Claude now signs its work with an invisible watermark
About 190 organizations signed the EU code that asks AI labs to mark what their models write, and Anthropic is one of them. As of November 16, text from four older Claude models carries Anthropic's watermark too. Only approved groups, like regulators and researchers, can check text for it today.
Claude now leaves a fingerprint on the words it writes, and you will never see it. Neither will your readers, which is sort of the point.
What Anthropic changed on November 16
On Monday, November 16, Anthropic's Claude Platform release notes added four older models to its text watermark: Claude Opus 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5. Text they write now carries the mark on supported platforms.
That brings the older models into the same system. Claude models launched on or after August 2, 2026 have marked their text from day one, and the model table in Anthropic's help center now shows a check mark next to every Claude model listed, from Fable 5.2 down to Haiku 4.5. The mark applies in the Claude apps, Claude Code, Claude Cowork, Claude Tag and the API, worldwide. It also applies when you reach Claude through AWS, Google Cloud or Microsoft Foundry.
One catch for teams building on the API. On the older models, setting temperature, top_p or top_k to anything other than the default may leave the text unmarked, and Anthropic's developer docs warn that your framework or client library might set those values for you.
How a watermark hides inside ordinary words
Anthropic explained the method in August. Claude writes one word at a time, and it often has two equally good options, like gray or overcast for a cold day. Normally a random number breaks the tie. With the watermark, a secret key breaks the tie instead. Over a long piece of text, those tiny choices form a pattern that only someone holding the key can test for.
Anthropic compares it to a board game where the digits of pi decide every move. The game feels exactly the same, but someone who knows pi can look at the moves afterward and spot it. Nothing gets added to the text: no hidden characters, no extra tokens and no extra cost. The mark also carries nothing about who you are or which company you work for.
The technique is a version of SynthID-Text, which Google DeepMind published in Nature in 2024. In that work, human raters who compared marked and unmarked answers saw no difference in quality, and Anthropic says its own testing found no change in readability or creativity.
Who can actually check for it
Not you, and not your competitor. Text detection is in private preview, open to groups Anthropic lists as eligible under EU law: regulators, law enforcement, media, fact-checkers, independent researchers, educational organizations and EU civil society groups. A few companies that must check their own Claude output can get access too. Anthropic says it plans to widen access over time.
Files work differently. When Claude creates a supported file such as a PNG, JPG or SVG, it attaches a Content Credential, a signed note in the file's metadata that follows the open C2PA standard. Anyone can test a file with Anthropic's free Content Checker, which runs in your browser and never uploads the file. It does not check text, and a screenshot or a format change can strip the note off a file.
A found mark also proves less than you might think. Anthropic says it only shows that Claude was likely involved at some point. It cannot tell whether Claude wrote the piece or edited it heavily, and a missing mark does not prove a person wrote it.
What weakens the mark
Short passages are hard to check, because there are too few word choices to carry a signal. Facts leave less room for a watermark too: when there is only one right next word, like the title of a famous book, the watermark has nothing to work with. The same goes for most code. And when Claude only fixes your grammar, nearly every word is still yours.
Heavy editing, paraphrasing or a full rewrite fades the mark or removes it. Translation does the opposite. Every word in a Claude translation is Claude's choice, so it carries the watermark in full.
One tell needs no private detector. Anthropic says outside AI detectors look for phrasing habits, and AI models use the word "quietly" far more than you would expect. If your About page says your team "quietly" does anything, everyone can see that one.
Why now: the EU rule behind it
This one comes from Europe. Under Article 50 of the EU AI Act, providers of generative AI must mark their output so it can be detected, starting August 2, 2026, according to the European Commission. Systems already on the market before that date have until December 2, 2026 for the marking part. That is the window for Anthropic's older models.
Anthropic is not the only lab doing this. About 190 organizations had signed the EU's Code of Practice on AI-generated content by the end of July, the Commission says, with Anthropic, Google, Meta, Microsoft, Mistral and OpenAI among the model providers. Anthropic applies its watermark worldwide because, it says, it does not yet have a reliable way to limit it by region.
The AI Act also asks some publishers to label AI text on matters of public interest, with an exception for text that gets real human review or editorial control. Spell-check alone does not count. If you publish in the EU, that one is a question for your lawyer, not your copywriter.
A watermark points to Claude. Readers judge the copy
The watermark tells a short list of approved checkers that Claude was probably involved. It says nothing about whether your copy is true, useful or worth a customer's time. Readers decide that part, and they decide it by the details.
If you want help building a content process where AI handles the repetitive work and your expertise carries the message, our founders can walk you through it. Talk to them below. See you in the next edition.
Sources
Sources, checked November 18, 2026. Every claim above links to its original source. We summarize and comment; the reporting belongs to the outlets named here.
- Anthropic: Claude Platform release notes
- Claude Help Center: How Claude marks AI-generated content
- Claude Platform Docs: Text watermark
- Anthropic: How Claude's text watermark works
- Anthropic: Check if a file was made with Claude
- European Commission: Transparency obligations under Article 50 of the AI Act
- European Commission: Strong backing for the Code of Practice on Transparency of AI-generated Content
AI consulting for businesses
Want AI working for your business, not just in the headlines?
Our founders look at how you sell today, then show you where AI earns its keep, where it is still a toy, and what to do first. For businesses of any size.
Let's talkMore AI news
- Google Search got a new Gemini brain. Here's what your site needsNovember 16, 2026
- ChatGPT can now pay for things. Make sure your site lets it inNovember 14, 2026
- Claude now joins your Google Meet calls, client calls includedNovember 13, 2026
- Claude gets Sheets and Whiteboards, and wants your campaign trackerNovember 12, 2026

