Paris: The European Commission can now investigate and penalise the companies behind Europe’s most powerful artificial-intelligence systems, after the enforcement phase of the AI Act opened on 2 August.
From that date, the Commission’s newly built AI Office holds the authority to demand technical documents, obtain access to models for evaluation, order corrective measures and levy fines. The powers apply to providers of general-purpose AI, the foundation models that sit beneath chatbots, coding assistants and image generators used across the continent.
The stakes are considerable. A provider that ignores the rules risks a penalty of up to 15 million euros or three percent of its worldwide annual turnover, whichever is higher. For the largest developers, that percentage dwarfs the fixed cap and turns compliance into a board-level concern rather than a legal footnote.
Officials designed the timeline to bite unevenly. Enforcement now covers general-purpose models that entered the EU market after 2 August 2025. Developers whose models arrived on or before that date receive a longer runway and face the same obligations only from 2 August 2027, giving established players time to document training data, energy use and safety testing.
To smooth the transition, the Commission published a voluntary General-Purpose AI Code of Practice a year earlier. Firms that sign and follow it earn a presumption of conformity, meaning regulators will treat adherence as evidence that a model meets the law. Signatories also shoulder lighter paperwork, a carrot the Commission hopes will pull the industry toward disclosure rather than confrontation.
Paris watches the moment closely. France has staked much of its digital ambition on home-grown model builders, and their leaders have argued for months that heavy rules could hand an advantage to better-funded American rivals. The government counters that clear obligations, applied to everyone at once, give European firms a stable market in which to compete.
Transparency requirements form the core of the new duties. Providers must summarise the content used to train a model, respect European copyright law, and share safety information with regulators and, in some cases, with the businesses that build products on top of their systems. Models judged to carry systemic risk face extra testing and incident-reporting obligations.
Enforcement will not arrive as a wave of fines. The AI Office has signalled that it prefers to open dialogue, request information and push for fixes before reaching for penalties. Yet the mere existence of binding powers changes the calculation for developers who until now treated European guidance as optional.
Industry groups warn that the rulebook still leaves grey areas, particularly over what counts as a modified model and how far obligations follow open-weight systems that anyone can download and adapt. The Commission promises further guidance, but the clock is already running for the newest and most capable models.
For Europe, the day marks a shift from writing rules to applying them. The bloc gambled that it could shape global AI through law rather than through the raw scale of its own companies. Whether that bet pays off now depends on how firmly, and how fairly, the AI Office chooses to act.




