OpenAI will add an invisible text watermark to ChatGPT and Codex in the EU over the coming weeks, and lets API customers worldwide opt in starting today

A detector that reads a signal hidden in word choices now exists for text written by OpenAI's models.

OpenAI said on 5 October 2026 that its text watermarking technology, called textGrain, will roll out to ChatGPT and Codex users in the European Union over the coming weeks, that API customers worldwide can opt in starting the same day, and that approved researchers can apply for detector access. The company's post frames the rollout as a response to the EU AI Act, which requires generative AI providers to make generated text identifiable in a machine-readable way.

How textGrain works and what it will not do

The post says textGrain adds "an invisible statistical signal to the model's word choices," and OpenAI's detector looks for that signal in a passage. In OpenAI's own evaluations, textGrain "matched or exceeded the performance of other approaches we tested, including SynthID for text," which is the technology Google DeepMind released and that Anthropic adopted for its own watermark in August.

OpenAI publishes the error profile. At a target false positive rate of 1 percent, the detector found watermarks in about 80 percent of 200-token passages and about 95 percent of 400-token passages on content such as psychology. On mathematics the detection rate was lower, because word choice is more constrained. Light editing drops the signal: replacing 10 percent of the words in a 400-token passage with synonyms moved detection from about 92 percent to 66 percent, and replacing 25 percent of the words dropped it to 17 percent.

OpenAI lists five things a watermark does not establish. It does not measure how much a human contributed. It does not say who owns the text or who is responsible for it. It does not identify the user, their account, their prompt, or their conversation. It does not verify accuracy. And the absence of a detected watermark does not prove human authorship, because short text, edited text, translated text, and output from older or other-vendor models can all fail detection.

The three parts of the rollout, with their dates

Text watermarking is "off by default in the API" and will stay that way. From 5 October, API customers in every country can opt in to watermarked outputs for select models. OpenAI says it is working with cloud partners to make the option available for OpenAI model outputs accessed through their services in the coming weeks.

In the European Union, OpenAI will add an invisible watermark to eligible ChatGPT and Codex text output "over the coming weeks," across all plans. The post is explicit that this is a regional step: "We are not making text watermarking a global default at launch."

Detector access opens to approved researchers and expert organizations from 5 October, "on a case-by-case basis." The detector reports whether an OpenAI watermark is present, without identifying the user or revealing their prompt. OpenAI says it is not making the detector publicly available at launch, because of the risk of missed watermarks and false positives. OpenAI says it also plans to open source textGrain so others can build on it.

Across OpenAI's benchmarks for Astra, its latest frontier model, the company does not report a meaningful difference between watermarked and unwatermarked output. Astra scored 53.90 percent unwatermarked and 56.06 percent watermarked on Terminal-Bench 4.0, and 94.44 percent against 93.94 percent on GPQA Diamond. These are OpenAI's own figures.

Source

Primary source: Our approach to EU text provenance rules, OpenAI, 5 October 2026. Technical report: textGrain, entropy-calibrated watermarking for language model text. Reporting: The Verge, Stevie Bonifield, 5 October 2026.