June 2026 · 6 min read

AI content labeling requirements from August 2026

Two separate duties hide under the phrase "label your AI content," and people mix them up constantly. One is technical and invisible. The other is visible and editorial. Both apply from August 2, 2026.

Duty 1: machine-readable marking — Article 50(2)

If your product generates text, images, audio or video, the output must be marked as artificially generated in a machine-readable format, and it must be technically detectable as such. This duty sits with the provider of the generative system, and it's about what a parser can read, not what a human can see.

What counts as machine-readable? The Act doesn't name one technology, and the draft Code of Practice (second draft, March 2026) says plainly that no single technique is enough on its own. The combination it points toward:

• Provenance metadata embedded in the file (C2PA-style content credentials)
• Watermarking where the medium supports it (images, audio)
• Detectability that survives common transformations like resizing or re-encoding, "as far as technically feasible"

The feasibility qualifier matters. Nobody has solved robust text watermarking, and the regulators know it. What they expect is a serious, documented attempt with current tools, not magic.

Duty 2: visible disclosure of deepfakes — Article 50(4)

Realistic AI images, audio or video of real people, places or events must be visibly disclosed as artificially generated. This catches more products than founders expect. Avatar generators. Voice cloning for podcast ads. A real-estate tool that renders photo-realistic staged interiors. If a normal person could mistake the output for a real photo or recording, plan for a label on the content itself.

The text exception: editorial review

AI-generated text published "to inform the public on matters of public interest" must also be disclosed, with one carve-out: it escapes the duty when a human reviewed it and someone holds editorial responsibility for publication. That carve-out fits newsrooms. It fits an AI-first blog farm much less well, and a thin "a human glanced at it" process is exactly the thing a regulator would test. If your content pipeline is AI-drafted with light review, the honest move is to disclose.

Who carries which duty

Marking (50(2)) is on the provider of the generative system. Deepfake and public-text disclosure (50(4)) is on the deployer, the company that publishes the content. A SaaS that builds a generative feature on top of a foundation model can end up wearing both hats at once: provider of its own system downstream, deployer of the content it publishes. Map this for your own product once, in writing. It takes an hour and it's the first thing anyone auditing you will ask for.

Practical starting point

Inventory every place your product or marketing emits AI-generated content. For each: does it need invisible marking, a visible label, or both? Then keep records of what you shipped and when. The free Article 50 checker walks through this in two minutes, and the DiscloseKit waitlist gets you the labels, the metadata and the audit trail as one script tag before the deadline.

General information, not legal advice. Sources: Regulation (EU) 2024/1689, Art. 50(2) and 50(4); draft Code of Practice on Transparency (December 2025 and March 2026 drafts).