ChatGPT, Claude, Gemini: your EU AI Act obligations if you use a GPAI model

General-purpose AI models (GPAI) are models trained on large amounts of data and capable of performing a wide range of tasks: writing, code, analysis, conversation, image generation. Examples: GPT-4o (OpenAI), Claude 3.5 (Anthropic), Gemini (Google), Llama (Meta), Mistral.

1. What is a GPAI model?

According to Article 51 of the EU AI Act, a GPAI model is an AI model trained with a large amount of data using self-supervision at scale, that displays significant generality and is capable of competently performing a wide range of distinct tasks. The technical criterion: a threshold of 10^23 FLOP of compute used for training distinguishes standard GPAI models from systemic risk models.

2. Obligations of GPAI model providers (OpenAI, Anthropic, Google...)

• Write and make publicly accessible the model's technical documentation
• Provide deployers with the information necessary to use the model in a compliant manner
• Implement a copyright compliance policy (Article 53)
• Publish a summary of content used for training

3. Your obligations as a deployer of a GPAI model

If you use ChatGPT, Claude or Gemini via an API to build a product or service for European users, you are a deployer. Your main obligations:

• Inform your users that they are interacting with AI (Article 50) - in force since August 2026
• Label AI-generated content if you publish it
• If your use of the model creates a high-risk system (recruitment, healthcare...), high-risk obligations apply
• Train your teams on AI literacy (Article 4)

4. The Article 50 transparency obligation - in force since August 2026

This is the most immediate obligation for GPAI model deployers: since 2 August 2026, your users must know they are interacting with AI. In practice:

• Mention 'AI' or 'Powered by AI' in your interface
• Not allow the AI to claim to be human if a user directly asks
• If you publish AI-generated content: label it as such

5. Systemic risk models (GPT-4, Claude, Gemini...)

Models exceeding 10^23 FLOP of compute (including GPT-4, Claude 3 Opus, Gemini Ultra) are subject to additional obligations imposed on their providers: adversarial evaluation, serious incident reporting, cybersecurity measures. For you as a deployer, these additional obligations do not directly change your responsibilities - they are the responsibility of the model provider.