Negotiations over the European Union’s AI Act had reached a difficult stage, with lawmakers split over how to protect fundamental rights while setting rules for a fast-changing technology. Disputes over prohibited uses, assessments of rights impacts and obligations for foundational models could determine whether negotiators reach an agreement.
Rights safeguards at the center of talks
The negotiations, known as trilogues, bring together the European Parliament, Member State governments represented by the Council, and the European Commission. Brando Benifei, a Parliament co-rapporteur for the AI legislation, said the talks were complicated and difficult, and warned that Parliament would not accept a deal that weakened protections too far.
Among the contested areas were the proposed list of prohibited AI practices, fundamental rights impact assessments, and exemptions for national security. Benifei said Parliament wanted movement from the Council on these points, particularly on law enforcement uses. He argued that negotiators should not conclude an agreement by conceding on protections for citizens.
Sarah Chander of EDRi described civil society recommendations that she said were facing resistance. They included a full ban on remote biometric identification in public, registration of high-risk AI used by law enforcement and immigration authorities, a clear risk classification process, and limits on exporting prohibited systems beyond the EU. The positions on proposed bans covering biometric categorization and emotion recognition were also unresolved.
Foundational models add another fault line
How the Act should regulate generative AI and foundational models became another major point of contention. Benifei said industry lobbying was putting pressure on governments, while insisting that lawmakers should maintain an ambitious approach.
TechCrunch reported that France and Germany had opposed Parliament’s proposed tiered approach to foundational models in a meeting of a Council technical body. It also reported that French startup Mistral and German startup Aleph Alpha were lobbying governments against dedicated measures for generative AI model makers. Corporate Europe Observatory told TechCrunch that France and Germany were among the Member States seeking a regulatory carve-out.
Mistral CEO Arthur Mensch rejected the claim that his company was blocking the talks. He said regulation should target applications rather than foundational models, and said providers could offer downstream users tools for evaluation, monitoring and guardrailing. The debate raised a practical question: how can companies deploying a model check for bias and other harms if they lack access to the underlying model and its training data?
Corporate Europe Observatory’s Bram Vranken argued that major technology companies were publicly calling for rules on dangerous AI while privately pushing for a lighter approach. He said European companies had joined that effort. The competing positions reflected a broader concern in the article: lobbying could shape safeguards before negotiators settle the law’s final provisions.
Impact assessments must have substance
Fundamental rights impact assessments were one area where Benifei expressed more hope for compromise. The proposed assessments would ask developers and deployers to consider in advance how an AI system might affect democratic freedoms and to take steps to prevent or reduce harm. The approach is intended to make rights risks part of decision-making before systems are used.
Lidiya Simova, an adviser to MEP Petar Vitanov, said private-sector opposition had framed assessments as burdensome. She cautioned that an obligation could become meaningless if companies could satisfy it with paperwork alone. In her view, assessments need consequences when organizations fail to meet the requirement.
Simova also pointed to a deeper challenge: the legislation seeks to protect fundamental rights through a product-safety framework. That tension, she suggested, helps explain why Parliament and other negotiators had repeatedly exchanged amendments and drafts. Agreement on individual rules may be difficult when the institutions differ over the underlying approach.
A deal remained uncertain
An EU source close to the Council described the disputed issues as tough points for Member States, with very little flexibility. The source said a conclusive trilogue on December 6 was still possible as Council preparatory discussions continued and Member States considered a revised mandate for the Spanish presidency. Technical teams were also searching for possible areas of agreement.
Benifei warned that if the Council did not move, the whole Act could fail. He said Parliament did not want a law unable to protect fundamental rights, but also recognized that without compromise the EU might deliver no AI Act at all. The negotiations therefore carried two risks: weakening protections in the name of a deal, or failing to reach a deal while AI adoption continued to expand.