General-Purpose AI (GPAI) – Art. 51–56
Definition
GPAI models are AI models that can be used for a wide variety of different tasks (e.g. Large Language Models such as GPT, Claude, Gemini). They are not trained for a specific purpose but can be adapted for various downstream tasks after training.
Distinction: GPAI Model vs. GPAI System
| GPAI Model | GPAI System | |
|---|---|---|
| What | The trained model itself | A system that integrates a GPAI model |
| Who regulates | The model provider (OpenAI, Anthropic, Google, etc.) | The system provider (whoever integrates the model into a product) |
| Obligations | Art. 53 (all), Art. 55 (for systemic risk) | Dependent on the risk level of the system |
Systemic Risk (Art. 51)
A GPAI model has systemic risk if:
- Cumulative computing power for training > 10²⁵ FLOPS, OR
- The EU Commission classifies the model accordingly (due to high capabilities/impact)
Providers of GPAI models with systemic risk have additional obligations (Art. 55).
BAUER GROUP Perspective
The BAUER GROUP is currently not a GPAI model provider. In-house model development in the future is not ruled out. Our current role is:
- GPAI deployer — we use GPAI models (Claude, GPT, etc.) as tools
- Potentially a system provider — if we integrate GPAI models into customer products
The GPAI provider obligations (Art. 53, 55) do not apply to the BAUER GROUP but to Anthropic, OpenAI, etc. Responsibility only shifts if we place a GPAI-based system on the EU market under our own name as a high-risk system.