Table of contents
The explosion of generative AI with ChatGPT, Midjourney, DALL-E and Stable Diffusion has posed a major challenge for the European legislator. The AI Act responds with a specific framework for General Purpose AI (GPAI) models and enhanced transparency obligations.
Generative AI in the AI Act
The AI Act does not directly classify generative AI into a single risk category. Instead, it introduces the concept of "General Purpose AI" (GPAI) - models capable of performing a variety of tasks without being designed for a specific use.
ChatGPT (OpenAI), Claude (Anthropic), Gemini (Google), Mistral and other foundation models fall into this category. Obligations vary depending on whether the model is considered to present a "systemic risk" or not.
blog_article10_info_text
Classification of GPAI models
The AI Act distinguishes two categories of GPAI models:
blog_article10_tool1_title
blog_article10_tool1_desc
blog_article10_tool2_title
blog_article10_tool2_desc
blog_article10_tool3_title
blog_article10_tool3_desc
GPT-4, Gemini Ultra and large-scale models are likely to be classified as "systemic risk". Smaller or specialised models generally fall under the standard category.
Transparency obligations (Article 50)
Article 50 of the AI Act imposes specific transparency obligations for generative AI systems:
blog_article10_quote
- blog_article10_transparency1
- blog_article10_transparency2
- blog_article10_transparency3
- blog_article10_transparency4
These obligations apply whether you are the model provider (OpenAI, Anthropic) or integrate it into your own products.
Specific rules for GPAI models
GPAI model providers must comply with a set of technical and documentary obligations:
| blog_article10_table_category | blog_article10_table_obligations | blog_article10_table_examples |
|---|---|---|
| blog_article10_row1_category | blog_article10_row1_obligations | blog_article10_row1_examples |
| blog_article10_row2_category | blog_article10_row2_obligations | blog_article10_row2_examples |
| blog_article10_row3_category | blog_article10_row3_obligations | blog_article10_row3_examples |
blog_article10_p4_2
Deepfakes and synthetic content
The AI Act pays particular attention to deepfakes. Anyone who generates or manipulates image, audio or video content with AI must clearly indicate that this content was artificially generated or manipulated.
Creating deepfakes without labelling constitutes a violation. Fines can reach 15 million euros. Platforms must implement detection systems.
- blog_article10_deepfake1
- blog_article10_deepfake2
- blog_article10_deepfake3
Exceptions exist for artistic and satirical uses, but the principle remains: transparency is the rule, opacity the exception.
Generative AI compliance checklist
Use this checklist to verify your compliance:
blog_article10_tip_text
Is your generative AI compliant?
Our audit evaluates the compliance of your generative AI uses and provides you with a detailed action plan.
Start free auditGenerative AI is at the heart of the AI Act's concerns. Whether you develop or use these tools, transparency and documentation are your best allies for staying compliant.
