The EU's Second Draft Code of Practice on Marking AI-Generated Content: Making Labelling Operable

On 5 March 2026 the European Commission published the second draft of its Code of Practice on Marking and Labelling of AI-Generated Content. The code is built around Article 50 of the AI Act and is designed to help providers and deployers of generative AI systems meet their marking and labelling duties in practice.

Voluntary in form, mandatory in target

The code is voluntary, but the obligation it serves is not. The AI Act’s transparency duties apply from 2 August 2026, requiring clear labelling of deepfakes and of AI-generated or AI-manipulated text published on matters of public interest.

The point of a voluntary code is the compliance presumption: companies that adopt it are more readily regarded as having discharged their duties. This is the EU’s familiar pattern of soft law first, hard law as backstop.

Three technical orientations in the draft

First, it distinguishes providers from deployers. Providers focus on machine-readable markings such as watermarks and metadata; deployers focus on user-facing disclosure. The granularity differs, so a company must first locate itself in the chain.

Second, it promotes open standards for content marking, avoiding a situation where proprietary watermarks cannot recognise one another.

Third, it introduces a common EU labelling icon — one visual symbol carrying the message “this is AI-generated”, reducing both user comprehension cost and company design cost.

What it means for Chinese companies going global

Those offering generative features to EU users. Two mechanisms are needed: machine-readable marking and user-visible disclosure. This converges in purpose with China’s synthetic-content labelling duty, but the technical definitions and presentation requirements differ — see our note on the CapCut enforcement case.

Content platforms and distributors. They need to assess their ability to detect and disclose labels on third-party uploads. Transparency responsibility does not rest on model providers alone.

Those operating in both China and the EU. The real work is mapping two sets of duties onto one engineering implementation. For a multi-jurisdiction approach, see our multi-jurisdiction service and our EU AI Act service.

Where it crosses the “simplification” agenda

In the same period, the EU is advancing a digital omnibus simplification package that includes adjustments to the timing of certain AI Act rules. Companies should read “finer labelling duties” and “delayed high-risk rules” on one timeline — see our notes on the EU AI Act transparency rules taking effect and on the EU agreement to ban deepfake pornography and simplify high-risk rules.

The direction of the labelling duty is settled; only the implementation path remains open. For companies going global, the earlier marking and disclosure become part of the product, the less exposed they are at the August deadline.

This article is based on public information concerning the European Commission’s second draft published on 5 March 2026. It is for general reference only and does not constitute legal advice.