Claude’s Text Watermark Explained: Which Models, How It Works and How to Check (Nov 2026)

If you use Claude, the text it writes for you now carries an invisible watermark. You cannot see it, it adds no characters, and nothing in your copy-paste changes. Anthropic can still detect it with a secret key.

The news this week: on 3 November 2026 Anthropic switched the watermark on for three older models, and four more follow on 16 November. After that, almost every Claude model you can call today marks its text.

This guide explains which models are covered, how the watermark works, who can detect it, how to check files with Anthropic’s free checker, what it means for students, writers and developers, and how it compares with ChatGPT and Gemini.

What Changed on 3 November (and What Comes on 16 November)

Anthropic first announced text watermarking on 14 August 2026. At that point it covered only models launched on or after 2 August 2026, the date the EU’s transparency rules took effect.

Older models sat in a transition period. Anthropic emailed API customers on 31 October and 4 November with the dates for those, and The Decoder reported the details. Here is the timeline:

Note: the 16 November list comes from customer emails reported by The Decoder. Anthropic’s public model table still showed those four models as “No” on 6 November. I will update this post once the table changes.

Which Claude Models Watermark Text?

This table follows Anthropic’s Help Center article “How Claude marks AI-generated content” as of 6 November 2026, with the 16 November changes added.

* Reported date and caveat from Anthropic’s 4 November email: the four 16 November models are expected to watermark only when sampling settings stay at their defaults.

Models such as Opus 3 that are no longer in the table are not covered, according to emergent.sh’s report.

How Claude’s Watermark Works, in Plain English

When Claude writes, it picks each next word from a list of good candidates. Often two or three words are almost equally likely, and the model normally picks between them at random.

The watermark replaces that coin flip. A secret key plus the few words just before decide which candidate wins. To a reader, the text looks normal. To someone holding the key, the pattern of choices is far too regular to be chance.

The technical basis

• Origin: Anthropic says its method is a version of Google DeepMind’s SynthID-Text (published in Nature in 2024), which builds on a 2022 idea by Scott Aaronson.
• Sampling, not training: the mark is applied when tokens are sampled, so the model itself was not retrained (Sebastian Raschka’s explainer).
• Tournament sampling: SynthID-Text pits candidate tokens against each other using several hidden scoring functions. Raschka notes he is not sure Anthropic uses exactly this variant.
• Detection: the detector re-scores the text with the key. It does not need to re-run Claude, and it gives a probability, not a yes/no verdict.

What it does not do

• No hidden characters: no zero-width spaces or odd Unicode. Copy-paste does not strip or add anything.
• No extra tokens: your bill, speed and output length stay the same.
• No personal data: Anthropic says the mark carries no information about you, your organisation or your chat.

On quality, Anthropic says internal tests and a side-by-side human rating study found no difference. It also points to Google’s live SynthID test, which found “no statistically significant differences” in thumbs-up and thumbs-down ratings. These are vendor-reported results.

Where the Watermark Is Weak

Anthropic is unusually open about the limits. Here is what its own post and independent explainers (Codemancers, Raschka) list:

Warning: a detected mark does not prove Claude wrote a text. Anthropic says the detector cannot tell “Claude wrote this” from “Claude heavily edited this”, and a missing mark never proves a human wrote it.

How to Check if Something Came From Claude

Text: you cannot check it yourself

Text detection runs only through Anthropic’s Detection API, which is in private preview. Eligible groups listed by Anthropic:

• Regulators and law enforcement
• Media organisations and fact-checkers
• Researchers and educational organisations
• EU civil society groups
• Enterprises checking their own outputs for compliance

To apply, use the request form linked from Anthropic’s Help Center article on marking AI content. Anthropic says access will widen over time but has not given dates or prices.

Files: use the free Claude Content Checker

Images, video and audio that Claude creates get a signed C2PA Content Credential in their metadata. Nothing in the picture or sound is changed. You can check a file like this:

1. Open the checker. Go to claude.com/check-content (Anthropic’s “Check if a file was made with Claude” page).
2. Drop in the file. Up to 100 MB. It runs in your browser and Anthropic says the file never leaves your device.
3. Read the result. If a Claude credential is found, the file was probably made or processed with Claude. The result says nothing about who made it.
4. Treat “nothing found” carefully. Screenshots, re-saving, format conversion and editing can all strip the credential.

Note that the checker can read these formats, but Anthropic’s Help Center names PNG and JPEG (and its news post SVG) as the files Claude itself credentials today.

Tip: online “Claude watermark detectors” that claim to check text are not using Anthropic’s key. At best they are general AI detectors, which have real false-positive problems. See my AI detectors guide before trusting any score.

What It Means for You

• Everyday chat users: nothing changes in how you use Claude. The watermark has no switch in Settings and no opt-out.
• Students: long Claude-written essays are now traceable by approved organisations, which can include schools. A positive result is evidence, not proof, so keep drafts and notes that show your own work.
• Writers and editors: using Claude only to proofread leaves little signal, but heavy rewriting by Claude can mark the text even when the ideas are yours.
• Businesses in the EU: if you publish AI-generated text to the public, Article 50 can create your own labelling duties. Anthropic tells deployers to assess their obligations and promises more technical guidance.
• Privacy: the mark identifies “Claude”, not your account. Anthropic says it contains no user or organisation data.

Not sure which Claude plan or model you are on? My Claude Free vs Pro vs Max vs Ultra comparison lists the models per plan.

Developer Guide: Sampling Settings and Compliance

For most API users, there is nothing to change. Anthropic’s docs say you do not need to edit requests or response handling.

The exception is if you rely on the watermark, for example for your own EU compliance, and you call Sonnet 4.6 (or, from 16 November, Opus 4.6, Opus 4.5, Sonnet 4.5 or Haiku 4.5). Then follow these steps:

1. Remove sampling parameters. Leave temperature, top_p and top_k out of the request entirely.
2. Check your wrapper. Some frameworks and client libraries set temperature for you. Look at their defaults and unset it.
3. Prefer newer models. Anthropic’s emails mention the sampling caveat only for the older models, so Opus 5.5, Sonnet 5.5 and Haiku 5.5 should not need this. That is an inference from the emails, not a documented guarantee.
4. Apply for detection access if you must verify your own outputs, using the form in the Help Center.

Anthropic’s own example, trimmed:

import anthropic

client = anthropic.Anthropic()
response = client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Summarize the water cycle."}],
    # no temperature, top_p or top_k if you rely on the watermark
)
print(response.content[0].text)

The watermark adds no tokens, so it does not change costs. For cutting spend, see my Claude API cost tracking guide and prompt caching guide.

Claude vs ChatGPT vs Gemini Watermarks

The big difference is scope. Anthropic watermarks everywhere because it says it has no durable way to limit it by region. OpenAI chose EU-only in ChatGPT and opt-in on its API.

For a wider look at the four chatbots, see ChatGPT vs Claude vs Gemini vs Grok.

Why Anthropic Is Doing This: the EU AI Act

• The law: Article 50(2) of the EU AI Act requires providers of generative AI to mark output in a machine-readable, detectable way. The duties applied from 2 August 2026.
• The code: in July 2026 Anthropic and about 190 other signatories signed the EU Code of Practice on Transparency of AI-Generated Content.
• Transition: models placed on the market before 2 August get a grace period that ends in December 2026. That is why the older models are being switched on now.
• Two methods: invisible watermarks for text, C2PA Content Credentials for files.

Criticism and Open Questions

Not everyone is happy. John Gruber of Daring Fireball called it a “perversion of writing”, arguing that synonyms are never identical and any nudge away from the best word is a cost to the user.

His other point: only the provider holds the key, so users cannot check their own text, and people who merely quote or proofread with Claude could be flagged.

Questions Anthropic has not answered publicly yet:

• Accuracy: no false-positive rate or minimum text length has been published.
• Access: no date or price for wider Detection API access.
• Schools: it is unclear how many educational bodies will qualify, and how results should be used in misconduct cases.
• 16 November wave: the public model table had not been updated for it at the time of writing.

Troubleshooting

• “My Sonnet 4.6 output was not detected as watermarked”: check whether your request or framework set temperature, top_p or top_k. Remove all three.
• “Detection results on Bedrock look inconsistent”: Anthropic expected the Bedrock rollout for Mythos 5, Opus 4.7 and Sonnet 4.6 to finish by 10 November. Results before that may vary.
• “The Content Checker found nothing on a Claude image”: the credential lives in metadata. Screenshots, re-saving in an editor or converting formats can remove it. Test the original download.
• “Claude’s answer looks different since the watermark”: Anthropic reports no measurable quality change, and model updates also shift style. If you need consistent wording, use a system prompt with style rules.
• “A website says my text has a Claude watermark”: only Anthropic’s key can detect it. Third-party tools are guessing. Ask which detector was used.
• “I want to turn it off”: there is no setting or API flag to disable it on supported models.

My Take

Technically, this is about as gentle as a watermark can be: no hidden characters, no cost, no personal data. For most people using Claude for work, it changes nothing day to day.

The weak part is trust. Only approved organisations can detect it, there is no published error rate, and a positive result can hit someone who just asked Claude to tidy their writing. That is risky in schools.

My advice: keep using Claude as normal, keep your drafts, and if you build on the API for the EU, drop custom sampling on the older models. I will update this page after the 16 November wave and when detection opens up.

FAQ

Does Claude watermark its text?

Yes. Supported Claude models add an invisible statistical watermark to generated text in the apps, API, Claude Code, Cowork and Claude Tag, worldwide.

Can I turn off the Claude watermark?

No. There is no setting or API flag. On Sonnet 4.6 and the older models joining on 16 November, non-default sampling settings can leave text unmarked, but that is a limitation, not an opt-out.

Can I check if text was written by Claude?

Not yourself. Text detection is limited to approved organisations through Anthropic’s private-preview Detection API. You can check images, video and audio with the free Claude Content Checker.

Does the watermark include my name or account?

No. Anthropic says it carries no information about the user, organisation or conversation. It only signals that Claude was involved.

Does the watermark survive copy and paste or editing?

It survives copy-paste because it is in the words themselves. Light edits usually leave it, while a full rewrite removes it.

Is code from Claude Code watermarked?

Barely. Code leaves few free word choices, so the signal is very weak. Comments and long prose explanations can carry more.

Does ChatGPT have a watermark too?

Yes, but only in the EU for ChatGPT and Codex, using OpenAI’s textGrain. On the OpenAI API it is opt-in for select models.

Will the watermark make Claude’s answers worse?

Anthropic reports no measurable difference in quality, creativity or readability in its tests. Critics argue that any nudge in word choice has a small cost.

Sources

• Anthropic: “How Claude’s text watermarking works” (14 Aug 2026, updated 1 Sep)
• Claude Platform docs: Text watermark page and release notes (3 Nov 2026 entry)
• Claude Help Center: “How Claude marks AI-generated content” (model table, detection form)
• Anthropic: “Check if a file was made with Claude” (Claude Content Checker)
• The Decoder: “Anthropic’s text watermark is easy to bypass on several older Claude models” (Nov 2026)
• TechCrunch: “OpenAI will start watermarking ChatGPT’s text in the EU” (5 Oct 2026)
• Sebastian Raschka, Ahead of AI: “How Claude Watermarks AI-Generated Text” (22 Aug 2026)
• Codemancers: “How Claude’s Text Watermark Works (And Where It Breaks Down)” (25 Aug 2026)
• Elephas: Claude watermark vs Gemini SynthID (Sep 2026)
• Daring Fireball: John Gruber on Claude’s watermark (Aug 2026)
• AI Weekly: Claude Content Checker launch (2 Sep 2026)