Key Takeaways
- Claude now leaves an invisible statistical watermark in the words it chooses, to comply with the EU AI Act.
- Readers cannot see it, it adds no characters or cost, and it carries no information about you or your company.
- Long, freely written text carries a strong mark; short, factual, proofread or code-heavy text carries little.
- Detection is limited to eligible organisations such as regulators, media and researchers for now.
- The practical task for businesses is a clear AI content policy: where AI writes, who reviews, and when to disclose.
Every email, proposal and social post your team drafts with Claude may now carry a mark nobody can see. It will not change a single word you read, but it means AI-written text is no longer untraceable.
Anthropic explains the change in its own watermarking FAQ. This article sets out what the watermark is, what it can and cannot prove, and what a Hong Kong business should decide before it matters.
What is Claude's text watermark?
Claude's text watermark is an invisible pattern built into the words Claude chooses when it writes. Anthropic added it to comply with the EU AI Act, which requires AI providers to mark generated content. Anyone holding Anthropic's key can later estimate how likely it is that Claude wrote a passage.
Anthropic, along with other major AI providers and around 190 signatories in total, signed the EU Code of Practice on Transparency of AI-Generated Content in July 2026. According to Eden AI's tracker, watermarking began rolling out on several newer Claude models in early November 2026, with older models following over the coming months.
Anthropic says it applies the watermark globally because it does not yet have a reliable way to limit it by region, so businesses in Hong Kong are covered too.
How does an invisible text watermark work?
When Claude picks its next word, it often has several equally good options, such as "grey" or "overcast". Normally a random number settles the choice. With watermarking, a secret key settles it instead, leaving a pattern across many choices that only the key can reveal, without changing the meaning.
Anthropic compares it to playing Monopoly with the digits of pi instead of dice: every move still looks random, but anyone who knows pi can check afterwards whether the game followed it.
The method is a version of SynthID-Text, published by Google DeepMind in Nature in 2024.
Does the watermark affect quality, cost or privacy?
According to Anthropic, no. Watermarked and unwatermarked text read the same, nothing is added to the text, there are no hidden characters, no extra tokens and no extra cost. The watermark carries no information about the user, organisation or conversation, so it cannot be traced back to you.
- Quality: Anthropic reports no change in content, creativity or readability in its testing.
- Cost and speed: no extra tokens, so the price and speed stay the same.
- Privacy: the mark only indicates Claude's involvement, not who used Claude.
When can the watermark be detected, and when can it not?
Detection works best on long passages that Claude wrote freely, because each word choice adds evidence. It is weak on short text, factual sentences with one correct answer, light proofreading and code. Light editing usually leaves the mark in place, while a complete rewrite removes it.
| Situation | Watermark strength |
|---|---|
| Long article or email written by Claude | Strong |
| Translation done by Claude | Present, since Claude chose every word |
| Short reply or factual statement | Weak |
| Claude only proofread your text | Very weak or none |
| Code | Mostly absent, except in comments |
| Text completely rewritten by a person | Removed |
Images and files Claude produces carry a separate C2PA content credential in their metadata, which records that Claude was involved.
Who can check whether a text was written by Claude?
Not everyone yet. Anthropic has released a detection API in private preview for organisations entitled under EU law, such as regulators, law enforcement, media, fact-checkers, researchers and educational bodies, plus enterprises that must verify watermarking for their own compliance. Wider access is planned over time.
The watermark can only answer one question: how likely it is that Claude was involved. It cannot prove that a person wrote something, and it cannot detect text from other AI tools, which use their own keys and methods.
What should Hong Kong businesses do about AI watermarks now?
Treat the watermark as a signal that AI-written content is becoming traceable, and set a clear policy before clients or regulators ask. The goal is not to hide AI use, but to make sure every piece of AI-assisted content is reviewed, accurate and disclosed where it should be.
- Map where AI writes: list the external content your team drafts with AI, such as proposals, emails, social posts and reports.
- Require human review: a named person checks facts, tone and promises before anything goes out.
- Decide disclosure rules: agree when you tell clients that AI assisted, for example in reports or customer service replies.
- Check supplier contracts: ask agencies and freelancers to state how they use AI in work they deliver to you.
- Keep records: note which tool and which reviewer were involved for important documents.
Trying to strip the watermark by heavy rewriting misses the point. Clients care whether the content is accurate and honest, not whether a machine touched it.
AI writing is no longer invisible
Claude's watermark changes nothing on the page, but it changes the rules around it: AI involvement can now be checked. Businesses that already review and disclose their AI-assisted content have nothing to worry about. Those that do not should set that policy now, while it is a choice rather than a correction.
Know AI, know you | With UD, AI never feels cold
Put AI to work with clear rules from day one
UD helps Hong Kong businesses deploy AI staff with review steps, permissions and content policies built in, so your team can use AI with confidence.
Frequently Asked Questions
Does the watermark show which company or person used Claude?
No. Anthropic says the watermark carries no identifying information, so it cannot be traced to a person, organisation or chat. It only indicates that Claude was likely involved.
Does Claude's watermark apply to businesses in Hong Kong?
Yes. Anthropic applies it globally at launch because it cannot yet limit watermarking by region, even though the requirement comes from EU law.
Do other AI tools watermark their text too?
Anthropic says other major model developers signed the same EU Code of Practice and will add their own watermarks, each with a different key and possibly a different method.
Will light editing remove the watermark?
Usually not. Anthropic says light editing probably will not remove it, while a complete rewrite where every word is replaced will.
Does a watermark change who owns the content?
No. Anthropic says the watermark only helps test whether Claude was involved; it does not change ownership, authorship or a user's rights under its terms.