Podcast
Folge 4: General-Purpose AI, die "Allzweck"-KI
Wenn von künstlicher Intelligenz die Rede ist, denken viele zunächst an KI-Modelle wie GPT von OpenAI oder Google Gemini. Solche "General-Purpose AI", kurz GPAI, findet etwa in Chatbots wie ChatGPT oder digitalen Assistenten Anwendung, und für sie gelten besondere Regeln. In dieser Folge der Podcast-Reihe erfahren Sie von Rechtsanwalt Matthias Orthwein und DIHK-Experte Arian Siefert, worauf Sie beim Einsatz von GPAI achten müssen.
A GPAI model is the foundation upon which complete AI systems are built. When a model is combined with a user interface, it becomes a fully functional AI system. The largest and most computationally intensive models are considered to pose systemic risks and are subject to additional safety and testing obligations.
Responsibilities primarily for GPAI providers
The AI Act focuses primarily on model providers rather than operators. This means most regulations affect companies that develop or offer an AI model. Those who merely use an existing model like GPT via an interface (API) to run their own chatbot are considered operators and are mainly required to fulfill transparency obligations. Important: Offering the model under your own brand or significantly altering it through fine-tuning makes you a provider in legal terms—with significantly more responsibilities.
Providers must ensure comprehensive documentation and transparency so that other businesses can understand the model’s capabilities and limitations. This includes a clear description of the model and a summary of the training content. Additionally, providers must ensure that copyright owners can exclude their content from training through machine-readable "opt-out" notices. Additional requirements for systemic models include extensive testing, reporting severe incidents to the EU AI Office, and cybersecurity measures.
What does this mean for SMEs?
Most small and medium-sized enterprises (SMEs) will not develop GPAI themselves but will use ready-made systems like ChatGPT or Copilot. It is crucial to check if a chosen system complies with legal requirements—such as labelling for AI-generated content. If you use AI for critical applications like emotion recognition or automated public information, you must clearly inform your customers that the content is AI-generated.
Top tips
What You Should Consider
Weigh Benefits and Risks: Consider whether you want to work with a GPAI model and develop your own user interface for the AI system, or whether you prefer to use a finished GPAI system. AI models offer more flexibility but require higher compliance efforts.
Ensure Transparency: AI-generated content must be visibly marked as AI only if it involves deepfakes (such as falsified but convincingly real-looking media) or texts about topics of public interest (e.g., PR releases). Otherwise, a non-visible, machine-readable AI mark is sufficient. Check whether your AI system supports this functionality. Chatbots directly interacting with humans as an interface must clearly identify themselves as AI systems.
Respect Copyrights: Make sure your AI model does not use copyright-protected content without permission. Implement machine-readable training opt-out provisions.
Follow Usage Guidelines: Providers of AI systems must offer guidance on their lawful application, which you, as the operator of the AI system, must adhere to.
- Relevant in topic:
- Innovation
- Key areas:
-
- Digitalisierung
Released 09.06.2026
Contact
Arian Siefert
Director Digital Economy
Jonas Wöll
Director Digital Single Market, EU Transport Policy, Regional Economic Policy