The EU AI Act: How the world's first AI law is changing the game for developers

On February 1, 2025, the European Union launched the AI Act, the world's first comprehensive law regulating artificial intelligence. The document introduces a classification system for AI systems based on risk levels and establishes requirements for developers. For companies working with neural networks, this means new rules of the game—even if you are not based in the EU, if you have users there, the law may apply to you.
Four risk levels: from minimal to unacceptable
The AI Act divides all AI systems into four categories based on their potential threat to human rights and safety. The higher the risk, the stricter the requirements.
Unacceptable risk — systems that are completely prohibited. These include:
- manipulating human behavior using subliminal techniques
- social scoring by governments or employers
- real-time biometric identification in public spaces (with rare exceptions for law enforcement)
- emotion recognition in the workplace or educational institutions
High risk — systems that could impact people's safety, rights, or life opportunities. This includes AI in medicine, transport, law enforcement, education, recruitment, and credit scoring. Mandatory requirements have been introduced for such systems: risk assessment, data transparency, activity logging, and human oversight.
Limited risk — chatbots, deepfakes, and content generators. The main requirement here is disclosure. The user must know they are interacting with an AI, and generated content must be labeled. For example, if you create an image using a neural network and publish it, you must clearly indicate that it is the result of an algorithm.
Minimal risk — everything else. Spam filters, recommendation systems, and gaming algorithms. There are no special requirements here, but general principles of transparency and data protection remain in effect.
What this means for model developers
If you are creating or deploying a high-risk AI system, you will have to document every stage: what data was used for training, how the model makes decisions, and what measures have been taken to prevent errors. You will need to register the system in a European database and undergo a conformity assessment before entering the market.
For generative models—text, image, or video—the key requirement is content labeling. If your service creates images or texts, users must be able to understand that it is the result of AI. This applies to watermarks, metadata, and explicit labels in the interface.
General-purpose models like GPT or Claude are a separate case. If a model can be used in high-risk systems, its creators are required to publish technical documentation, describe the training process, and implement measures to prevent the generation of harmful content.
Penalties and implementation timelines
Violating the AI Act carries fines of up to 35 million euros or 7 percent of a company's annual turnover, whichever is higher. The highest sanctions apply to the use of prohibited systems. Violating requirements for high-risk systems will be less expensive but still significant—up to 15 million euros.
The law is being implemented in stages. The ban on unacceptable risk systems has been in effect since February 2025. Requirements for general-purpose models will take effect in August 2025. The full set of rules for high-risk systems will begin to apply from August 2026. Companies have time to prepare, but it is not worth delaying—auditing and documenting processes require months of work.
How to prepare: five steps
- Classify your systems. Determine which risk level your product falls under. If in doubt, it is best to consult with a lawyer specializing in tech regulation.
- Document your processes. Collect descriptions of data, algorithms, and validation methods. For high-risk systems, this is a mandatory requirement; for others, it is a best practice.
- Implement content labeling. If your service generates text, images, or video, add clear indications that the content was created by AI.
- Set up logging. High-risk systems must maintain an activity log so that in the event of an incident, the chain of decisions can be reconstructed.
- Train your team. Developers, product managers, and lawyers—everyone must understand what requirements apply to your product and how to comply with them.
Why this is necessary
The AI Act is not an attempt to slow down technological development. It is an attempt to set rules that will protect people from abuse and errors while still leaving room for innovation. The law does not prohibit the development of new models—it requires transparency and accountability.
For users, this means more confidence: you know when you are talking to a bot, you see that an image is generated, and you understand how an algorithm made a decision about your loan. For developers, it is an incentive to build systems that can be trusted and a competitive advantage over those who ignore the standards.
Want to try models that comply with principles of transparency and safety? GPT, Claude, and Gemini are available at GEMERA AI—with intuitive interfaces, request logging, and support in Russian. Test the capabilities of modern neural networks in a convenient format.