On August 2, 2026, the European Commission's AI Office gains the power to demand documentation from any company that places a general-purpose AI model on the EU market, to evaluate those models directly, to order corrective measures, to restrict or withdraw models from the European market — and to impose fines of up to €15 million or 3 percent of global annual revenue, whichever is higher.
Most Americans have never heard of this date. Many American AI companies are subject to it anyway.
The rules have been law for a year. Now they get teeth.
The EU AI Act entered into force in August 2024, and its obligations for general-purpose AI — the category that includes every major foundation model — took legal effect on August 2, 2025. Since that date, providers have been required to do three things under Article 53:
Maintain a copyright compliance policy. Providers must have a policy for complying with EU copyright law, including honoring the rights reservations that European law allows creators and publishers to attach to their works. When a rights holder opts their content out of text-and-data mining in a machine-readable way, a model provider is expected to respect that reservation — regardless of where in the world the training happened.
Publish a training-data summary. Using a mandatory template issued by the European Commission in July 2025, providers must publish a “sufficiently detailed summary” of the content used to train their models. Models placed on the market before August 2025 have until August 2027 to comply, but every new model faces the requirement now.
Document and disclose. Technical documentation covering architecture, training methodology, and limitations must be maintained and made available — to the AI Office on request, and to downstream companies building on the model.
For the past year, these duties existed largely on paper, because the AI Office had no formal enforcement authority. That is what changes in two weeks. August 2, 2026 does not create new obligations; it makes the existing ones enforceable.
The delay that didn't happen
There has been genuine confusion about this, and it is worth clearing up. In June 2026, the European Parliament approved the so-called Digital Omnibus amendments by a wide margin, pushing many of the AI Act's high-risk system obligations — rules governing AI in hiring, credit scoring, biometrics — out to late 2027 and 2028.
What the amendments did not touch: the general-purpose AI obligations, the AI Office's enforcement powers, or the transparency duties around AI-generated content. Companies that read the headlines about “EU AI Act delayed” and concluded the training-data rules were postponed have read the situation wrong. The August 2026 deadline got narrower. It did not go away.
Why this reaches across the Atlantic
The AI Act binds any provider that places a model on the EU market, regardless of where the company is based or where the model was trained. A US lab serving European users is in scope. So is an American enterprise deploying a fine-tuned model to European customers.
This extraterritorial reach is deliberate — the EU did not want its own AI companies bound by rules their foreign competitors could ignore — and it has a practical consequence: EU transparency requirements are becoming the de facto global floor for how training data is documented. It is difficult to maintain one data-governance regime for Europe and a different one for everywhere else. In practice, the strictest jurisdiction sets the standard.
American creators should understand the other side of this coin. The rights-reservation mechanism at the heart of the Act's copyright provisions is not limited to European works. Content owned by American photographers, publishers, analysts, and engineering firms circulates in the same training corpora. As models placed on the EU market are required to demonstrate how they handled rights reservations, the question “whose work is in this model, and on what terms?” stops being rhetorical and becomes a documentation requirement with a fine attached.
An opt-out is not a business model
Here is the structural problem the AI Act exposes without solving. The law gives rights holders a way to say no — the machine-readable opt-out — and gives AI companies a duty to document what they used. What it does not provide is the machinery for saying yes, on these terms: a standard way for millions of individual rights holders to express licensing terms, for AI companies to license at scale without thousands of bilateral negotiations, and for usage and payment to be recorded in a form a regulator will accept.
Without that machinery, the likely outcome is the worst of both worlds: large publishers negotiate bespoke deals, everyone else opts out or is used without compensation, and AI companies assemble compliance documentation by hand, deal by deal.
The historical parallel is instructive. When commerce moved onto payment cards, the solution was not a million bilateral agreements between merchants and banks — it was shared clearing infrastructure that both sides could trust precisely because it belonged to neither. Radio faced the same problem a century ago and solved it with collective licensing and usage-based settlement. AI training and inference are simply the newest version of an old question: how do many parties who have never met transact at scale, with a record everyone accepts?
That is the infrastructure gap Alltio is built to close: a registry where any rights holder — not just the ones with lawyers — puts ownership and terms on the record; standardized contracts that let an AI company license the long tail in one agreement; and cryptographically signed usage reporting that produces, as a byproduct of normal operation, exactly the documented evidence the AI Act's transparency obligations require. Alltio is not the auditor. It is the infrastructure that makes auditing tractable — and that turns a compliance deadline into a functioning market.
August 2 is two weeks away. The companies treating it as a paperwork problem will spend the next several years reconstructing evidence. The ones treating it as market structure will spend that time building on it.