EU Lawmakers Vote For Tougher AI Rules As Draft Moves to Final Stage

European Union lawmakers on Wednesday agreed on changes to draft artificial intelligence rules to include a ban on the use of the technology in biometric surveillance and for generative AI systems like ChatGPT to disclose AI-generated content. The amendments to the EU Commission’s proposed landmark law aimed at protecting citizens from the dangers of the technology will set up a clash with EU countries opposed to a total ban on facial recognition or other tech that could be used for spying. However, they would still have the option to use such technologies for national security and military purposes. The compromise will also boost the right of citizens to file complaints against AI that negatively impacts them. It will also reform the role of the EU’s AI Office, which monitors how the rulebook is implemented.

The changes, approved by two committees overseeing the legislation, were primarily a response to concerns raised by civil society groups. The new text will emphasize the harms of AI and require that developers of high-risk systems are transparent about their use. The draft also sets out requirements for companies to train their employees in handling AI and to provide clear explanations of decisions based on such systems, even when they don’t directly affect the person making the decision.

Those are significant improvements over the original draft. But the new text also notably omits a ban on using AI for racist profiling or to forecast border movement, which could disproportionately target migrants and asylum seekers, who are often denied the right to seek protection. It also leaves it up to developers to decide if their AI is significant enough to be classified as high risk, which digital rights groups warn will be a significant red flag for enforcing the rules.

Centrist MEPs and Greens introduced more controversial changes to toughen the rules on AI used in marketing, censorship, and child development. These additions reflected a shift in approach from the more conservative bloc’s push for an outright ban on AI used for these applications, which was rejected.

The MEPs and the greens also pushed to extend the scope of the ban on using AI for invasive or discriminatory practices. That would include systems that use facial recognition, fingerprints, iris scans, or voiceprints to categorize people and those that identify their emotions in real-time. But this was defeated along party lines. The final version of the AI Act will now be voted on in plenary next week and is expected to pass with broad support. It will then be sent to the EU Council for review and negotiations with member states, possibly leading to the law taking effect in 2026. It will apply to any system that can be trained by an AI developer and will cover both commercial and open-source software. The new rules will be voluntary for most businesses but enforceable by regulators with fines up to 40 million euros ($43 million) or 7% of a company’s annual global revenue, which for the tech giants would mean billions.