AI Image Labelling Rules: The Laws, Treaties and Agreements Behind Every AI Tag

Every AI label you have ever seen — the grey line under a post, the badge in the corner, the tag inside the file — exists because of paperwork. Not product design. Not branding. Rules.

That distinction matters for one practical reason: product features come and go, but compliance requirements only accumulate. Understanding what is driving the marking tells you which way this is heading, and it is not toward less.

This is a spoke of our AI image watermarks pillar guide.

The four layers of obligation

There is no single global treaty on image labelling. What there is instead is this stack — and in practice the stack behaves like one, because a provider serving a global market implements the strictest requirement everywhere.

EU AI Act Article 50: mark it, and say so

Article 50 carries the transparency obligations, and two parts concern images directly.

Marking synthetic output. Providers of AI systems that generate synthetic image, audio, video or text are expected to ensure the output is marked in a machine-readable format and detectable as artificially generated or manipulated. The marking is expected to be effective, interoperable, robust and reliable as far as technically feasible.

Read that phrasing carefully, because it explains the entire watermarking landscape:

• *Machine-readable* → metadata and invisible watermarks, not just a visible logo.
• *Interoperable* → a shared standard, which is why C2PA won.
• *Robust* → survives editing, which is why invisible watermarks exist at all.

Disclosing deepfakes. Those deploying systems that generate or manipulate content constituting a deepfake are expected to disclose that the content has been artificially generated or manipulated. This is the duty most likely to reach an ordinary publisher rather than a model provider.

California SB 942: provenance plus detection

California's AI Transparency Act approaches the same problem from the provider side, pushing large generative AI providers toward attaching provenance information to generated content and making detection tooling available so that content can be verified.

The detection half is the interesting bit. It answers a fair objection to watermarking: marking is useless if nobody can check it. It also explains the asymmetry users run into — providers hold the verification capability, so you generally cannot independently confirm whether an invisible watermark is present in your own file. Our SynthID guide covers that asymmetry honestly.

C2PA: the standard everything converges on

The law says "machine-readable and interoperable" without naming a format. Industry answered with C2PA: a cryptographically signed manifest travelling inside the file, recording how it was created and what was done to it since.

That is why the same provenance data turns up across completely different products — it is one standard, adopted widely. And it is why the data is fragile in one specific way: a signed manifest is part of the file container, so decoding an image to raw pixels and re-encoding it leaves the manifest behind entirely. Full detail in C2PA Content Credentials explained.

Voluntary commitments came first

Before any of this was law, major AI providers publicly committed to developing mechanisms enabling users to tell whether content is AI-generated, including provenance and watermarking work. Those commitments are why marking appeared in products well ahead of legal deadlines — and why providers who never faced a binding obligation still implemented it.

What this means for you, practically

Where the tooling fits

Knowing the rules is only useful if you can see what your files actually declare. Inspect before you publish: the report lists every provenance signal present, what removal cleared, and an honest estimate of how the pixels themselves are likely to read.

Read next: how platforms turn these rules into a label on your post →