General-Purpose AI (GPAI) Under the EU AI Act

Key takeaways

• A general-purpose AI model is the EU AI Act’s legal category for broadly capable models, defined in Article 3(63): significant generality, able to perform a wide range of distinct tasks, and built to be integrated into other systems.
• The European Commission treats a model as general-purpose when its training compute passes roughly 10^23 floating point operations and it can generate language, images, or video, though genuine generality matters more than the number.
• A second, higher line at 10^25 floating point operations marks general-purpose AI with systemic risk, which carries extra duties under Article 55.
• Every provider faces four baseline obligations under Article 53: technical documentation, information for downstream users, a copyright policy, and a public summary of training content.
• The General-Purpose AI Code of Practice, published on 10 July 2025, is the practical route to showing compliance, and the core obligations have applied since 2 August 2025.

What is a general-purpose AI model?

A general-purpose AI model, often shortened to GPAI, is a model that learns broad capability from large volumes of data and can then be pointed at many different tasks. The EU AI Act gives it a precise legal meaning in Article 3(63): a model that “displays significant generality” and is “capable of competently performing a wide range of distinct tasks,” that can be integrated into a variety of downstream systems or applications. The large language models behind familiar assistants are the clearest examples, alongside image and multimodal models from the major labs. The definition matters because it decides who carries the obligations. The Act regulates the model itself, separately from any finished product built on top of it. A laboratory that trains a model and places it on the EU market is a provider of a general-purpose AI model, even if that organisation never ships a consumer application. For a wider view of how the Act fits together, see the AI Sigil platform.

GPAI vs foundation models vs generative AI

These three terms describe the same technology from different angles, and mixing them up leads to compliance mistakes.

• General-purpose AI is the legal term used by the EU AI Act. It is what triggers the obligations in this article.
• Foundation model is the technical description: a large model trained on broad data that serves as a base for many downstream uses. Most foundation models are general-purpose AI models in the legal sense.
• Generative AI describes a capability, producing new text, images, audio, or code. A generative system is usually built on a general-purpose model, but “generative” is a feature, not the legal category.

When you map obligations, anchor on the legal term. A model is in scope because it is general-purpose under Article 3(63), not because someone called it generative in a product brochure.

How the AI Act decides if your model is general-purpose

Compute is the first signal. In its guidelines published on 18 July 2025, the European Commission set an indicative line: a model is presumed to be general-purpose when the cumulative training compute is above 10^23 floating point operations and the model can generate language (text or audio), text-to-image, or text-to-video output. That roughly corresponds to models trained with around a billion parameters on large datasets. The number is a starting point, not a verdict. The Commission is explicit that generality is decisive. A model that crosses the compute line but only does one narrow job, such as transcribing speech or generating music, may sit outside the category. A model below the line that shows real breadth can still qualify. The test reads capability and range first, and uses compute as evidence rather than as the rule. For most organisations that buy or fine-tune models rather than train them from scratch, the practical question is not whether they crossed 10^23 floating point operations. It is whether they did enough to become a provider in their own right, which the value-chain section below addresses.

Two tiers: standard GPAI and GPAI with systemic risk

The Act splits general-purpose AI into two tiers, and the second carries materially heavier duties. Article 51 sets the line for systemic risk. A model is presumed to have “high impact capabilities” when “the cumulative amount of computation used for its training measured in floating point operations is greater than 10^25.” A model can also be designated by a Commission decision, made on its own initiative or after an alert from the scientific panel, using the criteria in Annex XIII. Article 51(3) lets the Commission move these thresholds as hardware and algorithms improve, so the line is meant to track the frontier rather than stay fixed. The two numbers do different jobs, and they are easy to confuse:

• 10^23 floating point operations: the indicative trigger for being a general-purpose AI model at all, from the Commission guidelines.
• 10^25 floating point operations: the trigger for the systemic-risk tier, written into Article 51.

A handful of frontier models cross the second line. The vast majority of general-purpose models sit in the standard tier, with the Article 53 obligations only. Knowing which tier applies is the first compliance decision, because it sets the size of everything that follows.

What every GPAI provider must do (Article 53)

Article 53(1) sets four baseline obligations that apply to every provider of a general-purpose AI model, regardless of tier.

1. Technical documentation. Providers must “draw up and keep up-to-date the technical documentation of the model, including its training and testing process,” following Annex XI, and make it available to the AI Office and national authorities on request.
2. Information for downstream providers. Providers must prepare and maintain documentation for the organisations that build on the model, set out in Annex XII, so those downstream providers understand the model’s capabilities and limitations and can meet their own obligations.
3. Copyright policy. Providers must “put in place a policy to comply with Union law on copyright and related rights,” including identifying and respecting the rights reservations expressed under Article 4(3) of the Copyright in the Digital Single Market Directive.
4. Public training-content summary. Providers must “draw up and make publicly available a sufficiently detailed summary about the content used for training,” using the template provided by the AI Office.

Together these turn a model from a black box into a documented, accountable artefact. Two of the four obligations face outward, toward downstream builders and the public, which is why they tend to take the most preparation.

The open-source carve-out (and its limit)

The Act gives open models a partial break. Under Article 53(2), a model released under “a free and open-source licence” that allows access, use, modification, and distribution, with its parameters, architecture, and usage information made public, is exempt from the first two obligations: the technical documentation and the downstream information pack. The copyright policy and the training-content summary still apply. There is a hard limit. The carve-out “shall not apply to general-purpose AI models with systemic risks.” A frontier open-weight model that crosses the 10^25 line carries the full obligation set, open licence or not. Open release reduces paperwork at the standard tier; it does not exempt a systemic-risk model from anything.

Extra duties for systemic-risk models (Article 55)

Providers at the systemic-risk tier take on four further obligations on top of Article 53, set out in Article 55.

• Model evaluation. Providers must “perform model evaluation in accordance with standardised protocols and tools,” including “adversarial testing of the model with a view to identifying and mitigating systemic risks.” Red-teaming moves from good practice to a legal expectation.
• Risk assessment and mitigation. Providers must “assess and mitigate possible systemic risks at Union level,” including their sources, across development, market placement, and use.
• Serious-incident reporting. Providers must “keep track of, document, and report, without undue delay, to the AI Office and, as appropriate, to national competent authorities, relevant information about serious incidents,” along with possible corrective measures.
• Cybersecurity. Providers must “ensure an adequate level of cybersecurity protection for the general-purpose AI model with systemic risk and the physical infrastructure of the model.”

These duties read like a risk-management programme, which is why frameworks such as the UC Berkeley Center for Long-Term Cybersecurity profile for general-purpose and foundation models, which extends the NIST AI Risk Management Framework, map cleanly onto them. For organisations already running structured risk processes, the work is to extend them to the model level rather than to start over. Our AI risk management approach follows the same logic.

The GPAI Code of Practice: the fast path to compliance

The General-Purpose AI Code of Practice, published on 10 July 2025, is the instrument most providers will actually use. It is voluntary, but signing up gives a presumption of conformity with the Article 53 and Article 55 obligations until harmonised European standards exist. In practice it is the shortest credible route to demonstrating compliance. The Code has three chapters:

• Transparency. This chapter provides a Model Documentation Form that providers complete once and share with the AI Office and downstream providers. Completing it acts as presumptive compliance with the documentation duties in Article 53 and Article 54.
• Copyright. This chapter sets out how providers operationalise the copyright policy required by Article 53, including respecting text-and-data-mining reservations.
• Safety and Security. This chapter applies to systemic-risk providers. It describes a Safety and Security Framework covering model evaluations, red-teaming, post-market monitoring, cybersecurity, incident reporting, and accountability, which together satisfy the Article 55 duties.

Signing the Code does not remove the underlying obligations. It gives providers a recognised way to evidence them, which is far easier than defending a bespoke approach to a regulator.

Provider, deployer, or downstream provider?

The most common practitioner question is not about thresholds at all. It is “which role are we?” The Act distinguishes the provider of a model from the deployer who uses it, and the answer changes the obligations dramatically. Most organisations are deployers. They license a general-purpose model through an API or an enterprise agreement and build on top of it. A deployer does not inherit the Article 53 provider duties, but it does need the provider’s downstream documentation to meet its own obligations, especially if its finished system is high-risk under other parts of the Act. The trap is fine-tuning. A downstream actor who modifies a general-purpose model can themselves become a provider of a general-purpose AI model, and then carries the Article 53 duties for the modified model. The Commission’s guidelines use an indicator: a modification that uses more than roughly a third of the original model’s training compute is the kind of change that crosses the line. Most light fine-tuning stays well below that. Substantial retraining can tip an organisation from deployer to provider without anyone deciding it on purpose, which is exactly why the role assessment belongs in the governance process and not in a single engineer’s head. A practical rule: document the role decision, record the compute used in any fine-tuning, and keep the provider’s downstream information pack on file. Those three habits answer most of the questions a regulator or a customer will ask.

Timeline, enforcement and penalties

The dates are settled. The AI Act entered into force on 1 August 2024. The general-purpose AI obligations began to apply on 2 August 2025, one year later. Models already on the market before that date have a longer runway, with compliance required by 2 August 2027. Enforcement of the general-purpose rules sits with the European AI Office, which oversees providers directly rather than through national authorities alone. The Office can request documentation, evaluate models, and require corrective action. The penalties are sized to matter. Under Article 101, the Commission can fine a provider of a general-purpose AI model up to 3% of its total worldwide annual turnover or 15 million euros, whichever is higher, for breaching the obligations or failing to respond to a request. For a large model provider, that is a board-level number, and it is the reason GPAI compliance has moved from legal review into core product planning.

Operationalising GPAI compliance

The obligations look manageable once each one is tied to a concrete artefact:

• Technical documentation maps to a maintained model dossier (Annex XI).
• Downstream information maps to an information pack you can hand to every customer (Annex XII).
• The copyright policy maps to a written, enforced data-sourcing rule.
• The training summary maps to a published document on the AI Office template.
• For systemic-risk models, evaluation, incident reporting, and cybersecurity map to an ongoing risk-management programme.

The difficulty is not creating these once. It is keeping them current as the model is retrained, as customers change, and as incidents occur. A governance platform that stores each artefact, tracks its owner, and flags when it falls out of date turns a set of legal duties into a repeatable process. That is the work AI Sigil is built for: holding the documentation, the risk records, and the evidence in one place so a compliance claim can be produced on demand.

FAQ

Is ChatGPT a general-purpose AI? The model behind ChatGPT is a general-purpose AI model under the EU AI Act, because it shows significant generality and performs a wide range of distinct tasks. The chatbot product is a system built on that model. The provider of the underlying model carries the Article 53 obligations; an organisation that merely uses the assistant is a deployer. What is the difference between general-purpose AI and AGI? General-purpose AI is a legal and practical category for today’s broadly capable models. Artificial general intelligence (AGI) is a hypothetical future system that matches or exceeds human ability across essentially all tasks. The EU AI Act regulates general-purpose AI that exists now; it does not define or regulate AGI as a separate thing. What counts as a general-purpose AI model with systemic risk? A general-purpose model is presumed to have systemic risk when its cumulative training compute exceeds 10^25 floating point operations, under Article 51. The Commission can also designate a model based on the criteria in Annex XIII. Systemic-risk models carry the extra Article 55 duties on top of the baseline obligations. Do open-source AI models have to comply? Partly. Under Article 53(2), a genuinely free and open-source model is exempt from the technical documentation and downstream information obligations, but still needs a copyright policy and a public training-content summary. The exemption does not apply at all to systemic-risk models, which must comply in full. Does using a general-purpose model make my company a provider? Usually not. Licensing a model and building on it makes you a deployer. You can become a provider if you substantially modify or fine-tune the model, with the Commission pointing to changes using more than about a third of the original training compute as the kind of modification that crosses the line. Record what you do so the role is clear. How do providers show they comply? The most direct route is the General-Purpose AI Code of Practice, published on 10 July 2025. Adhering to it gives a presumption of conformity with the Article 53 and Article 55 obligations until harmonised standards arrive, and its Model Documentation Form provides a ready template for the documentation duties.

Conclusion

General-purpose AI is no longer a loosely defined buzzword. Under the EU AI Act it is a legal category with a clear definition, two compute thresholds, and a defined stack of obligations that has applied since August 2025. The path through it is orderly: confirm whether your model is general-purpose, decide whether it reaches the systemic-risk tier, work the Article 53 and, where relevant, Article 55 duties, and use the Code of Practice to evidence them. The organisations that handle this well treat it as a documentation and risk-management problem with owners and deadlines, not a one-off legal exercise. Start with classification, then build the artefacts once and keep them current.