Episode Details

Article 53 of Regulation (EU) 2024/1689, the European Union’s Artificial Intelligence Act, bound general purpose AI model providers to new duties on August 2, 2025, a year before the date most coverage names. A founder searching for when the Act binds a product built on a commercial model usually finds the louder, later date first, but that date governs a different set of duties. The one that actually created general purpose AI model obligations already passed, and the European Commission published interpretive guidelines for it in July 2025.

The confusion has a specific cause. Regulation (EU) 2024/1689 entered into force August 1, 2024, and Article 113 phased its provisions across four dates rather than switching the whole statute on at once: prohibited practices and AI literacy duties on February 2, 2025; the AI Office, national authorities, and the general purpose AI model chapter on August 2, 2025; the bulk of the remaining Act, including the high risk AI system rules, on August 2, 2026; and the high risk rules for AI built into already regulated products on August 2, 2027. Two of those four dates fall on the same calendar day a year apart, and coverage of the louder 2026 date has crowded out the quieter 2025 one that actually governs general purpose AI models.

Whether any of this reaches a particular startup turns on a distinction the statute draws with more precision than the headlines suggest: provider versus deployer. Article 3(3) defines a provider as whoever develops a general purpose AI model, or has one developed, and places it on the market under its own name. A startup that sends API calls to GPT, Claude, Gemini, or Llama is, in the overwhelming majority of cases, a deployer of someone else’s model rather than the provider of it, because it neither trained the model nor placed it on the market under its own brand. Deployers sit outside the documentation, copyright policy, and training data summary duties Article 53 places on providers. That is not a loophole. It is the structure the statute chose, because the entity with the technical ability to describe how a model was trained is the entity that trained it, not the entity that calls its endpoint.

The line moves, and startups doing serious model work should know exactly where. The European Commission’s guidelines on the scope of general purpose AI obligations set an indicative marker: a fine tuning run that consumes more than one third of the original model’s training compute produces a new model, and the party that ran that fine tuning becomes its provider. Article 25 supplies a parallel test for the broader category of high risk AI systems. A company becomes a provider if it puts its own name or trademark on a system already placed on the market, if it makes a substantial modification within the meaning of Article 3(23), or if it changes a system’s intended purpose from something that was not high risk into something that is. A startup training a domain specific model on top of an open weight base, or stripping a foundation model’s branding to ship it as a proprietary product, should expect the same answer under either provision: it built enough of the model to own the duties that come with building it.

August 2, 2026 still matters, just not for the reason most search results imply. It is the date Article 113 set for the general application of the remaining provisions of the Act, including the Article 50 transparency duties that require disclosure when a person is interacting with an AI system or viewing content it generated. The high risk AI system rules tied to Annex III, covering uses such as employment screening, credit scoring, and access to essential services, were supposed to apply on the same date until the Digital Omnibus on AI, a set of amendments that entered into force July 27, 2026 after publication in the Official Journal three days earlier, pushed the compliance date for stand alone high risk systems to December