop, train and make the models available, whether openly or behind an API ^[1]. A downstream provider who takes a general-purpose model and adapts it to a specific use has obligations under other parts of the Act, but not usually under Articles 53 and 54 unless the adaptation is substantial enough to count as a new model ^[1].

The guidance gives examples. If you fine-tune a publicly available language model on your own data to answer customer questions, you are a downstream provider, not a provider of the underlying general-purpose model ^[1]. If you build a model from scratch or train a foundation model and release it, you are.

The distinction matters because the transparency and documentation duties are on the provider of the general-purpose model, and downstream providers need the information the guidelines say the model provider must give them.

Systemic risk

General-purpose AI models with systemic risk face stricter rules, including adversarial testing and the duty to report serious incidents to the AI Office ^[1]. The Act says a model presents systemic risk if it has high-impact capabilities, evaluated by the total compute used in training, or if the AI Office designates it based on its capabilities and reach ^[1].

The guidelines explain how the Office will assess this. The primary indicator is compute: a model trained with more than 10^25 floating-point operations is presumed to have systemic risk ^[1]. For models below that threshold, the Office can still designate them based on factors like the number of users, the model's reasoning or autonomous capabilities, and whether it can cause harm at scale ^[1].

Once designated, the provider has six months to comply with the additional obligations ^[1]. The guidelines list what those obligations are: model evaluation, adversarial testing, tracking and documenting serious incidents, ensuring an adequate level of cybersecurity and reporting on energy consumption ^[1].

Transparency on training data

Article 53 requires providers to make publicly available a sufficiently detailed summary of the content used to train the model, respecting trade secrets and confidential business information ^[1]. The guidelines say that means listing the main data sources, not every URL, and describing the kind of content: text corpora, image sets, code repositories ^[1].

If copyrighted material was used, the provider must put in place a policy to comply with EU copyright law, including the obligation in the Copyright Directive to allow rights holders to reserve their works from text and data mining ^[1]. The guidelines explain that this means having a technical means for rights holders to express that reservation, and a process for honouring it ^[1].

This is the copyright transparency obligation that applies from 2 August 2025, earlier than most other provisions of the Act ^[1]. It does not require providers to list every work, but the summary must be detailed enough that rights holders can assess whether their content might have been used ^[1].

Documentation and downstream information

Providers must draw up technical documentation that includes a description of the model, its capabilities and limitations, the training process, data sources and evaluation results ^[1]. The guidelines give a list of what that documentation should contain, cross-referenced to the template in Annex XI of the Act ^[1].

Downstream providers, the ones who take a general-purpose model and put it to a specific use, need information to meet their own obligations. If they are using the model in a high-risk system, they need to know its capabilities, its known limitations and what it was trained on ^[1]. Article 53 says the general-purpose AI provider must give them that information, and the guidelines clarify what counts as sufficient: enough for the downstream provider to do their conformity assessment and comply with the logging and human oversight requirements in Articles 12 and 14 ^[1].

The guidelines note that much of this can be done through model cards or API documentation, as long as it is kept up to date and accessible to the downstream providers who need it ^[1].

A record of what was released, and when

Here is where the evidence question comes in. The obligations are on the provider at the time the model is made available. A year or two later, when a downstream provider asks what the model's known limitations were when they started using it, or when a supervisor asks what data sources were listed in the summary at a given release, the provider needs to be able to show it.

That means keeping a record of each model release: the version, the documentation, the transparency summary and the information given to downstream providers. What matters is being able to show that the record is the one from that date, unchanged since.

Here, the thing to seal is the release package: the model card, the transparency summary, the documentation and the downstream information sheet, each time a version is made available. WitniumChain takes a fingerprint of each document and seals it, so you can later show anyone that the version you are presenting is exactly the one published on that day. The content stays with you; only the fingerprint is sealed.

Sealing happens at release. If you update the documentation or the summary, that becomes a new sealed version. The point is that you can show a regulator, a downstream provider or a court which version applied when, and that nobody has quietly rewritten it in the meantime.

It doesn't prove that the documentation was complete, or that the training complied with copyright law. It shows that the record you kept at the time is the record you are presenting now.

Anyone you share the record with can check it: calculate the fingerprint of the document again, fetch the sealed one using the witness ID and compare them. If they match, nothing has changed. The first 2,000 witnesses a month are free, through the REST API, the SDK or the MCP server ^[2].

How it would work

1. At each model release, seal the model card, the transparency summary, the technical documentation and the downstream information sheet.
2. Store the witness ID with the release metadata.
3. When asked, share the document and the witness ID. The recipient calculates the fingerprint and checks it against the sealed one.

What this means for you

If you provide a general-purpose AI model, these are the questions the guidelines suggest you should have answers to now:

• Do you have a technical documentation template that covers the items in Annex XI, updated for each release?
• Have you published a sufficiently detailed summary of your training data, and do you have a policy for respecting opt-outs under copyright law?
• Do you give downstream providers the information they need to comply with their obligations, and can you show which version of that information applied at a given time?
• If your model is above the compute threshold, or if the AI Office might designate it, do you have a process for adversarial testing and incident tracking?
• Can you demonstrate, if asked, what your documentation and transparency summary said at the time a particular version was released?

The last question is the one a digital witness addresses. Keeping the documentation is required; being able to prove it hasn't changed is a shorter conversation than explaining why you can't.

What to watch

The copyright transparency obligation applies from 2 August 2025 ^[1]. The wider obligations under Articles 53 and 54 apply from 2 August 2025 as well, except for those in Article 53(1)(d) and (e), which take effect from 2 August 2026 ^[1]. The AI Office has indicated it will issue further guidance on model evaluation and codes of practice, so this is not the final word ^[1].

The first reports from providers on compliance with these obligations are likely to emerge in the second half of 2025. That will be the first test of how the guidance works in practice.