The AI Act is now fully in force. Who kept up, and who fell behind

The second half of 2026 closes out the timetable of the EU Artificial Intelligence Act. Under the official implementation roadmap, the rest of the regulation took effect on August 2, 2026. That is the deadline for companies that had been putting off preparations for two years.

Transparency, or Article 50 in practice

The most visible effect is the duty to label AI-generated content. From August 2, providers of generative systems must make sure that what they produce, text included, carries a machine-readable mark and is recognisable as generated. The rule makes no exception for models from outside Europe, so it also covers Chinese and American models entering the EU market.

Anthropic's watermark is one of the first implementations. The company said the mark is carried by Claude Fable 5.1, Claude Mythos 5.1, Claude Opus 5.5 and Claude Opus 5, and that older models will carry it by December 2. Anthropic applies the watermark at the model level, and the user cannot switch it off. The company notes that its evidentiary value is limited.

The EU AI Act was the direct reason for introducing the watermark, though the company applies it more broadly than the law requires.

Diplomacy runs ahead of technology

At the same time, the dispute over oversight moved to the diplomatic stage. On September 21, on the margins of the UN General Assembly, leaders from more than twenty countries called for stronger human oversight of the most advanced AI systems. Germany and Spain were among the signatories. France, despite being one of the main centres of European AI, did not sign, which the French trade press noted as a surprise.

The tension between Paris and the rest of the Union has a practical dimension. France is pushing for its own AI champions and for infrastructure funding, and every new obligation means another cost for domestic laboratories. Other states now wonder whether fragmented positions will weaken Europe's hand in negotiations with Washington and Beijing.

Poland, like any market without a frontier model of its own, is a recipient of other people's decisions in this puzzle. What counts, then, is not so much who wrote the rule as whether it can be enforced against a provider headquartered on another continent. For now the answer is: partly. The debate over oversight runs in parallel in Brussels, New York and Beijing, but it is settled in practice in the model laboratories.

Article 50 of the act obliges providers to mark AI-generated content in a machine-readable way, to label deepfakes and to inform users when they are talking to a system rather than a human. The obligations apply to every model placed on the EU market, regardless of where it was trained. For the same reason the requirement also covers open models, such as DeepSeek V4.1 Flash, released on September 10 under the MIT licence, which the FLASH24 editorial team refers to.

Anthropic says the watermark is applied at the model level and the user cannot switch it off. It will cover text first, and the company is working on extending the mechanism to images and video. European oversight of synthetic content is therefore no longer a declaration but a question of auditability: without a readable marker, the supervisory authority has nothing to check.