Updates to the proposed law would require developers to meet transparency requirements before releasing foundational models; create regulatory sandboxes, allowing firms to test AI products in real-world conditions; and limit but not ban police services from using live biometric identification systems such as facial recognition. (The Logic)

Talking point: The bloc began working on the act in April 2021, but had to rework it for ChatGPT and other generative systems. France, Germany and Italy have lately been pushing for a more innovation-facilitative (or industry-friendly, depending who you ask) approach, particularly on foundation models. Rights groups panned the final version as insufficiently protective, while tech groups and executives panned it as overly restrictive. Both Parliament and Council still need to adopt the text, but the political deal struck late Friday means it will take effect in 2026. That would put it in line with or ahead of the fastest timeline for Canada’s proposed Artificial Intelligence and Data Act.