The Pulse
OpenAI Starts Adding Invisible Text Watermarks in the EU
OpenAI will add invisible statistical watermarks to eligible ChatGPT and Codex text output in the European Union over the coming weeks. It is also opening opt-in watermarking for select API models worldwide, while limiting access to its tex
AI.info Team · Written with AI ·
“Text watermarking and detection remain early technologies with significant limitations, and views about their benefits and responsible uses are still developing.”
OpenAI
ChatGPT’s rollout is limited to the EU
OpenAI says it will begin adding invisible watermarks to eligible text produced by ChatGPT and Codex users in the European Union over the coming weeks. The rollout will cover eligible users across all plans, but the company is not making watermarking a global default for ChatGPT at launch. OpenAI announced the plan on October 5, 2026, alongside changes for API customers and researchers.
Starting October 5, API customers worldwide can opt in to watermarked text for select models; the feature stays off by default. OpenAI says it is also working with cloud partners to offer watermarking for its models accessed through their services in the coming weeks. Researchers and expert organizations can apply for access to a detector, but the company says it will approve access case by case and will not release the tool publicly at launch.
The move comes after the EU AI Act’s transparency requirements began applying on August 2, 2026. The European Commission says the requirements call for clear labels in key cases, including AI-generated or AI-manipulated text published on matters of public interest. OpenAI describes its approach as a response to the law, while phasing in text watermarking by region and product rather than making it a universal ChatGPT setting.
textGrain changes token choices, not visible text
OpenAI calls its system textGrain. It adds a statistical signal by influencing the model’s word choices; the signal is invisible to a reader, and a detector checks whether a passage contains it. The company’s technical report describes a method that uses a secret key and preceding text to shape token sampling, then tests the generated passage for statistical evidence of that signal.
OpenAI says its evaluations matched or exceeded approaches it tested, including Google DeepMind’s SynthID for text, but warns that results under test conditions do not guarantee reliable detection in everyday use. At a target false-positive rate of 1%, the detector identified the watermark in about 80% of 200-token passages and about 95% of 400-token passages for material such as psychology. Detection fell substantially on mathematics, where the model has less room to vary its word choices.
Editing can weaken the signal, too. In tests on 400-token passages, replacing 10% of words with synonyms lowered detection from about 92% to 66%; replacing 25% lowered it to 17%. OpenAI says it plans to publish textGrain as open source, and that its technical report will receive further detail in the coming weeks.
OpenAI reports little change on model benchmarks
The company also compared watermarked and unwatermarked output from Astra, which it describes as its latest frontier model, across its selected benchmarks. OpenAI says it found no meaningful performance differences. On the Artificial Analysis Intelligence Index, Astra scored 49.57 without watermarking and 49.76 with it; on DeepSWE v1.1, the respective scores were 72.80 and 71.68.
Those figures are OpenAI’s own benchmark results, not a guarantee that every output will be unaffected. The comparison shows small differences in both directions across the listed tests, while the detection results highlight a separate limit: shorter, constrained or edited passages can be harder to classify. The company says these uncertainties help explain its decision to keep detector access restricted while it evaluates the technology.
A watermark cannot establish who wrote a passage
OpenAI says a positive result can indicate that one of its systems generated or processed some text, but cannot measure how much a person contributed. It does not identify the user, establish ownership or responsibility, or verify whether the passage is accurate. A negative result does not prove human authorship: the text may be too short, edited, translated, from an unsupported model, or produced before watermarking began.
Those limits make text provenance different from a simple authorship label. OpenAI’s plan offers a machine-readable signal for some of its outputs, but the signal can weaken as the text changes and cannot resolve who shaped the final passage. At launch, the company is putting watermarks into eligible EU ChatGPT and Codex output, keeping API use opt-in, and reserving its detector for approved applicants.