Report from EuroNews
In Brief – The latest phase of the multi-year implementation timeline for the EU AI Act took effect on August 2, including rules regulating new General Purpose AI (GPAI) models and the requirement that EU member states appoint their compliance oversight authorities. Despite the deadline, many member states have not yet notified the European Commission of their roster of AI Act regulators. The challenges facing member state regulators, called “surveillance authorities”, are significant given the technical complexity of the AI systems themselves, the many requirements of the AI Act, and the fact that the new AI regime will interact with existing regulations like the GDPR, Digital Services Act, and Digital Markets Act. Rules regarding GPAI models such as ChatGPT, Gemini, and Grok are also technically taking effect, but they now only apply to GPAI models released after August 2, 2025. Models already in operation are not required to comply with the new law until August 2027.
Context – The AI Act was initially designed as a regulatory regime for applications using AI, not to police the underlying AI technologies. Applications were categorized into four risk tiers — unacceptable, high, limited, and low. Rules prohibiting unacceptable risk systems, such as social scoring regimes or AI intended to deceive or exploit people, went into effect in February. The EU Parliament changed the framework of the AI Act after the emergence of ChatGPT, adding a regulatory regime for general purpose “foundation models”. It was a divisive change, and concerns have grown as other major markets, including the US, UK, and Japan have moved in a deregulatory direction. Top EU business and tech leaders have called on the Commission to delay implementation of the AI Act, but the Commission recently released its GPAI Code of Practice that provides guidance to companies creating advanced models as it keeps to the implementation schedule. Meta is one major foundation model provider that has announced it is not signing onto the code while xAI is signing onto only the chapter on Safety and Security, but not the sections on Copyright or Transparency.