Short answer: Anthropic says it began applying an invisible statistical text watermark on 3 November 2026 to four more Claude models (Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6), after newer models already carried it. It says the mark adds no hidden characters, identifies no user or chat, and can only show that Claude was likely involved in a passage, not who wrote it or whether it was heavily edited. Detection is limited to a private-preview API for eligible organisations. A trade report says xAI’s documentation lists a default watermark on seven Grok API models from 7 November. Writers and students cannot check a watermark themselves, so the practical step is to keep dated drafts and notes of how you used any tool. A TimeBind dates the files you hold. It does not show whether a text carries a watermark, or who wrote it.
Last checked 13 November 2026. General information for writers, students and developers, not legal advice. Details come from Anthropic’s own pages and one trade report on xAI’s documentation; we have not read xAI’s documentation ourselves. Vendors change these settings, so check their current pages.
Table of Contents
- What changed on 3 and 7 November
- How the watermark works, in plain terms
- What it cannot show
- Who can check it today
- Why people are searching for this
- Checklist for writers, students and developers
- What a TimeBind does and does not show
- Limits
- Tools that can help
- Frequently Asked Questions
- Sources checked
What changed on 3 and 7 November
Anthropic told Claude Platform customers that from 3 November 2026 it would apply its text watermark to outputs from Claude Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6, the same mark that newer models such as Opus 5.5 already carry. AI Lately, which quoted the email, also reports that Amazon Bedrock customers would wait a few more days. Anthropic’s help page says the marks apply wherever supported Claude models are offered, worldwide, because it has no durable way to limit them to the EU.
The reason given is the EU AI Act. Article 50 started applying on 2 August 2026, and Anthropic says it signed the EU Code of Practice on Transparency of AI-Generated Content in July with around 190 signatories in total. AI Lately reports that xAI’s developer documentation, updated on 31 October, lists a default watermark for seven Grok API models from 7 November and says it is there to comply with the EU AI Act. We have not checked that documentation directly.
How the watermark works, in plain terms
A language model picks its next word from several that would fit. Anthropic says its watermark replaces the random number that settles those low-stakes choices with one derived from a secret key and the few words before. The text still reads normally, but someone with the key can check whether the word choices fit the pattern and put a probability on it. Anthropic says its method is a version of Google DeepMind’s SynthID-Text, published in Nature in 2024.
Anthropic is explicit about the limits. The test answers “how likely is it that Claude wrote part of this?” It does not confirm that text is human-written, cannot spot a different AI’s output, works poorly on short passages, and has less to attach to on factual lines, proofreading and code. A complete rewrite removes it. Light editing probably does not. Translations by Claude carry it, because Claude chooses every word.
What it cannot show
Anthropic’s help page says a detected mark “provides a signal” and “is not fully conclusive.” Claude may not be the original author, since people use it to proofread, translate or summarise. The text may have changed afterwards. And no mark does not mean no AI: heavily edited, paraphrased or very short text may carry nothing detectable, and older models, some platforms and Sonnet 4.6 with non-default sampling settings may not carry it reliably.
Anthropic also says the mark does not change who owns an output or who is responsible for it. That matters for students and employers: a watermark result is a probabilistic signal about one vendor’s tool, not a finding about a person. Our guide to AI detector accuracy covers the separate problem of detectors that guess from style.
Who can check it today
Anthropic says detection is a private-preview API for eligible organisations, such as regulators, law enforcement, media, fact-checkers, independent researchers, educational organisations and EU civil society groups, plus enterprises that must verify marks for their own compliance. Ordinary users cannot test a document themselves. For files such as PNG or JPEG, Claude attaches a C2PA Content Credential instead, and Anthropic offers a free checker for those. See our Content Credentials field guide and the content credentials checker.
Why people are searching for this
Measured on 13 November 2026. A 30 October post saying Claude would start watermarking text on Tuesday had 4,753,588 impressions, 4,424 likes and 575 reposts. A 1 November explainer-style post had 3,103,505 impressions and 14,086 likes, and a joke about retyping text to bypass the mark had 2,254,865 impressions. A 2 November post asking whether fixing spelling and grammar would flag a whole document had 280,178 impressions. X’s trending summaries covered the rollout on 2 November. Counts keep moving.
Checklist for writers, students and developers
- Keep dated drafts. Save each version as you write, with your own notes and outlines. They show how a piece developed, which a vendor watermark cannot.
- Note how you used any tool. Write one line per use: tool, what you asked, what you kept. Do this when you do it, not afterwards.
- Know your rules. Check your school, employer or publisher policy on AI use and whether it asks you to disclose it.
- Do not rely on a watermark either way. Its presence is a signal that a tool touched the words. Its absence shows nothing.
- Developers: record settings. If you build on Claude or Grok, note the model, parameters such as temperature, and any opt-out choices. Anthropic says Sonnet 4.6 may not carry the mark reliably with non-default sampling, and xAI reportedly makes opt-outs your responsibility.
- Check your own duties. If you deploy a model in a product, work out what Article 50 asks of you. Anthropic says to assess this independently.
- Keep the final text. Save the exact version you submitted or published.
- Fingerprint the set. Date drafts, notes and the final file together.
What a TimeBind does and does not show
A TimeBind is a dated, checkable record that an exact file existed no later than a recorded time and still matches. The SHA-256 fingerprint is made in your browser and the file is never uploaded. The record is anchored to Bitcoin (a Merkle root in an OP_RETURN). Your first Pooled TimeBind is free; after that, TimeBinds start from US$5, and an ASAP TimeBind is US$15. See how it works.
For writing it can date your drafts, notes and final version. It does not read or detect watermarks, show whether a passage was written by a person or a model, or show that you wrote a text. It cannot detect AI.
Limits
This page relies on Anthropic’s published pages, one trade report on xAI’s documentation, and public posts. We have not tested any detector. The EU rules are summarised from the vendors’ descriptions; applying Article 50 to your situation is a question for a lawyer in the EU. Other countries differ.
Tools that can help
Use the bulk SHA-256 manifest to fingerprint a folder of files, and the process promise tool to write a plain note about how something was made. For related reading see Substack’s AI detector, a manuscript workflow with dated drafts, the draft streak tool and the glossary entries for watermark and C2PA.
Frequently Asked Questions
-
When did Claude start watermarking text on more models?
Anthropic told customers it would apply its text watermark to Claude Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6 from 3 November 2026. Newer models already carried it.
-
Does the watermark show who wrote a text?
No. Anthropic says it can only indicate that Claude was likely involved, cannot tell whether Claude wrote or heavily edited a passage, and carries no identifying information about users or chats.
-
Can I check whether my text has a watermark?
Not yet, according to Anthropic. Detection is a private-preview API for eligible organisations such as regulators, media and researchers.
-
Does editing remove the watermark?
Anthropic says light editing probably will not and a complete rewrite will. Short passages may be too small to detect either way.
-
Does Grok watermark text too?
A trade report says xAI’s documentation lists a default watermark on seven Grok API models from 7 November 2026. We have not read the documentation ourselves.
-
Can a TimeBind show a text is human-written?
No. It dates exact files and shows they still match. It does not read watermarks or show who wrote something.
Sources checked
- Anthropic, “How Claude’s text watermark works,” 14 August 2026, https://www.anthropic.com/news/claude-text-watermark (read 13 November 2026)
- Claude Help Center, “How Claude marks AI-generated content,” https://support.claude.com/en/articles/16266773-how-claude-marks-ai-generated-content (read 13 November 2026)
- AI Lately, Ryan Elliott Dennis, “Claude and Grok Start Watermarking Text This Week,” 2 November 2026, https://ailately.com/articles/claude-grok-text-watermarks (read 13 November 2026; source for the customer email wording and xAI documentation details)
- X posts by @petesena (30 October), @_junaidkhalid1 (1 November), @redactedrain (1 November), @alex_prompter (2 November), https://x.com/petesena/status/2116243133956796806 (metrics read 13 November 2026)
- X trending summaries on the Claude watermark rollout, updated 2 November 2026 (read 13 November 2026; not independently verified)
Date your drafts and notes
Your first Pooled TimeBind is free, with no credit card required. After that, TimeBinds start from US$5, and an ASAP TimeBind is US$15. Click here to start, or see plans and pricing.
