Neural networks will start leaving hidden markers: new content-labeling rules take effect in the EU

New transparency requirements for artificial intelligence came into force in the European Union on 2 August 2026. Article 50 of the AI Act requires developers of certain AI systems to ensure that content created or modified by artificial intelligence can be detected by machines. This includes digital watermarks and other machine-readable information that may be invisible to users.

Such labeling does not necessarily look like the familiar «Made with AI» notice. For example, Google already uses SynthID technology, which embeds an invisible digital signal in images, video, audio and text, while Microsoft applies watermarks and Content Credentials to some content created or modified by AI.

The aim of the new rules is to make the origin of digital content more transparent and make it harder to pass generated materials off as entirely human-made. At the same time, the claim that ChatGPT, Gemini, Suno and all other neural networks are now required to add hidden watermarks to absolutely everything, including any code, would be too broad. The specific requirements depend on the type of AI system and content, and the rules provide for exceptions.

It is telling that the industry is already moving towards this model: in August 2026, Anthropic announced the introduction of invisible labeling for text and provenance data for images associated with Claude. The company said that the new measures would be applied globally, not only to users in the EU.

In other words, this is not about a single «secret watermark» from the European Union, but about a transition to a system in which the origin of AI-generated content should become technically verifiable. In the future, this could significantly change attitudes towards images, video, audio and texts that are currently almost impossible to distinguish from those created by humans.