The EU has required AI companies to warn users and label content created by neural networks.

New requirements are coming into effect for companies in the European Union that develop or use artificial intelligence systems. Starting from August 2, users must receive a clear warning if they are interacting with AI, undergoing biometric analysis, or viewing materials created by neural networks.

One of the main principles of the new rules is quite simple: a person must understand that they are conversing not with a living interlocutor, but with a program. According to the draft recommendations from the European Commission for complying with Article 50 of the EU Artificial Intelligence Act, developers can use banners, separate icons in the interface, text messages, or voice notifications for this purpose.

However, a warning is not required in all cases. If it is clear to the user that they are facing an automated system without additional explanations, special labeling may not be necessary. The European Commission’s documents provide the example of non-playable characters in video games: most people already understand that they are controlled by a program.

Separate requirements concern images, videos, audio recordings, and other materials created by artificial intelligence. Companies must provide technical labeling of such content so that its origin can be determined automatically.

For this, machine-readable metadata, digital watermarks, cryptographic information about the file's origin, and other technical solutions may be used. One well-known example is the SynthID technology, which allows for the addition of a special hidden tag to materials created by neural networks.

The labeling must appear directly as a result of the system's work and cannot be added later by the website owner or user. In other words, the obligation to foresee this possibility lies primarily with the developers of AI services.

However, it will not be necessary to label the output of every tool with elements of artificial intelligence. An exception is provided for services that only slightly alter the original material and do not affect its core meaning.

For example, this may involve correcting grammatical errors, removing noise from audio recordings, or enhancing image resolution. If the tool only improves the technical quality of the file but does not create new content, labeling requirements may not apply.

Stricter rules are established for systems working with biometric data. Users must be warned in advance if a program recognizes them by their face, voice, or other traits, analyzes facial expressions and intonation, or attempts to determine emotions and categorize the individual demographically.

This warning must be highly visible. In a digital service, it could be a pop-up window or a message before starting work. If the system is used in a public place, information can be placed on a physical sign.

Separate exceptions are provided for law enforcement agencies. During an active criminal investigation, the police may, in some cases, be able to use AI surveillance systems without prior notification to individuals. However, this does not exempt such systems from other legislative requirements.

Reporting the use of artificial intelligence will also be necessary in cases where the created or significantly altered material appears realistic. If a neural network depicts an existing person, a real place, object, or event, the viewer must understand that they are looking at synthetic content.

This rule aims to reduce the risk of spreading deepfakes and other materials that could be mistaken for a real photo, video, or audio. Labeling is especially important in situations where artificially created material could affect a person's reputation or mislead the audience.

A more lenient approach is provided for films, music, satire, and other artistic works. Notification of AI use should not interfere with viewing or undermine the author's intent.

For instance, if an actor has been digitally de-aged for a role, this information can be included in the credits or description of the film rather than displayed as a noticeable caption over each frame. A similar principle can apply to parody music and other works where the fictional nature is clear to the audience.

Companies face serious fines for non-compliance with the requirements. The maximum amount can reach 15 million euros or 3% of the company's global annual turnover, depending on which value is greater.

The new rules will affect not only developers of large language models. Companies that integrate AI into their websites, applications, surveillance systems, support services, and content creation tools will also have to comply with the requirements.