Until now, the EU AI Act has largely been a distant prospect. That will change on August 2: despite the postponement of the strictest rules, providers of AI models will be required to submit detailed information to the European Commission starting next Monday if the Commission requests it. The goal is to ensure that AI systems remain transparent, measurable, and safe. What do these changes mean in practice?
Starting August 2, this coming Sunday, the Commission may request all kinds of information about AI models. This includes, among other things, summaries of AI training methods, safety mechanisms, and protections against copyright infringement. The Commission may require changes, remove LLMs from the EU market, and impose fines of up to 3 percent of annual revenue.
An evolving law
It is noteworthy that the EU AI Act was proposed well before the introduction of ChatGPT. Approval and the first tangible impact followed in 2024. Much changed between the initial presentation of the EU AI Act and its enactment into law: LLMs evolved from experimental tools to the engine driving the ever-expanding global expansion of IT infrastructure. The advantage for lawmakers: although AI is already having a societal impact, the technology is still in its infancy. Many organizations have yet to deeply integrate AI, which gives regulators the opportunity to establish frameworks for when its adoption does gain momentum.
Even superficial use of AI can be significantly influenced by EU law. We’ve seen before that existing legislation, such as the GDPR, put a stop to the simultaneous rollout of a new LLM in Europe and North America, where restrictive laws in this area have not been enacted.
In the U.S., until recently, the policy was dominated by the major “frontier labs,” namely OpenAI, Anthropic, and (occasionally) Google. Since the temporary export ban on Claude Fable 5, the U.S. government has become more active and restrictive in regulating AI models. For now, however, this remains an ad hoc process without clear rules. In Europe, the law is clearer but still evolving.
The rules effective August 2
As of August 2, several new rules that had been missing until now have taken effect. High-risk AI systems must now meet strict compliance criteria. This includes AI used to calculate credit scores, determine school grades, support police work, screen resumes, and manage critical infrastructure. In other words, AI that can have a decisive impact at socially and economically significant intersections. AI providers must conduct impact assessments, adhere to protocols, log various data, and confirm compliance through CE marking.
Systems that interact directly with people, such as customer service chatbots or standard chat apps, must make it clear that it is AI, not a human, communicating with the user. AI-generated media must also include a label indicating that it is AI. Examples of the above include the fact that Claude, Gemini, and ChatGPT display a short message below the chat window stating that the application is AI-driven and may make errors, or the Gemini star located in the bottom right corner of AI-generated images on Google.
EU member states must use their own AI “sandbox” to conduct controlled studies, so that national regulators have the opportunity to assess whether AI systems require further local intervention.
Not the end
AI systems that indiscriminately scan faces or other software that uses AI for far-reaching purposes have been banned under the EU AI Act since February. Previous laws already covered such uses, but under this new legislation, AI systems are classified based on risk, with the least desirable forms of AI excluded from the outset. Incidentally, for low-impact uses of AI, such as spam filters, search engines, or AI in games, nothing changes under the EU AI Act.
Existing AI systems deployed before August 2 have until December 2 to comply with the new rules. Exactly one year from now, a wave of AI regulations will follow for critical components, such as medical devices and cars. It won’t be until 2030, again on August 2, that legislation will be introduced to regulate high-risk AI systems introduced before August 2, 2026.
See also: As Fable 5 returns, Anthropic wants to write the frontier AI rulebook