Sources: EU lawmakers can't agree on how to regulate systems like ChatGPT, with foundation models becoming the main hurdle in talks over the proposed AI Act
EU lawmakers cannot agree on how to regulate systems like ChatGPT, in a threat to landmark legislation aimed at keeping artificial intelligence … Forums: Beehaw Forums: Beehaw : Generative AI a stumbling block in EU legislation talks
Context & Ripple Effects
The AI Act’s earlier legislative arc had built momentum around disclosing copyrighted training material for generative AI and a broader political agreement on restrictions for facial recognition and transparency requirements for generative AI systems.
The new impasse exposes a harder unresolved question: whether the Act’s risk-based framework can assign meaningful responsibilities at the foundation-model layer, rather than chiefly to individual applications.
First-order effects
- AI Act negotiators must resolve rules for foundation models before completing a framework that still awaits formal parliamentary approval.
- Providers and deployers of ChatGPT-like systems lack clarity on which model-level disclosures and compliance duties the eventual Act will impose.
Second-order effects
- Companies preparing EU AI governance programs may have to keep compliance plans flexible, because requirements aimed at training data and generative-model transparency remain unsettled.
- The dispute puts pressure on lawmakers to reconcile application-based risk tiers with obligations that apply upstream across many downstream uses of a single model.
Third-order effects
- If foundation-model obligations become a durable part of the Act, AI regulation will increasingly target shared infrastructure and model providers alongside the businesses that deploy AI applications.
- The outcome will help determine whether operational AI governance becomes centered on documenting specific deployments or on continuous accountability throughout the model supply chain.
The trend: This is one data point in the shift from regulating discrete AI use cases toward governing general-purpose models whose capabilities flow into many products.