The draft law that outlines a risk-based approach for regulating artificial intelligence applications has received strong support from the European Parliament’s committees on civil liberties (LIBE) and internal market (IMCO). The committees voted 71-8 (with 7 abstentions) this morning to approve the compromise reached with EU Member States after long and complex ‘trilogue’ negotiations at the end of last year.
The EU AI Act, first proposed by the Commission in April 2021, sets rules for AI developers based on the power of their models and/or the purpose for which they want to use AI. It bans some uses of AI (such as social scoring), and sets rules for some high risk uses (e.g. education, health or employment), such as duties in areas like data quality, testing and risk assessment. It also imposes some transparency requirements on general purpose AIs and tools like deepfakes and AI chatbots.
Apart from that, most AI apps will be considered ‘low risk’ — meaning they are not covered by the law. The plan also allows for the creation of regulatory sandboxes at the national level to help developers to create, train and test risky apps in a supervised “real-world” setting.
The Commission’s proposal for an AI rulebook did not attract much attention when the EU presented it three years ago. But with the emergence of generative AI over the past year+ the plan has gained global attention — and created big divisions among the bloc’s lawmakers. MEPs amended the proposal to make sure it covers powerful general purpose AIs, while some Member States, led by France, have been pushing for the opposite — seeking a regulatory exemption for advanced AIs in the hopes of promoting national leaders.
The compromise text that resulted from long trilogue talks in December still had some provisions for general purpose AIs, which caused some governments to remain opposed. And until last month these disagreements seemed like they could still stop the bill’s progress. But after a crucial Member State vote on the compromise text earlier this month the bloc is almost sure to adopt its flagship AI rulebook in a few months.
That said, the draft law still has a few challenges left to overcome before adoption: There’s a plenary vote in the parliament in the next weeks, where MEPs will be asked to officially adopt it. Then there will be a final Council approval after that.
However these last few steps look less likely to cause any conflicts between EU co-legislators. Any such move would be a disaster for the bill in the current cycle, with parliamentary elections coming up and the end of the current college’s term — meaning both legislative time and reputational space are limited.
Today’s strong support by the two parliamentary committees, which were involved in thorough examination of the legislative proposal over years, also gives a strong indication that MEPs will follow through with the absolute majority support needed — which would clear the way for the law to be adopted and come into force later this year. The first provisions (bans on prohibited practices) would then apply six months after that (so probably in the second half of this year).
The EU agreed on a phased implementation of the Act which is likely to see legal requirements increasing on in-scope developers between 2024 and 2027. (The EU is also allowing nine months after entry into force for a code of practice to apply; 12 months after entry into force for rules on general-purpose AI, including governance, to apply. It also says the law will be fully applicable 24 months after entry into force — although obligations for high-risk systems have a longer application timeframe still (of 36 months).)
While the parliament plenary vote looks likely to succeed, some opposition remains. The Pirate Party, for example, is refusing to support what its MEPs — who account for a few of today’s committee votes against the Act — are calling a “flawed” law. In a statement, Pirate Party MEP Marcel Kolaja, also a member of the IMCO committee, said: “Unfortunately, despite the good position of the European Parliament, the national governments managed to weaken the AI Act. Hence, the Pirates cannot support it.”
Another Pirate Party MEP and LIBE committee member, Patrick Breyer, also cautioned: “The EU’s AI Act allows for constant facial surveillance in real time: More than 6,000 people are sought by European arrest warrant for the crimes listed in the AI Act. Any public area in Europe can be subjected to permanent biometric mass surveillance for these reasons. This law authorises and normalises a culture of suspicion. It takes Europe towards a dystopian future of a suspicious high-tech surveillance state.”