stobart.ai
@ai4lawyersTikTok
All transcripts

Writing & authorship

Claude text watermarking: what lawyers should know

An introduction to token-level text watermarking, followed by an important update on rollout, detection, and what a watermark can actually show.

Watch on TikTok

Update · September 4, 2026

A watermark is a limited signal, not proof of authorship.

Anthropic says future Claude models will watermark generated text, while coverage for older models is still rolling out. Detection works better on longer, free-form passages, and Anthropic's watermark-detection API is currently in limited-access private preview. A result can suggest only that Claude was likely involved; it cannot identify the user or prove who authored the final document. Google also watermarks text in the Gemini app and web experience, but coverage varies by provider and product. The video's statements that every current Claude sentence is watermarked and that courts or clients can necessarily detect it are too broad.
Anthropic: How Claude's text watermark worksGoogle DeepMind: SynthID overviewOpenAI provenance coverage

What a text watermark can and cannot establish

FactorA watermark can indicateA watermark cannot establish
Model involvementA provider-specific detector may find a signal in qualifying text.Every sentence was generated by AI or carries a detectable signal.
IdentityA result may suggest that a particular model was involved.Who used the model or who authored the final document.
Professional obligationsA result can prompt a fact-specific review.Whether disclosure, verification, or another duty was satisfied.

Example conversion prompt

Assess the limits of a text-watermark result

Explain what a text-watermark result can and cannot show about this passage. Identify the provider, model, product, and detection method that would be required before drawing a conclusion. Do not treat a result as proof of authorship, user identity, or compliance.

Edited transcript

Transcript

Hey lawyers, AI is now watermarking generative text. Here's what that means for you.

1. What Text Watermarking Is

When it comes to generative text, watermarking is the practice of an AI model—let me get this right—consistently choosing low-stakes words over others in a statistical pattern that can be detected by a key holder, but not by a user.

The example they give is where “overcast” and “gray” are equally good, equally applicable. It'll choose one of those, and it'll keep doing that with different words in a pattern that can be detected—not by you, but by what it calls a key holder.

It can't even be detected by the AI that's writing it. It doesn't know what the words are. It just chooses them because the pattern is deeply embedded in the program.

2. The Signal Has Important Limits

What does that mean for you? Well, it applies to any generative text. So if you ask, for example, Claude to write a paragraph, it's going to be watermarked. A sentence—it's going to be watermarked. The more of these you use, the stronger the pattern and the more likely it is to be flagged as AI-written.

Now, does this matter? Maybe so, maybe not. I don't personally care if people know that I use AI, obviously.

3. Confirm the Actual Rule Before Drawing a Conclusion

But if I'm in a jurisdiction that requires disclosure of AI, a court will be able to tell. And if you haven't disclosed it, that can be a problem.

You should be disclosing the use of AI to your clients. If you don't and they run one of these programs, they'll be able to tell, and that might be a problem.

If you want to avoid the watermark altogether, watch my next video. I'm going to tell you how to do it.

Keep following me for more practice tips on how to integrate AI into your legal workflow.

More transcriptsTikTok