◆ THE LAB · Safety
OpenAI will watermark EU ChatGPT text without a public detector
The News
OpenAI said on October 5, 2026 that it will add an invisible watermark to eligible ChatGPT and Codex text output in the European Union over the coming weeks, in response to the EU AI Act, which OpenAI says requires generative AI providers to make generated text identifiable in a machine-readable way. From the same day, API customers worldwide can opt in to the watermark for select models; it stays off by default in the API. OpenAI opened applications for its text watermark detector; access will initially be limited to approved researchers and expert organisations, granted case by case, and the detector is not public at launch. The watermarking technology is called textGrain.
Chandni Melwani
Founder & Editor
Oct 7, 2026 · 2 MIN READ
Know More
- — Languages: across the 24 official EU languages, OpenAI's test found detection at a 1% false-positive rate ranging from 69.0% in Spanish to 42.2% in Romanian; OpenAI says it then raised the watermark's strength for languages below 60%.
- — Length: at a target false-positive rate of 1%, the detector identified watermarks in about 80% of 200-token passages and about 95% of 400-token passages for content such as psychology. Detection was substantially lower for content such as mathematics.
- — Threshold: the EU code of practice exempts very short text, currently defined as under 200 tokens (about 150 words in English, by OpenAI's estimate), from watermarking. The Commission's separate Article 50 guidelines treat source code as outside the AI Act's marking duty.
- — API switch: Organization settings, then Data controls, then Text provenance, with a per-project override. Customers choose which models carry the watermark.
- — Next: OpenAI says it will add detail to the textGrain technical report in the coming weeks and plans to release the technology as open source. Its research partners include John Thickstun of Cornell, Martin Vechev of ETH Zurich and INSAIT, and researchers at the Kempelen Institute of Intelligent Technologies.
In an editing test on 400-token English passages, OpenAI’s detector found its new invisible watermark for AI-generated text in about 92% of them. Swapping one word in ten for a synonym cut that to 66%, and swapping a quarter of the words cut it to 17%. OpenAI published the results on October 5 with its plan to add the watermark, called textGrain, to eligible ChatGPT and Codex text in the European Union over the coming weeks.
OpenAI writes the mark into the model’s word choices. As the model picks each next word among plausible options, a secret key tilts its choices into a pattern that a detector holding the same key can test for, OpenAI’s help page explains. The mark is expected to stay with any text a person leaves unchanged. Code and very precise answers, such as math solutions, leave the model few alternatives to tilt. OpenAI reports no meaningful performance difference from the watermark on its GPT-6 Astra benchmarks.
API customers who switch the watermark on do not get access to the detector, the help page says. An agency in the EU drafting client copy in ChatGPT could be handing over marked text that neither it nor its clients can test, unless OpenAI approves them.
The watermark does not replace any visible label a business may be required to show, OpenAI says, leaving that question to customers’ legal teams. The European Commission’s code of practice has a separate section for deployers, the AI Act’s term for anyone using an AI system other than for personal, non-professional activity. The section covers labelling deepfakes and AI-generated text published to inform the public on matters of public interest without human review or editorial control. Platforms add rules of their own: Amazon told sellers in July to tag product images and videos showing photorealistic AI-generated people with a metadata keyword, which Amazon uses to show shoppers an indicator where applicable.
Related
Room for Disagreement
OpenAI's own results make its case for holding the detector back. Detection falls on short, edited, translated or very factual text, and detectors can also report a watermark where none is present. OpenAI cites the risk of both errors in keeping the text detector private at launch, and says it will widen access "when we believe results can be interpreted responsibly." Its checks for images and audio are already open to the public at openai.com/verify.
Frequently Asked Questions
Will ChatGPT text written outside the EU carry the watermark?
Not at launch. OpenAI says ChatGPT and Codex watermarking applies in the EU only and is not a global default. API customers anywhere can switch it on for select models from their project or organisation settings, and it stays off unless they do. OpenAI also says it is working with cloud partners to make watermarking available for its models accessed through their services in the coming weeks.
Does the watermark show who wrote the text?
No. OpenAI says the watermark does not associate a person, organisation, account, prompt or conversation with the text, and a detection does not show how much a person contributed, who owns the text or whether it is accurate. A missing watermark does not prove a human wrote the text either: the passage may be too short, edited or translated, may predate watermarking, or may come from another company's tools.
How is this different from the AI detectors already on the market?
Third-party detection tools typically use classifiers that look for patterns such as word choice in finished text, according to OpenAI's help page. textGrain embeds a signal while the text is being generated and checks for it with a secret key, and OpenAI says the EU AI Act specifically requires a signal embedded in generated text. OpenAI's detector looks only for OpenAI's own watermark and is not designed to identify text from other companies' models.
Sources
- OpenAI, “Our approach to EU text provenance rules” (October 5, 2026): the rollout plan, the API opt-in, detector access, and OpenAI's robustness and benchmark results
- OpenAI Help Center, “Provenance signals in OpenAI-generated content”: how textGrain works, detector access for API customers, results by language, the 150-word estimate for 200 tokens and the guidance on visible labels
- OpenAI technical report, “textGrain: Entropy-Calibrated Watermarking for Language Model Text” (October 5, 2026)
- European Commission, “Code of Practice on Transparency of AI-generated Content”: the exemption of very short text from watermarking, and the deployer section on labelling deepfakes and AI-generated text published on matters of public interest
- European Commission, “Guidelines on transparency obligations for providers and deployers of AI systems” (July 20, 2026): paragraph 68, which treats source code as outside the Article 50(2) marking duty
Chandni Melwani
Chandni Melwani is the founder and editor of New in AI, covering AI agents, M&A, and enterprise adoption. She holds a Master's in Management of Artificial Intelligence from Queen's University and brings a practitioner's perspective from her work in Data and AI leadership.
Follow ↗The Lab Briefing
Get the five-beat briefing in your inbox.
The big story, frontier labs, AI at work, e-commerce and the capital crossing borders, each with the part that tells you why it matters, not just that it happened.
Free · No spam · Unsubscribe in one click
More from The Lab
Labs
Anthropic is turning away distillers' paying traffic
Anthropic says early evidence suggests its new defences have cut the data harvested by distillers, who covertly extract Claude's capabilities to replicate them in other models, by around 90 to 99%, and that it has turned away meaningful revenue and trillions of tokens' worth of paying traffic from distillers.
Chandni Melwani · Global · Oct 21, 2026
Models
Meta releases Muse Spark Open under its own licence
Meta's open-weights Muse Spark Open is free of charge under its Muse Model License, but if a company or any of its affiliates offers AI models to third parties via API or a similar interface, and together they gross over $20m in any consecutive 12 months, the company must notify Meta and sign a separate written agreement before offering the model or a derivative as part of that model-as-a-service business. Meta's own table puts it behind GLM 5.3, Kimi K3 and Qwen3.8-2.4T-A95B on the Artificial Analysis Intelligence Index.
Chandni Melwani · Global · Oct 20, 2026
Models
Google's Argon beats Opus 5.5 on LiveBench reasoning, not coding
Gemini 4 Argon costs $2 per million input tokens and $10 per million output through November 30, then $4 and $20, the same as Claude Opus 5.5. On LiveBench 2.0 it beats Opus 5.5 on reasoning but trails it on coding, and it ranks sixth on Vals' Terminal-Bench 4.0.
Chandni Melwani · Global · Oct 18, 2026