EU moves from lawmaking to implementation
The European Union is shifting the AI Act from headline rules to detailed implementation, issuing obligations and guidance for providers of general‑purpose models and closing consultation on procedural safeguards. That transition makes compliance more operational—companies will soon face actionable expectations on model access, documentation and risk controls rather than only broad regulatory principles. Firms selling or using large models in Europe should expect a wave of rule‑level detail to arrive alongside enforcement timelines. (dig.watch (dig.watch))
Europe’s Artificial Intelligence Act is no longer just a big rulebook passed in Brussels. Since 2 August 2025, the European Union’s obligations for general-purpose artificial intelligence models have been in force, and the Commission has started publishing the documents companies will actually use to comply. (digital-strategy.ec.europa.eu) That changes the question for companies like OpenAI, Google, Meta, Anthropic, and Mistral. The old question was whether the law would pass; the new one is what files, notices, summaries, and risk controls they must hand over when they sell a model into Europe. (eur-lex.europa.eu) (digital-strategy.ec.europa.eu) The European Union uses “general-purpose artificial intelligence model” for a model that can be reused for many jobs, the way one engine can power many different cars. The law puts one layer of duties on all of those model providers, then adds a tougher layer for the small set of models the law treats as carrying “systemic risk.” (digital-strategy.ec.europa.eu 1) (digital-strategy.ec.europa.eu 2) For the broad group of model providers, the Act points to concrete paperwork. The European Union’s own summary says providers need technical documentation, information for downstream developers, and a public summary about the data used to train the model. (eur-lex.europa.eu) For the most powerful models, the list gets heavier. The Code of Practice says the safety and security chapter is aimed only at providers covered by Article 55, which is the part of the law for general-purpose models with systemic risk. (digital-strategy.ec.europa.eu) The Commission moved this from theory to operations in July 2025. On 18 July 2025 it published guidelines on the scope of the obligations for providers of general-purpose models, and on 10 July 2025 the General-Purpose Artificial Intelligence Code of Practice was published as a voluntary compliance tool. (digital-strategy.ec.europa.eu 1) (digital-strategy.ec.europa.eu 2) That code was not written by one office behind closed doors. The Commission says the final version it received on 24 September 2025 was developed by 13 independent experts with input from more than 1,000 stakeholders, including model providers, small businesses, academics, safety experts, rightsholders, and civil society groups. (digital-strategy.ec.europa.eu) Now the next layer is arriving through procedure. A Commission consultation on a draft implementing regulation covering detailed arrangements for certain Artificial Intelligence Act proceedings closed on 8 April 2026, and the draft covers evaluation of general-purpose models under Article 92, including how independent experts can be involved and selected. (dig.watch) That sounds dry until you translate it into company behavior. If Brussels writes the steps for model access, evidence handling, expert selection, and case procedure in detail, an investigation stops being an abstract threat and starts looking like a checklist with deadlines. (dig.watch) (digital-strategy.ec.europa.eu) The office running much of this is the European Artificial Intelligence Office inside the Commission. The Commission says that office plays a key role in implementing the Act, especially for general-purpose artificial intelligence, and it enforces the rules for those models while national authorities handle other parts of the system. (digital-strategy.ec.europa.eu 1) (digital-strategy.ec.europa.eu 2) The timetable is staggered, not one big launch day. The Commission says the ban on unacceptable-risk practices and the obligations for general-purpose models are already applicable, while many other provisions, especially for high-risk systems, start later, with all rules applying by 2 August 2027. (eur-lex.europa.eu) So the story in Europe is no longer “the Artificial Intelligence Act exists.” The story is that the Commission is filling in the operating manual line by line, and companies that build or package large models for the European market are moving from reading principles to preparing evidence. (digital-strategy.ec.europa.eu) (digital-strategy.ec.europa.eu)
