EU AI Act: General-Purpose AI Models
You'll learn how the EU AI Act treats foundation models and large-scale general-purpose AI systems differently from task-specific AI applications.
What Makes General-Purpose AI Different?
The EU AI Act recognizes that general-purpose AI models (GPAIs)—like large language models, multimodal foundation models, and other broadly capable systems—don't fit neatly into traditional risk categories. Unlike a loan approval system designed for one task, a GPAI can be adapted for countless uses, some harmless, some high-risk.
Think of it like selling kitchen knives versus selling drones: a knife has clear, limited uses, but a GPAI is more like a drone that could deliver packages, film weddings, or conduct surveillance—its risk profile depends entirely on downstream application.
Two-Tier Framework
The Act creates two regulatory tiers for GPAIs:
Standard GPAI Models
Must provide technical documentation, transparency about training data (copyright compliance), and energy efficiency information. These are baseline requirements for responsible development.
GPAI Models with Systemic Risk
Models that pose systemic risks—typically those trained with compute exceeding 10²⁵ FLOPs or meeting other capability thresholds—face stricter obligations:
- Model evaluation and adversarial testing
- Tracking and reporting serious incidents
- Ensuring adequate cybersecurity
- Reporting energy consumption
This recognizes that ultra-capable models could enable mass disinformation, cybersecurity threats, or other society-wide harms even before specific downstream deployment.
Provider vs Deployer Responsibilities
Foundation model providers bear these GPAI obligations, while deployers using these models for specific applications must still meet requirements based on their use case's risk level.
Key Takeaway: The EU AI Act introduces special governance for general-purpose AI models, with heightened scrutiny for systemically risky foundation models based on their scale and capabilities, not just their deployment context.