OpenAI will start invisibly watermarking ChatGPT text in the EU

The plan has three parts, according to Unite.AI's summary of OpenAI's announcement:

• ChatGPT and Codex in the EU: OpenAI will introduce text watermarking for eligible users across all plans in the European Union, in the coming weeks.
• API, worldwide: from October 5, API customers anywhere can opt in to text watermarking for select models. The setting is off by default.
• The detector: OpenAI opened applications for access to its text watermark detector, but only for approved researchers and expert organizations, granted case by case.

OpenAI is also working with cloud partners to offer watermarking for its models accessed through their services in the coming weeks.

Not a global default: OpenAI says it isn't turning on text watermarking everywhere at launch. A regional rollout lets it learn from real-world use and feedback first.

The EU AI Act. Its Article 50 requires providers of generative AI to make generated text identifiable in a machine-readable way.

Those transparency rules have applied since August 2, 2026, Unite.AI reports. Existing systems got a transition window until December 2, 2026, according to Crypto Briefing.

OpenAI signed the EU's Code of Practice on Transparency of AI-Generated Content in June 2026, Crypto Briefing reports. The code was drawn up by independent experts in a process run by the EU's AI Office.

OpenAI describes text watermarking and detection as early technologies with significant limitations. Its phased approach reflects both the legal requirements and those limits, the company says.

OpenAI's technology is called textGrain. It embeds an invisible statistical signal in the words a model picks. The detector then checks a passage for that signal, Unite.AI reports.

The short version: when the model chooses between words that fit equally well, a secret key nudges which ones it picks. Over enough text, that pattern becomes detectable, but only to someone with the key.

A technical report dated October 5 lists authors from the University of Pennsylvania, Yale University and OpenAI. The detector needs only the generated text and the secret key.

OpenAI plans to release the technology as open source, and says the report will be updated with more details in the coming weeks.

Under good conditions, pretty well. Under real-world conditions, less so. OpenAI itself cautions that strong lab performance doesn't guarantee reliable detection in everyday use.

Detection rates at a 1% false positive rate, for topics like psychology:

• About 80% of 200-token passages.
• About 95% of 400-token passages.

Rates are substantially lower for subjects like mathematics, where word choice is less flexible.

Editing hurts: in tests on 400-token passages, replacing 10% of words with synonyms cut detection from about 92% to 66%. Replacing 25% of words cut it to 17%.

Quality impact: OpenAI says it saw no meaningful difference in performance with and without watermarking on the benchmarks it uses for Astra, its latest frontier model. On GPQA Diamond, for example, Astra scored 94.44% without watermarking and 93.94% with it.

OpenAI says textGrain matched or beat other approaches it tested, including Google's SynthID for text.

Less than people might hope. OpenAI is unusually clear about this.

A text watermark does not:

• Measure how much a human contributed.
• Establish ownership or responsibility.
• Identify the user.
• Verify whether the text is accurate.

And no watermark doesn't mean human-written. Text might be too short, edited or translated for reliable detection, come from an unsupported model, predate watermarking, or come from another company's tools, OpenAI says.

The detector will report only whether it finds an OpenAI watermark, without identifying the user or revealing their prompts or conversations. Because of the risk of missed watermarks and false positives, OpenAI isn't making it public at launch.

Article 50 of the AI Act covers marking and detection of AI-generated content, plus labelling of deepfakes and certain AI-generated publications, Unite.AI reports.

The code is voluntary, the law isn't: signing the Code of Practice is optional, but Article 50's transparency requirements are legal obligations. The European Commission and the AI Board have confirmed the code is an adequate voluntary tool to show compliance. About 190 companies and organizations had signed it by the end of July 2026, according to the Commission.

The final code was published on June 10, 2026, Unite.AI reports.

Editing help is treated differently: Article 50 includes an exception where AI performs an assistive editing function without substantially changing the original material or its meaning, City AM reported.

Who's responsible: a provider can supply watermarking switches, but companies deploying the models still have to decide how to use them in their own products, Crypto Briefing notes. Many details are still being clarified, including how regulators will judge opt-in tools versus default watermarks.

OpenAI says it will keep improving detection, study how watermarks hold up through editing and translation, and explore ways to tell AI assistance apart from AI authorship. It plans to widen access to the detector once it believes results can be interpreted responsibly, according to Unite.AI.

OpenAI's internal prototype was reported to be about 99.9% effective back in 2024, Crypto Briefing reports. But the company held back a broad release.

Two worries, according to Crypto Briefing:

• Robustness: text is easy to paraphrase or run through another tool, and each edit weakens the signal.
• Fairness: non-native English speakers who use AI to polish their writing could be unfairly flagged.

Others moved faster, Crypto Briefing notes.

• Google has offered SynthID-Text in Gemini since 2024.
• Anthropic introduced invisible text watermarks in new Claude models starting August 2, 2026, the day Article 50 took effect. City AM reported that Anthropic applies the marking worldwide, not just in the EU.
• xAI was the only major model maker that didn't sign the EU code, City AM reported, though Grok still has to follow applicable AI Act rules.

Images and audio from OpenAI already carry provenance signals, including C2PA metadata and SynthID watermarks, Crypto Briefing reports. OpenAI says its verification tools for audio and images, including its openai.com/verify web tool, stay publicly available.

• In the EU: your ChatGPT and Codex text will soon carry an invisible watermark. It won't change how the text looks.
• Developers: you can switch on text watermarking in the API now, which may help with your own EU obligations before the December deadline.
• Teachers and employers: don't expect a public AI-text checker from OpenAI yet, and treat any detector result with caution.