Guide · Updated 19 November 2026

ChatGPT text watermarks

Since 5 October 2026, OpenAI has been rolling out an invisible text watermark, textGrain, to ChatGPT and Codex users in the EU, with an opt-in for API customers worldwide. Only approved researchers and expert organizations can run the detector.*

ChatGPT watermark rollout

The watermark covers eligible ChatGPT and Codex users across all plans in the EU. In the API it stays off by default, and organizations can opt in for select models under Data controls, Text provenance.* OpenAI published the change on 5 October 2026 as its approach to the EU AI Act's text provenance rules.

The textGrain method

OpenAI's technical report describes entropy-calibrated watermarking: a keyed Gumbel method, in the line of Scott Aaronson's 2022 scheme, tuned to high-entropy word choices. The pattern is statistical and invisible to readers.

ChatGPT watermark reliability

  • About 80% detection at 200 tokens and 95% at 400, at a 1% false-positive rate.*
  • Replacing 10% of words with synonyms drops detection from 92% to about 66%, and a quarter of words to 17%.*
  • Detection in other EU languages ranges from 42.2% (Romanian) to 69.0% (Spanish).*

EU AI Act requirements

Article 50(2) of the EU AI Act requires providers to mark AI-generated text in a machine-readable format. It applies from 2 August 2026, and systems already on the market have until 2 December 2026.

Marketing copy

Readers never see the watermark, and the public can't run the detector. What readers notice is how AI copy reads. See why AI marketing copy sounds generic.

ChatGPT watermark FAQ

Does ChatGPT watermark text in the US?

Not by default as of November 2026. The rollout covers ChatGPT and Codex users in the EU, and API customers anywhere can opt in.*

Can I check whether text has a ChatGPT watermark?

No. OpenAI gives detector access only to approved researchers and expert organizations.*

Do other AI models watermark text?

Yes. Claude marks text worldwide, Gemini uses SynthID Text, and Grok's API offers it per model. See AI text watermarks compared.

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 scope, settings, and figures come from coverage quoting it (ActuIA, Nemko). The textGrain technical report is dated 5 October 2026.

Sources, checked 19 November 2026

  • cdn.openai.com (textGrain: Entropy-Calibrated Watermarking for Language Model Text, 5 Oct 2026)
  • openai.com/news/rss.xml (Our approach to EU text provenance rules, 5 Oct 2026)
  • actuia.com (OpenAI will watermark ChatGPT in the EU); digital.nemko.com (OpenAI EU rollout)
  • scottaaronson.blog/?p=6823
  • EU AI Act Articles 50 and 113; Regulation (EU) 2026/1744