Summary
On December 8, 2023, after three days of negotiations, the European Parliament and the Council of the EU reached a provisional political agreement on the Artificial Intelligence Act. The deal kept the Commission's risk-based structure, added bans on practices such as untargeted scraping of facial images and emotion recognition in workplaces and schools, set narrow exceptions for police use of real-time biometric identification, and added obligations for general-purpose AI models, with stricter rules for the most capable ones. It settled the political content of the law that Parliament adopted in March 2024.
What Happened
The final round of negotiations between Parliament, the Council presidency held by Spain and the European Commission ran for three days and ended late on Friday, December 8. The Council described the talks as a "3-day 'marathon'," and BBC News reported 36 hours of talks. The Parliament's lead negotiators, Brando Benifei and Dragos Tudorache, Spain's secretary of state for digitalization and artificial intelligence Carme Artigas and Commissioner Thierry Breton held a joint press conference after the agreement.
The agreed text banned biometric categorization using sensitive characteristics such as political or religious beliefs, sexual orientation or race; untargeted scraping of facial images from the internet or CCTV footage to build recognition databases; emotion recognition in the workplace and in educational institutions; social scoring; and AI systems that manipulate behavior or exploit people's vulnerabilities. Police use of remote biometric identification in public spaces was allowed only under prior judicial authorization and for defined lists of serious crimes, and real-time use was limited to purposes such as searching for victims of abduction or trafficking and identifying suspects of crimes such as terrorism or murder. The Council said the act would not apply to systems used exclusively for military or defense purposes.
Rules for general-purpose AI and foundation models, which the Commission's 2021 proposal had not covered, were new in the agreement. Under the agreement, general-purpose AI systems and the models they are built on had to meet transparency requirements, including technical documentation, compliance with EU copyright law and detailed summaries of the content used for training. High-impact models with systemic risk faced further obligations: model evaluations, assessment and mitigation of systemic risks, adversarial testing, reporting of serious incidents to the Commission, cybersecurity measures and reporting on energy efficiency. High-risk systems required a fundamental rights impact assessment, and fines ranged from 35 million euros or 7% of global turnover to 7.5 million euros or 1.5% of turnover.
The Council said the act would apply two years after its entry into force, with exceptions for some provisions. The agreed text still had to be formally adopted by both Parliament and the Council.
Why It Matters
The provisional agreement was the point at which the world's first comprehensive AI law became politically settled. The later steps, the Parliament's vote in March 2024, entry into force in August 2024, the prohibitions applying from February 2025 and the general-purpose AI obligations from August 2025, implemented a bargain struck in this session.
The deal's most consequential addition was the tier of obligations for general-purpose models with systemic risk, added after ChatGPT and GPT-4 were released. It meant that the frontier developers who had made voluntary commitments to the US government in July 2023 would face binding evaluation, risk-mitigation and incident-reporting duties in Europe. The exceptions for police use of biometric identification remained one of the most disputed parts of the law.
§ How to read the metadata
- Landmark
- Fundamentally alters the trajectory; 2–5 per year.
- Major
- Meaningfully shifts the landscape; 2–4 per month.
- Notable
- Worth documenting; significance can be upgraded later.
- Confidence
- High = primary sources corroborate. Medium = credible secondary only. Low = provisional. Disputed = credible sources disagree.
- Contestation
- Uncontested = no formal challenge. Contested = at least one challenge filed and not resolved by a correction. Corrected = a challenge was upheld and the entry corrected; see its revision history. Superseded = replaced by a later entry. Unresolved = dispute still open.