When Integrating a GPAI Model Makes You the Provider

Most companies building on a general-purpose model assume the model provider carries the regulatory weight. Article 25 of the EU AI Act is the provision that decides when it moves to you, and two of its three triggers are things product teams do without noticing.

The three triggers

Article 25(1) of Regulation (EU) 2024/1689 says that a distributor, an importer, a deployer or any other third party shall be considered a provider of a high-risk AI system for the purposes of the Regulation — and so subject to the provider obligations in Article 16 — in any of three circumstances:

• (a) Name or trade mark. They put their name or trade mark on a high-risk AI system already placed on the market or put into service, without prejudice to contractual arrangements stipulating that the obligations are otherwise allocated. This is the white-label and reseller case.
• (b) Substantial modification. They make a substantial modification to a high-risk AI system that has already been placed on the market or put into service, in such a way that it remains high-risk under Article 6.
• (c) Changed intended purpose. They modify the intended purpose of an AI system — including a general-purpose AI system — which has not been classified as high-risk and has already been placed on the market or put into service, in such a way that it becomes high-risk under Article 6.

Limb (c) is the one that catches integrators. It does not require you to modify the model at all; modifying the intended purpose is enough. Take a general-purpose assistant that is not high-risk, point it at CV screening, and you have moved a system into Annex III point 4 — employment — and become its provider, with conformity assessment, technical documentation, quality management, registration and post-market monitoring attaching to you. Regulation (EU) 2024/1689 on EUR-Lex.

Building on a GPAI model

Separate two things that get merged. A general-purpose AI model attracts the Chapter V obligations — Article 53 documentation, copyright policy, training content summary, and the Article 55 systemic-risk duties above the threshold — and those stay with the model provider. An AI system built on that model and placed on the market under your own name is a different artefact with a different provider, and that provider is you. Article 25 is not what makes you the provider of your own system; it is what makes you the provider of somebody else’s.

The practical consequence is that most product companies are already providers of an AI system before Article 25 is reached, and the live question is only whether that system is high-risk. If it is not, the duties are light: the Article 50 transparency obligations if it generates content or interacts with people, and the Article 4 AI literacy duty. If it is, Article 16 is the whole list. The gap between those two outcomes is the Article 6 classification, and it is the decision that deserves the documentation.

Article 6(3) contains a narrow derogation: a system referred to in Annex III is not high-risk where it does not pose a significant risk of harm to health, safety or fundamental rights, including by not materially influencing the outcome of decision-making, and where one of the listed conditions applies — narrow procedural task, improving the result of a previously completed human activity, detecting decision patterns without replacing or influencing human assessment, or a preparatory task. Profiling of natural persons is always high-risk regardless. Relying on the derogation requires a documented assessment before placing on the market, and registration under Article 49(2). It is a real route and it is not a shrug.

What counts as substantial modification

Article 3(23) defines a substantial modification as a change to an AI system after its placing on the market or putting into service which is not foreseen or planned in the initial conformity assessment carried out by the provider, and as a result of which either the system’s compliance with the Chapter III Section 2 requirements is affected, or the intended purpose for which it was assessed is modified.

Two limbs, and the first has a hinge most summaries miss: not foreseen or planned in the initial conformity assessment. A change the provider anticipated and assessed — retraining on a schedule, a documented parameter range, an update within a declared envelope — is not a substantial modification even if it changes behaviour. This is why Article 13(3)(c) requires the instructions for use to describe predetermined changes to the system and its performance: that disclosure is what defines the envelope you can operate in without becoming a provider. Read it as a boundary, not as a footnote. Detail in the substantial modification page.

Against that definition, the common integration changes sort roughly as follows — and this is analysis of the wording, not settled law. Changing a system prompt or adding retrieval over your own documents does not obviously affect Chapter III Section 2 compliance and does not change the intended purpose, so it is unlikely to be substantial on its own. Fine-tuning on your own data is much closer to the line, because accuracy and robustness characteristics under Article 15 are exactly what the requirements cover. Changing a decision threshold in a scoring system, or applying it to a population materially unlike the one it was assessed on, engages the first limb. Deploying it for a new purpose engages the second limb and limb (c) of Article 25(1) directly.

What happens to the original provider

Article 25(2) handles the handover and it is more consequential than it looks. Where one of the Article 25(1) triggers fires, the provider that initially placed the system on the market is no longer considered the provider of that specific system. It does not become jointly responsible; the role transfers. What remains is a cooperation duty: the initial provider shall closely cooperate with the new provider, make available the necessary information, and provide the reasonably expected technical access and other assistance required for the new provider to fulfil its obligations.

There is an exception, and it is the sentence to read twice. That cooperation duty does not apply where the initial provider has clearly specified that its AI system is not to be changed into a high-risk AI system. A vendor can, by saying so plainly in its terms and documentation, switch off its own obligation to hand you the documentation you would need. Many general-purpose model terms now contain exactly such a statement. Check for it before you plan a high-risk deployment on somebody’s general-purpose product, because discovering it after launch means you owe Article 16 duties and have no route to the technical documentation to discharge them.

The related supply-chain provision is Article 25(4): third parties supplying AI systems, tools, services, components or processes used or integrated in a high-risk AI system must specify by written agreement the information, capabilities, technical access and assistance the provider needs to comply — excluding third parties making tools, services, components or processes other than general-purpose AI models accessible under a free and open-source licence. Article 25(3) separately makes the product manufacturer the provider where a high-risk system is a safety component of a product covered by Annex I Section A and placed on the market under the manufacturer’s name.

What is not settled

Several questions here are genuinely open, and a page that pretends otherwise is not useful.

What the contractual carve-out in 25(1)(a) does. The words “without prejudice to contractual arrangements stipulating that the obligations are otherwise allocated” sit inside a provision that otherwise allocates a public-law obligation. One reading is that parties can allocate the work and the cost between themselves while the regulator can still address the party whose name is on the product; another is that the carve-out is capable of shifting the designation itself. The first reading is the safer basis on which to plan. What would settle it is Commission guidance or a market surveillance decision on a white-label case.

Where fine-tuning sits. There is no threshold in the text distinguishing a light adaptation from a modification affecting Section 2 compliance, and the answer plainly depends on what the tuning does to accuracy and robustness. Until guidance or standards give a test, the defensible approach is to evaluate against the provider’s declared metrics before and after, and keep the comparison.

How the roles compose across a long chain. A model provider, a platform that hosts it, an integrator that builds a system, a reseller that rebrands it and an enterprise that deploys it can be five entities, and the Act allocates roles pairwise rather than across the chain. The provider and deployer page sets out the base roles, and the importer and distributor obligations cover the two intermediate ones. In a chain of five, expect at least one boundary where both parties believe the other is the provider; the cheapest time to resolve that is in the contract, in writing, with the Article 25 limbs named.