Where labelling is already the law

The shift from ethics to obligation

Earlier modules argued for disclosure as good practice. In several places it is now a legal duty, and the trend is one-directional.

The rules differ in a way that matters operationally: they attach to audience location as much as to publisher location, so a small studio publishing internationally can be subject to several at once.

The European Union

The AI Act creates transparency obligations for synthetic content. Providers of systems generating synthetic audio, image, video or text must mark output in a machine-readable way so it can be detected as artificially generated. Deployers who generate or manipulate content constituting a deepfake must disclose that it has been artificially generated or manipulated, with carve-outs including artistic and satirical work where the disclosure must not spoil the display of the work, and adjustments for law enforcement.

These transparency provisions apply from August 2026, with the Act's other obligations phased across earlier and later dates. Penalties are set by member states within ranges the Act specifies, and they are substantial.

China

Measures on labelling AI-generated synthetic content took effect on 1 September 2025. They require both an explicit label that users can see or hear and an implicit label embedded in file metadata. Obligations fall on generation service providers, on distribution platforms, and on users who publish. Platforms must check for the implicit label and act where content is unlabelled or falsely labelled.

This is the most comprehensive labelling regime currently in force anywhere, and it is worth knowing about even if you never publish there, because it is the closest thing to a working implementation of the idea.

The United States

No general federal statute at the time of writing. Instead:

• State laws on synthetic media in elections, on non-consensual sexual imagery, and on voice and likeness replicas, differing considerably between states.
• Federal Trade Commission enforcement on deceptive advertising, which reaches fabricated testimonials and endorsements without any AI-specific statute.
• Proposed federal bills on likeness and on election content, introduced and not enacted.

Elsewhere

Several countries have election-specific rules requiring disclosure of synthetic content in political advertising within defined periods. India's information technology rules and advisories have addressed synthetic media and labelling, with proposals to tighten them. Broadcast regulators in several countries have applied existing accuracy rules to synthetic material without new legislation.

What this means practically

Identify the strictest rule that applies to your audience and build to that. For anyone publishing internationally, that currently means a visible disclosure plus preserved provenance metadata, which is a reasonable specification to adopt generally.

Political and election content is the sharpest edge. Rules here are stricter, timed to election periods, and enforced. If you take political work, read the specific rules for that election, in that country, for that period.

Advertising has its own regime. Consumer-protection law reaches deceptive synthetic testimonials and endorsements everywhere, whether or not there is a labelling statute.

Machine-readable marking is becoming a compliance requirement, not just good practice. Keep the credentials and watermarks your tools produce, and choose tools that produce them.

The gap between the law and reality

Two honest observations.

These rules bind compliant actors. The material they are aimed at — fraud, non-consensual imagery, election manipulation — is produced by people who will not comply, using open models that mark nothing. The realistic effect is to establish norms and enable platform enforcement for the vast ordinary volume, not to stop the deliberate misuse.

And enforcement is barely begun. Most of these regimes are new, the authorities are staffing up, and how the thresholds work in practice will be established case by case over years.

A note on why the details in this lesson will date and the structure will not. Specific dates, thresholds and penalty figures change; three features have been stable across every regime introduced so far. Duties attach to the audience rather than the publisher. Machine-readable marking sits alongside human-visible labelling rather than replacing it. And the trigger is whether a reasonable person could take the content for a real record. If you build to those three, a change in the detail is an adjustment rather than a rebuild.

None of this is legal advice, and the applicable rules for your specific work in your specific country are worth checking directly, with a professional where the stakes justify it. What this lesson gives you is the map: there are duties, they differ, they attach to your audience, and they are getting stricter rather than looser. Building an honest disclosure habit now is much cheaper than retrofitting one later.