August 2, 2026: The AI Act GPAI becomes enforceable – day one of a still under-equipped regime

Since today, transparency in training, disclosure of protected sourcing, and systemic risk management are no longer recommendations but legal obligations for general-purpose AI models. The question remains whether the AI Office, barely established, will be able to translate this mandate into case law—under the explicit threat of an American response.

In plain terms. Starting August 2, 2026, European rules on general-purpose AI models (GPAI) become enforceable. In concrete terms: large models must document their training, disclose protected content used, and manage systemic risks. On paper, a turning point; in practice, enforcement is the real debate.

Context

The AI Act came into force in 2024, with a phased implementation schedule. August 2, 2026 marks the specific shift for GPAI obligations from theoretical to enforceable. Ahead of this, the Commission adopted a voluntary Code of Practice; according to Euronews, most major Western labs have signed it—with the notable exception of Meta.

The facts (day one)

Three immediate concrete requirements for GPAI developers:

1. transparency about model construction;
2. disclosure of copyrighted content used for training;
3. documentation of capabilities and limitations for downstream users.

For frontier models, an additional obligation to identify and mitigate societal risks. The European AI Office is the body created to oversee enforcement. OpenAI, cited by Euronews, states it has worked closely with the Commission and intends to continue collaboration.

Analysis

The text is not the debate; enforcement is. Euronews highlights two structural frictions:

• a strained AI talent pool, making recruitment for the AI Office difficult;
• rapidly evolving technology—frontier models remain a moving target that outpaces doctrine.

Beyond technical friction, geopolitical lines are drawn: MEP Michael McNamara warns, in the same article, of the risk that Washington treats the regime as an attack on U.S. commercial interests. This risk is new in this cycle and directly impacts the enforcement window.

Probabilized scenarios

• Progressive and negotiated enforcement (~55%): the voluntary Code serves as a ramp, few sanctions in the first year, doctrine built case by case.
• Transatlantic blockade (~30%): U.S. commercial retaliation, delayed enforcement or dilution of frontier requirements.
• Emblematic case sanctioned (~15%): hard jurisprudence against a test lab—likely a non-signatory.

Implications for the pro

If you deploy an LLM in Europe, the question is no longer “am I covered?” but “can I document my training pipeline, sourcing, and risks?” Mapping data becomes a legal requirement, not just ethical. Non-signatories of the Code—Meta first among them—are now exposed.

To watch

First hearings by the AI Office; Washington’s response by mid-August; Meta/xAI positioning; first emblematic case.

So what. The regulator shifts from text to test. The real challenge: its ability to sanction a transatlantic giant without triggering a rupture.