Where AI-Generated Text Required to Be Watermarked by Law?
Short answer: only in the European Union. Google's Gemini has watermarked text since 2024 with SynthID, and Anthropic added invisible watermarking to Claude's text and code in August 2026 — but California, China, South Korea, and Vietnam all stop short of making it mandatory. Here's the full comparison, and why a watermark alone still isn't something you can rely on to catch AI-written text.
UncovAI is an AI content detection platform that checks text, images, video, and audio for signs of AI generation. Unlike a watermark, UncovAI's text detector doesn't need the model provider's cooperation or a secret key — it works on text with no watermark at all, text from models that don't mark their output, and text where a watermark has already been stripped by editing or translation.
What triggered this
Since August 2, 2026, Article 50 of the EU AI Act has required providers of generative AI systems to make their outputs machine-readable and detectable as AI-generated, wherever technically feasible. Anthropic confirmed in August 2026 that every Claude model released from that date weaves an imperceptible, machine-readable pattern directly into its generated text, applied globally rather than only for EU users. Google's Gemini has done something similar since 2024 through SynthID. OpenAI has signed the same EU Code of Practice and stated its goal is to extend provenance signals to text, with the underlying transition period for existing systems running until December 2, 2026 — but has not yet shipped a text watermark of its own.
That timeline made us curious about a narrower question: which governments, besides the EU, are actually requiring this by law?
The comparison: five jurisdictions, one clear outlier
We compared the regulatory frameworks of the European Union, California, China, South Korea, and Vietnam on five specific points for AI-generated text. Only one requires an actual technical watermark.
| Requirement | EU — AI Act & Code of Practice (2026) | California — SB 942 (2026) | China — Content Labeling Measures (2025) | South Korea — AI Basic Act (2024) | Vietnam — AI Law (2026) |
|---|---|---|---|---|---|
| Text covered by the law? | Yes | No — SB 942 covers image, video, and audio only | Yes | Yes | Yes |
| Visible disclosure / label required? | Yes (with editorial exceptions) | No | Yes — indication or symbol in the text or interface | Yes | If risk of confusion about authenticity (events or people) |
| Machine-readable metadata mandatory for text? | No — a machine-readable watermark remains mandatory instead | No | Yes, if the format allows it | Partially — planned, not yet detailed for text | No |
| Invisible text watermark mandatory? | Yes — above a 200-token threshold | No | No — not required, citing technical difficulty | No | No |
| Detection mechanism planned? | Restricted access — regulators, fact-checkers, labs | No | Yes — metadata check, user declaration, or other generation traces | Encouraged — companies expected to take reasonable measures | No |
Only the European Union imposes an actual technical watermark on AI-generated text. Every other jurisdiction here either stops at a disclosure requirement, leaves detection as encouraged rather than mandatory, or excludes text from the law altogether.
Why the EU ended up here — and California didn't
While the Code of Practice for Article 50 was being drafted, several stakeholders raised doubts about how useful and technically feasible text watermarking actually is. But the AI Act explicitly covers text as one of the content types in scope, which left the Code with little room to leave it out.
California took the opposite path. Early drafts of SB 942 included text, but lawmakers removed it during the legislative process — leaving the final law covering image, video, and audio only. That single difference is the clearest sign of how contested textual watermarking still is, even among regulators trying to solve the same underlying problem.
The core technical problem: text has nowhere to hide a signal
The difficulty is well understood in the field: text carries far less entropy than an image or an audio waveform, leaving little room to embed a robust, hard-to-remove signal. A paraphrase, a rewrite, a translation, or simply passing the text through a different model can substantially degrade or completely erase the watermark. Anthropic's own documentation on Claude's watermark acknowledges this directly — a heavy rewrite or translation can knock it out, and the company still has no scoped way to apply it only to EU traffic, which is why it rolled the watermark out globally instead.
Watermarking vs. detection: two different tools solving two different problems
A watermark and a detector answer different questions. A watermark asks: "did this specific model produce this text, and can I prove it with a key I control?" Detection asks: "does this text show the statistical fingerprint of AI generation, regardless of which tool made it or whether that tool cooperated?" Relying on watermarking alone leaves several gaps that only detection can close:
This is exactly the gap UncovAI's text detection is built to close. It analyzes the writing itself — structure, word choice, statistical patterns typical of generated text — rather than depending on a cooperating provider's key. That means it can flag AI-written content whether or not the source model watermarks anything, and whether or not a watermark that was once there survived editing. For anyone who needs to check text today rather than wait for every model provider's watermark rollout to finish, that's the difference between a tool that works right now and one that works only in the cases regulation has already caught up to.
A fast-moving landscape
Google has run invisible watermarking on Gemini's text output since 2024 through SynthID. Anthropic became the first major lab to deploy production-scale text watermarking across its entire product line at once, in August 2026, applying it globally rather than only where legally required. OpenAI has signed the same Code of Practice and stated its intent to extend provenance signals to text, but has not shipped one as of this writing — leaving text watermarking as, for now, a two-out-of-three-labs story rather than an industry norm.
Textual watermarking is clearly a live scientific, technical, political, and even sociological experiment — user reaction to Claude's watermark announcement made that last part especially visible. Expect this comparison to look different within a year.
Frequently asked questions
Which countries legally require AI-generated text to be watermarked?
Only the EU, under Article 50 of the AI Act and its Code of Practice. China, South Korea, and Vietnam require some form of disclosure, but not a mandatory invisible watermark. California's SB 942 excludes text entirely.
Does California require watermarking of AI-generated text?
No. SB 942 covers image, video, and audio only — text was in early drafts but was removed during the legislative process.
Can AI text watermarks be removed or bypassed?
Yes. Text has little entropy to hide a robust signal in, so paraphrasing, translation, heavy rewriting, or a different model can significantly degrade or remove a watermark.
If watermarking is unreliable, how can AI-generated text be detected?
Tools like UncovAI analyze the writing itself for statistical patterns typical of AI generation, so they can flag AI text even with no watermark, from a model that never watermarks, or where a watermark has already been stripped.
Sources
Regulation (EU) 2024/1689 (AI Act), Article 50 · European Commission, Code of Practice on Transparency of AI-Generated Content, June 2026 · California SB 942, AI Transparency Act · Cyberspace Administration of China, Measures for Labeling AI-Generated Content, 2025 · Republic of Korea, AI Basic Act, 2024 · Anthropic, "How Claude's text watermarking works," August 2026 · Google, SynthID documentation.
Don't wait for regulation to catch up
Most AI text in circulation today carries no watermark at all — and the one region that requires it still lets a rewrite defeat it. UncovAI checks text for signs of AI generation directly, no watermark or provider cooperation required.
Try it free for 30 days →