OpenAI plans invisible text watermarks for ChatGPT and Codex in EU

On Oct. 5, 2026, OpenAI said it intends to embed undetectable-by-eye, machine-readable markers in qualifying EU text generated through ChatGPT and Codex. The company said the change would cover eligible users across all plans in the EU, while watermarking will not become a global default at launch.¹²³

A statistical pattern in the words

The system, called textGrain, places a statistical signal in generated text by subtly influencing the model’s word choices, according to OpenAI. TechCrunch reported that a secret key is used to organize possible next-word selections; a detector can then examine the passage with that key for evidence of the pattern. Because the signal is carried in the chosen words, it can remain when text is copied and pasted.¹³

OpenAI said textGrain performed as well as or better than approaches including Google DeepMind’s SynthID for text. The Verge reported that OpenAI also published benchmark results showing comparable performance for marked and unmarked output. OpenAI nevertheless cautioned that the method cannot ensure dependable detection.²³

Detection becomes more effective as passages grow longer, according to OpenAI’s testing. With a targeted false-positive rate of 1%, its detector identified the watermark in about 80% of 200-token psychology passages and about 95% of passages containing 400 tokens. OpenAI also said short passages, mathematical answers and translated material are more difficult to identify.¹³

Editing can weaken the signal substantially. OpenAI found in a test of 400-token texts that detection fell from roughly 92% to 66% after 10% of the wording was changed, and to 17% when 25% was changed. OpenAI has also warned that broader rewriting and translation can make detection less effective.¹³⁵

Restricted access to detection

The detector will not be publicly available at launch because of the risks of missed watermarks and false positives. Approved researchers and expert organizations may apply for access, which will initially be granted case by case. The tool reports whether it finds an OpenAI watermark but does not identify a user or expose prompts or conversations, the company said.¹²³

A positive result has limited meaning. OpenAI says it can show that one of its systems generated or processed at least part of a passage, but it cannot establish how much human editing, judgment or creative work was involved. An absent signal also “does not prove human authorship”. The Verge reported that such marks do not establish accuracy, ownership or the level of human contribution.¹²³

OpenAI will let API customers worldwide opt into marked output for supported models, with the setting disabled by default. It is also working with cloud providers to make marked output available through their services. The company plans to publish textGrain as open-source software, though the dossier provides no release date.¹²³⁵

European rules and wider scrutiny

The rollout follows the Aug 2, 2026, start of Article 50 obligations under the EU AI Act. According to the cited accounts of the law, providers of generative AI must mark produced content and make it possible for systems to identify it as AI-generated. OpenAI is among the companies that committed to the EU code of practice on AI-generated content.¹³⁴

Other providers have pursued similar systems. An arXiv paper reported on Sept 9 that Claude models released after Aug 2 embed SynthID-Text by default without an opt-out. The same paper said Google had used SynthID-Text in Gemini since 2024. TechCrunch reported that Anthropic’s approach applies worldwide, contrasting with OpenAI’s EU-focused default rollout.¹⁴

The arXiv authors argued that important vendor claims and user objections cannot currently be independently checked, partly because deployed detectors are not public. Having evaluated an open-source version of SynthID-Text with two open-weight models, the authors called for certified audits, standardized testing, compatible detectors and publication of watermark settings. They identified the inability to verify claims, rather than watermarking itself, as the central governance problem.⁴