OpenAI's EU text watermark: 92% detection, 66% after 10% of words are swapped
What happened
OpenAI's own tests show its new text watermark, textGrain, is detected about 92% of the time, but only about 66% after 10% of words are replaced with synonyms. The company said Monday it will add the invisible watermark to ChatGPT and Codex output in the European Union over the coming weeks, to meet the EU AI Act's transparency rules. API developers worldwide can switch it on for select models, but it is off by default, and OpenAI is not making it a global default at launch. Detector access will initially go only to approved researchers and expert organizations.
What we know
- OpenAI said textGrain shapes the model's word choices to leave a statistical pattern that a detector can read, and that the pattern stays with the text when it is copied and pasted, per TechCrunch and Engadget.
- In OpenAI's test, replacing 10% of words with synonyms cut detection from about 92% to 66%, per TechCrunch. OpenAI also said short passages, math answers and translated text are harder to detect.
- Engadget reports that short text is detected at roughly 80%, and that OpenAI plans to release the technology as open source so others can modify it.
- The watermark rolls out by default only for eligible ChatGPT and Codex users in the EU; API use is opt-in for select models and off by default, per TechCrunch and Engadget.
- OpenAI said the watermark does not identify the user and that a missing watermark "does not prove human authorship", per TechCrunch. TechCrunch adds that Anthropic announced worldwide watermarking of Claude text about two months earlier.
What we don't know yet
- The two outlets give different compliance timelines: TechCrunch says the AI Act's transparency rules took effect August 2, while Engadget says existing providers have until December 2 to comply under Article 50. Neither cites the legal text.
- Which API models will support the watermark has not been specified.
- Whether the roughly 92% figure holds outside OpenAI's own test conditions is untested; OpenAI itself says strong performance under ideal conditions does not guarantee reliable everyday detection.
- The detector is limited to approved researchers, so no outside party can yet check the false-positive or false-negative rates independently.
- It is unknown whether OpenAI will extend the watermark beyond the EU. TechCrunch notes the company previously held back a text watermark over fears users would move to rivals that did not use one.
Why it matters
The EU rules require machine-readable marking of AI text, and OpenAI's own numbers show that light editing weakens the mark. Publishers, schools and regulators that planned to rely on detection will get a signal that works on unedited output but that OpenAI does not claim can prove human authorship, and for now only approved researchers can run the detector.
Claims
- ConfirmedOpenAI will add an invisible textGrain watermark to ChatGPT and Codex text output in the EU.
- ConfirmedIn OpenAI's test, replacing 10% of words with synonyms cut watermark detection from about 92% to 66%.
- ConfirmedThe watermark is off by default for API developers, who can enable it for select models.
- ConfirmedDetector access is initially limited to approved researchers and expert organizations.
- UnconfirmedDetection of short text succeeds about 80% of the time.
- UnconfirmedOpenAI plans to release the textGrain technology as open source.
- UnconfirmedEU AI Act transparency rules took effect on August 2.
- UnconfirmedExisting AI providers have until December 2 to comply with Article 50 of the EU AI Act.
Related coverage
Why you can trust this story
63%Source map · 2 outlets / 2 articles
- TechCrunch
- Article 1techcrunch.comreport
- Engadget
- Article 1engadget.comreport
How this credibility score is calculated
- Source reliability
- Corroboration
- Primary source
- Atom-verified claims
- No contradiction
- Settled
- Claim attribution
- AI disclosure
Atom-verified claims matched cited source text. Weights are fixed and explainable.
Human accountability
- 2 independent origins / 2 sources
- Drafted by the Newsmesis judgment agent (agent-cli); human approval required before publishing.
Verification ledger · 8
- Claim: OpenAI will add an invisible textGrain watermark to ChatGPT and Codex text output in the EU.Status: ConfirmedVerification trailVerifiedChecked by verification engine
At least one extracted atom matched cited source full text.
- Claim: In OpenAI's test, replacing 10% of words with synonyms cut watermark detection from about 92% to 66%.Status: ConfirmedVerification trailVerifiedChecked by verification engine
At least one extracted atom matched cited source full text.
- Claim: The watermark is off by default for API developers, who can enable it for select models.Status: ConfirmedVerification trailVerifiedChecked by verification engine
At least one extracted atom matched cited source full text.
- Claim: Detector access is initially limited to approved researchers and expert organizations.Status: ConfirmedVerification trailVerifiedChecked by verification engine
At least one extracted atom matched cited source full text.
- Claim: Detection of short text succeeds about 80% of the time.Status: UnconfirmedVerification trailUnverifiedChecked by verification engine
No atom-level source-text verification is available.
- Claim: OpenAI plans to release the textGrain technology as open source.Status: UnconfirmedVerification trailUnverifiedChecked by verification engine
No atom-level source-text verification is available.
- Claim: EU AI Act transparency rules took effect on August 2.Status: UnconfirmedVerification trailUnverifiedChecked by verification engine
No atom-level source-text verification is available.
- Claim: Existing AI providers have until December 2 to comply with Article 50 of the EU AI Act.Status: UnconfirmedVerification trailUnverifiedChecked by verification engine
No atom-level source-text verification is available.
Update log · 0
AI accelerates. Humans approve.