On December 9, the EU Commission announced a political agreement between the European Parliament and the European Council regarding the proposed Artificial Intelligence Act (AI Act).  The agreement is provisional and is subject to finalizing the text and formal approval by lawmakers in the European Parliament and the Council. The AI Act will regulate the development and use of AI systems, as well as impose fines on any non-compliant use. The object of the law is to ensure that AI technology is safe and that its use respects fundamental democratic rights while balancing the need to allow businesses to grow and thrive. The AI Act will also create a new European AI Office to ensure coordination, transparency, and to “supervise the implementation and enforcement of the new rules.” According to this EU Parliament press release, powerful foundation models that pose systemic risks will be subject to specific rules in the final version of the AI Act based on a tiered classification.

Except with foundation models, the EU AI Act adopts a risk-based approach to the regulation of AI systems, classifying these into different risk categories: minimal risk, high-risk, and unacceptable risk. Most AI systems would be deemed as minimal risk since they pose little to no risk to citizens’ safety. High-risk AI systems would be subject to the heaviest obligations, including certifications on the adoption of risk-mitigation systems, data governance, logging of activity, documentation obligations, transparency requirements, human oversight, and cybersecurity standards.  Examples of high-risk AI systems include utility infrastructures, medical devices, institutional admissions, law enforcement, biometric identification and categorization, and emotion recognition systems. AI systems deemed “unacceptable” are those that “present a clear threat to the fundamental rights of people” such as systems that manipulate human behaviors, like “deep fakes,” and any type of social scoring done by governments or companies. While some biometric identification is allowed, “unacceptable” uses include emotional recognition systems at work or by law enforcement agencies (with narrow exceptions).

Sanctions for breach of the law will range from a low of €7.5 million or 1.5 percent of a company’s global total revenue to as high as €35 million or 7 percent of revenue. Once adopted, the law will be effective from early 2026 or later. Compliance will be challenging (the law targets AI systems made available in the EU), and companies should identify whether their use and/or development of such systems will be impacted.