EU publishes AI content labelling code

The European Commission has released the finished version of its Code of Practice covering how AI-generated content gets marked and labelled. It is a voluntary code, but it maps directly onto the AI Act’s transparency obligations, which start applying on 2 August 2026.

The substance is practical. Providers and deployers of generative AI systems get two recommended mechanisms for marking content: digitally signed metadata and imperceptible watermarking, with fingerprinting or logging as an optional third. Deepfakes need clear labels, so does AI-written or AI-altered text on public-interest topics, and people must be told when they are talking to a chatbot rather than a human.

Signing is optional, but the alternative is not nothing: a provider or deployer that does not sign will, after August 2, need to demonstrate compliance to market surveillance authorities by other means. The code is the path of least paperwork.

Why this matters

Most small operators using AI for marketing copy, product images, or support bots have been waiting to see what “transparency” concretely requires. This document is the closest thing to an answer so far, from the regulator itself. If you publish AI images for business use, the labelling guidance applies to you directly, and if AI drafts your customer-facing text, keeping your own voice now has a compliance angle too. The same labelling expectation is showing up in products, like Deezer’s new AI music detector, so the direction of travel is consistent.