New EU Guidelines
Rules for AI Models
with Systemic Risk

The European Union, under the leadership of the newly created European AI Office, has published official guidelines for providers of general-purpose AI models (GPAI). The aim is to help artificial intelligence (AI) developers comply with the far-reaching AI law and provide clarity, especially for companies developing AI models with 'systemic risks'. Guidelines for providers of general-purpose AI models

The AI law, which already entered into force last year, categorizes AI systems into four risk levels: unacceptable risk, high risk, limited risk, and minimal risk. AI applications classified as unacceptable, such as social scoring or certain forms of remote biometric identification, are prohibited in the EU. The new guidelines also specify the conditions under which open-source models may be exempt from certain obligations to promote innovation and transparency.

AI models with systemic risks are defined as those trained with a computing power of more than 10^25 FLOPs (Floating Point Operations per Second). This includes well-known models such as GPT-4 from OpenAI, Gemini from Google, the newer Claude models from Anthropic, and Grok from xAI. These models are considered potentially far-reaching and with possible negative impacts on society.

Developers of AI models with systemic risks must fulfill a number of requirements. Non-compliance can lead to significant penalties, ranging from 7.5 million euros to 35 million euros or 1.5% to 7% of a company's global annual turnover.

The deadlines are staggered: The rules for new AI models with systemic risks apply from August 2, 2025. From August 2, 2026, penalties can be imposed for non-compliance. Providers of models that were already on the market before August 2025 must fully comply with the regulations by August 2, 2027.

Reactions to the AI law and the new guidelines are mixed. Critics, including Meta, fear legal uncertainties and hindrance of innovation. Leading European companies had even demanded a suspension of regulation in the past.

On the other hand, there is also agreement. The AI Office promotes the 'General-Purpose AI Code of Practice', a code of conduct that serves as a practical tool for voluntary compliance with the rules. Companies like Mistral and OpenAI have signaled their willingness to participate. Supporters of the law see it as an important step to protect the safety and rights of consumers while creating a framework for trustworthy AI applications. 'General-Purpose AI Code of Practice'