ChatGPT watermark

What the evidence says

OpenAI announced a text watermark called textGrain on 5 October 2026, in response to the EU AI Act. The scope is phased, and narrower than a headline suggests. From that day, API customers anywhere can opt in for select models, and text watermarking stays off by default in the API. Eligible ChatGPT and Codex text output in the European Union is to carry the mark “over the coming weeks”, on all plans. OpenAI states that it is not making text watermarking a global default at launch, so the answer for a given passage depends on where, when and through which product it was generated.

The method is a statistical signal in word choice. OpenAI describes textGrain as adding an invisible statistical signal to the model’s word choices, and the technical report published with the announcement specifies it: token sampling is coupled to pseudorandom values derived from a secret key and a window of preceding tokens, within a budget that limits how much sampling entropy the watermark may remove. The detector needs only the generated text and the key. Neither document describes inserting any character; that nothing is added to the text follows from the published mechanism rather than from a statement about Unicode.

Detection is not open to the public. OpenAI opened applications for its text watermark detector on the day of the announcement, with access limited at first to approved researchers and expert organisations, and says it is not making the tool publicly available at launch because of the risk of missed watermarks and false positives. Its verification tools for images and audio stay public. Those cover media files, through Content Credentials and SynthID watermarks on supported images and audio, and none of them reads a paragraph pasted into a document.

The stated limits are measured. At a target false positive rate of 1 percent, the detector identified the mark in about 80 percent of 200-token passages and about 95 percent of 400-token passages on psychology questions, and in substantially fewer on mathematics, where there is less flexibility in word choice. In 400-token passages, replacing 10 percent of the words with synonyms reduced detection from about 92 percent to 66 percent, and replacing 25 percent reduced it to 17 percent. OpenAI adds that a watermark does not measure human contribution, establish ownership, identify the user or verify accuracy, and that the absence of a detected watermark does not prove human authorship.

A separate observation still causes confusion. In April 2025 the education startup Rumi reported that OpenAI’s o3 and o4-mini reasoning models were leaving unusual Unicode characters in long answers, most often the narrow no-break space, U+202F, and a developer forum thread opened in October 2025 reports the same character from GPT-5. Rumi later added that OpenAI had contacted them and indicated the characters are not a watermark, quoting OpenAI calling them “a quirk of large-scale reinforcement learning”. That exchange was reported by one party, and the textGrain announcement does not mention these characters. The watermark OpenAI did announce works through word choice and uses none of them.

This entry will keep changing. OpenAI signed the European code of practice on transparency of AI-generated content, says it expects to revisit each part of its approach, plans to make the technology available in open source, and is working with cloud partners to offer the watermark for its models on their services. The European Commission’s FAQ on the code, last updated on 29 July 2026, describes a rule agreed by the EU co-legislators and still to be adopted at that date, which would give systems already on the market until 2 December 2026 to meet the marking obligation of Article 50(2).

Sources

1. Our approach to EU text provenance rulesOpenAI · 2026-10-05 · Provider documentation
2. textGrain: Entropy-Calibrated Watermarking for Language Model TextOpenAI, technical report · 2026-10-05 · Provider documentation
3. New ChatGPT models seem to leave watermarks on textRumi · 2025-04-20 · Independent report
4. GPT-5 (non-reasoning) outputs U+202F (narrow no-break space) instead of normal spacesOpenAI Developer Community · 2025-10-13 · Independent report
5. Strong backing for the Code of Practice on Transparency of AI-generated ContentEuropean Commission · 2026-07-31 · Independent report
6. Signing the Code of Practice on Transparency of AI-generated ContentEuropean Commission · 2026-07-29 · Independent report