Guide · Updated 19 November 2026

AI text watermarks

An AI text watermark is an invisible statistical pattern in word choice, detectable only with the lab's key. Claude, ChatGPT, Gemini, and Grok all have one now, and the EU AI Act requires marking from 2026. Every watermark weakens under heavy editing, paraphrasing, or short text.

AI text watermarks by model

Checked 19 November 2026.

ModelStatusScopeDetection
Claude (Anthropic)On worldwideAPI, Claude apps, Claude Code, Cowork, and TagDetection API in private preview for approved groups
ChatGPT (OpenAI)Rolling out to EU users since 5 Oct 2026*ChatGPT and Codex in the EU; API opt-in worldwide*Approved researchers and expert organizations*
Gemini (Google)SynthID Text in the Gemini app since 2024Gemini app and web; API unconfirmedNo public text detector
Grok (xAI)Off by default in the APIGrok API models, enabled per model by adminsx-watermark header; detection API by application

AI text watermarking algorithms

  • Green lists (Kirchenbauer et al., 2023). A hash of the previous word splits the vocabulary into green and red lists, the model favors green words, and a statistical test detects the excess without access to the model.
  • Tournament sampling (SynthID-Text, Nature, 2024). Candidate words compete in a keyed knockout tournament. Google tested it on about 20 million Gemini responses, and Claude and Grok use versions of it.
  • Gumbel sampling (Scott Aaronson, 2022). The model picks the word with the highest sum of its log probability and a keyed random value. OpenAI's textGrain report builds on this approach.
  • Distortion-free watermarks (Kuditipudi et al., 2023). The text distribution stays unchanged, and the mark survived 40 to 50% random token edits.

The entropy problem

A watermark hides in high-entropy word choices. Predictable phrases, facts, and code leave almost no room, so short, formulaic, or factual text carries a weak signal. Kirchenbauer et al. name low-entropy text as a core problem, SynthID performs better with more entropy, and Anthropic says its watermark is sparser on factual passages and code.

AI watermark robustness

  • Length. OpenAI reports about 80% detection at 200 tokens and 95% at 400.*
  • Edits. Replacing a quarter of the words drops OpenAI's detection from 92% to 17%.*
  • Paraphrasing. The SynthID-Text paper says LLM paraphrasing weakens it, though one green-list study still detected the mark after about 800 tokens of strong human paraphrase.
  • Translation. Claude's own translations carry the mark, and text translated afterward can lose it.

EU AI Act requirements

Article 50(2) requires providers of AI systems that generate text to mark it in a machine-readable format, detectable as AI-generated. It applies from 2 August 2026, and systems already on the market have until 2 December 2026. Article 50(4) also requires deployers who publish AI text on matters of public interest to disclose it, unless a human reviewed it with editorial responsibility.

Watermarks and marketing copy

Readers never see a watermark, and detection is restricted to approved organizations. What readers notice is how AI copy reads: the default voice, hedged claims, and uncanny phrasing. See AI slop, and why cleaning it up takes weeks.

AI text watermarks FAQ

Can I check whether text has an AI watermark?

Not as a member of the public yet. OpenAI, Anthropic, and xAI limit text detection to approved organizations, and Google's public SynthID Detector covers images, video, and audio.

Does Claude watermark its text?

Yes, worldwide, on every current model. See Claude's text watermark.

Does ChatGPT watermark its text?

For EU users of ChatGPT and Codex, since October 2026, and the API is opt-in. See ChatGPT's text watermark.

3 agency-grade versions of your site, with conversion copy and art, for a one-time $399.

Generate 3 versions

*OpenAI's announcement blocked our fetch, so these figures come from coverage quoting it (ActuIA, Nemko). OpenAI's textGrain technical report is dated 5 October 2026.

Sources, checked 19 November 2026

  • anthropic.com/news/claude-text-watermark; support.claude.com (How Claude marks AI-generated content)
  • cdn.openai.com (textGrain technical report, 5 Oct 2026); actuia.com and digital.nemko.com (OpenAI EU rollout)
  • docs.x.ai/developers/advanced-api-usage/watermarking
  • deepmind.google (SynthID text in Gemini, May 2024); blog.google (SynthID Detector, Oct 2026)
  • Kirchenbauer et al., A Watermark for Large Language Models, 2023 (arxiv.org/abs/2301.10226)
  • Kirchenbauer et al., On the Reliability of Watermarks for LLMs, 2023 (arxiv.org/abs/2306.04634)
  • Dathathri et al., Scalable watermarking for identifying LLM outputs, Nature 634, 2024
  • Aaronson, My AI Safety Lecture for UT Effective Altruism, 2022 (scottaaronson.blog/?p=6823)
  • Kuditipudi et al., Robust Distortion-free Watermarks for Language Models, 2023 (arxiv.org/abs/2307.15593)
  • EU AI Act Articles 50 and 113 (ai-act-service-desk.ec.europa.eu); Regulation (EU) 2026/1744