AI Watermarking Laws in 2026: What the EU AI Act Actually Requires

In August 2026, invisible watermarking of AI-generated text stopped being a research topic and became a legal obligation in the world's largest single market. If you write with AI tools — or publish content made with them — here's what changed.

What Article 50 says

The EU AI Act's transparency article requires that content generated or manipulated by AI systems be marked as AI-generated in a machine-readable way. For text, in practice, this has pushed major providers toward statistical watermarking: patterns embedded in word choice that specialized detectors can verify. Anthropic's Claude watermarking, which began rolling out in August 2026, is the most visible example; other providers are implementing their own approaches, with a December 2026 deadline for older models to comply.

Two design choices in these rollouts matter to users:

• They apply globally, not just in the EU. Providers are implementing marking once, everywhere, rather than maintaining separate behavior per region.
• There is generally no opt-out. Marking applies across consumer apps, APIs, and enterprise tiers alike.

Who is actually obligated?

The law puts the technical marking duty on AI providers, not on you. But there's a second obligation that does touch writers and publishers: anyone who deploys AI-generated content professionally — publishing it, distributing it as part of a service — is expected not to strip or obscure the marking, and in many contexts to disclose AI involvement.

For an individual editing their own documents, nothing is required of you. For a business publishing AI-assisted content at scale, disclosure duties are real and worth reading carefully.

What a watermark does and doesn't establish

Worth repeating, because enforcement debates will hinge on it:

• A watermark signal indicates AI involvement, not AI authorship. A human-written text that was proofread or translated by an AI can carry a signal.
• Detection currently requires the provider's cooperation — their secret key, their detector. Third-party "AI detectors" you can use today are statistical guessers that work on entirely different (and much noisier) signals.
• Regulators have published little guidance on accuracy thresholds or dispute procedures. Expect the first years of enforcement to be messy.

What this means in practice

If you're a writer using AI assistance: Understand that your tool's output may be marked, and that the mark says "AI touched this," nothing more. If your context requires AI-free text — some academic and journalistic settings do — the reliable path is writing without AI assistance, not trying to launder AI text into invisibility.

If you're a publisher or editor: Detection will become available, but treat positive signals as the start of a conversation, not proof. False positives — especially on human-written, AI-edited text — are a structural feature of the current technology.

If you're building with AI APIs: Assume everything your product outputs is or will be marked, globally, within the year. If your product rewrites AI text (as ours does), be honest with users about what transformation can and can't guarantee.

Where this is heading

The EU rarely stays alone. Expect similar transparency provisions to spread — and expect a long, unresolved debate about where "AI-assisted" ends and "AI-generated" begins. The watermark answers a technical question. The policy question it raises is much harder.

Working with AI-generated text? Our rewriting tool transforms text at the vocabulary, syntax, and structure level — with honest metrics about how much changed.