• Verdict: the EU AI Act's Article 50 transparency duties are binding soon — weeks out on the framework's 2026 schedule — not noise.
• Two separate duties: providers must machine-mark AI-generated output; deployers must add a visible disclosure to deepfakes and AI-written public-interest text.
• Chatbots sit in the same limited-risk tier: a system that talks to a person must tell that person it is an AI.
• Prioritise the machine-readable mark at generation — it is the load-bearing obligation and the hardest to retrofit.
• Caveat: machine marks are strippable and enforcement is untested at scale; treat provenance as defense-in-depth, not a guarantee.

Verdict: binding soon — ship the machine-readable mark first

Binding soon, not noise. The EU AI Act's transparency duties for AI-generated content are not live everywhere yet, but the European Commission's framework schedules them among the obligations that apply in 2026 — roughly two years after the Act entered into force, and imminent as of this writing. If you generate or publish synthetic media or AI-written text, your first move is the machine-readable mark at generation time. It is the load-bearing obligation and the one you cannot bolt on later.

Two duties, two different owners

The framework splits the work. Providers — the people who build and ship the generative system — must make outputs detectable as artificially generated or manipulated, in a machine-readable form. Deployers — the people who use that system to publish — must add a visible disclosure when the output is a deepfake or is AI-written text put out to inform the public on matters of public interest (s1). Chatbots sit in the same limited-risk tier: if a system talks to a person, it must say it is an AI. Miss the split and you can satisfy one duty while breaching the other.

Which surfaces need a visible label now — and the catch

Prioritise, in order: synthetic audio, image or video passed off as real (deepfakes); AI-generated text on public-interest topics; and any AI system that converses with users. Purely internal drafts and clearly artistic or satirical work sit lower. The honest caveat: the machine-readable mark is the strong part on paper and the weak part in practice. Provenance metadata — including C2PA Content Credentials, the emerging standard for this — is strippable by a screenshot or a re-encode, and enforcement of these duties is untested at scale. Treat marking as defense-in-depth, not a guarantee. Use the time before it binds to wire labeling into the generation pipeline, not the publish button.