Brussels: Europe’s AI rulebook is about to stop being theoretical. From 2 August the Commission’s AI Office gains the power to enforce the general-purpose obligations of the AI Act against the largest model providers, a full year after those rules first took effect. The date marks the first real test of whether Europe can police the technology it moved so quickly to regulate.
The obligations themselves are not new. Since August 2025, providers of general-purpose models have faced requirements on transparency, copyright and, for the most capable systems, safety and security. What changes now is teeth. Until this point the rules lived largely on paper, cushioned by a voluntary code of practice that many big labs signed but few regulators could yet compel.
Voluntary code, real deadline
That code, published last year, gives companies a way to show they comply while formal European standards are still being drafted. It splits into chapters on transparency, copyright and safety, and adherence lets a provider demonstrate good faith to the AI Office. Crucially, it is not binding. A firm can ignore it and argue compliance another way, but from August the Office can start asking hard questions either way.
The copyright chapter is the most contested. It asks model builders to respect text-and-data-mining opt-outs and to document their training sources, a demand that cuts to the heart of how these systems are built. Publishers and rights holders want proof their work was not scraped without consent. Developers counter that full disclosure is impractical and commercially ruinous. The Office now sits in the middle of that fight.
Enforcement without a template
The deeper challenge is capacity. The AI Office is young, staffing up against the best-resourced engineering teams in the world, and it must interpret rules that reach across borders and business models. Enforcement against existing models is deferred to 2027, giving incumbents breathing room, but new releases from August must comply from day one. That two-speed design invites disputes over what counts as new.
What Europe does next will be watched far beyond its borders. No other major jurisdiction has tried to regulate foundation models this directly, and every fine or forbearance sets a precedent others will study. Move too hard and the bloc risks pushing model releases offshore and confirming complaints that it regulates what it cannot build. Move too softly and the rulebook looks toothless, its credibility spent before it is tested.
Industry has spent the year lobbying over precisely these questions, and the responses have split the sector. The largest developers, keen to shape the rules they will live under, largely signed the code and now press for predictable, proportionate enforcement. Smaller open-source projects worry that the same obligations will land hardest on those least able to afford armies of compliance lawyers, chilling the very research Europe says it wants to keep at home.
The likeliest path is cautious. Expect the Office to open with information requests and dialogue rather than headline penalties, building cases slowly while it finds its feet. The August deadline is less a starting gun than a signal that the grace period is ending, and that Europe’s ambition to be the world’s AI referee is finally meeting the messy business of enforcement.
