EU negotiators emerge from a 38-hour trilogue with a political agreement on the AI Act
France, Germany and Italy had pushed back against general-purpose AI obligations until late in the night. The compromise text contained a chapter that did not exist when the Commission's original proposal was tabled in 2021.
On the afternoon of Friday 8 December 2023, negotiators from the European Parliament, the Council of the European Union and the European Commission entered the final trilogue session on the Artificial Intelligence Act. They emerged thirty-eight hours later, on Saturday evening, with a political agreement. The text would not be finalised at the technical level until February, would not pass Parliament in plenary until March, would not be adopted by Council until May, and would not enter into force until August. But the file was effectively closed in those thirty-eight hours.
The contested chapter was general-purpose AI. It had not existed in the Commission's original 2021 proposal. As reported by Politico Europe and the Financial Times that weekend, France, Germany and Italy had argued through the autumn for a soft-law code-of-conduct approach to GPAI providers, citing the competitive position of the European labs Mistral and Aleph Alpha. The Parliament's negotiators, led by the rapporteurs Brando Benifei and Dragos Tudorache, pushed for binding obligations including a tiered systemic-risk regime.
What the compromise actually was
The agreement adopted a two-tier GPAI architecture. All GPAI providers face baseline transparency, copyright and safety obligations. Models meeting a defined systemic-risk threshold, set at ten to the twenty-fifth FLOPs of training compute, face additional model-evaluation, incident-reporting and cybersecurity obligations. The threshold was deliberately set to capture the largest existing models without sweeping in everything in the wider open-source ecosystem. As Brando Benifei put it in remarks the following week, the negotiation had produced "the regulation that was politically possible, not the one that was technically optimal".
The negotiating positions of the three resistant member states broke late on the second night. The compromise that closed the gap, a code-of-conduct path that would map onto the same obligations until standardised conformity assessments existed, was politically face-saving but legally identical in effect.
The text closed the file. The standards work to operationalise it had not begun.
The trilogue agreement marked the end of a three-year political negotiation that had outlasted three Commission AI commissioners, two German governments, one French presidential election cycle, and the public release of ChatGPT. What it produced was a regulatory architecture written, fundamentally, before the technology it covered existed in the form it now takes.
Discussion