What the ChatGPT text watermark is, and what Grok has instead

On 5 October 2026 OpenAI settled a question it had left open for years: ChatGPT text will carry a watermark in the European Union, and nowhere else by default. The mark is not the stray space that character scanners have reported since 2025. Two different things now answer to one name, and telling them apart decides what any tool can show you about a draft.

What did OpenAI announce about watermarking ChatGPT text?

OpenAI’s text watermark is called textGrain, and the company’s post of 5 October 2026 sets out a phased start. API customers anywhere could opt in for select models from that day, with the setting off by default. Eligible ChatGPT and Codex text in the European Union is to get the mark over the coming weeks, on all plans. Applications also opened for a detector, with access limited at first to approved researchers and expert organisations.

The scope is narrower than the headlines. OpenAI says it is not making text watermarking a global default at launch, and that it is working with cloud partners to offer the option on their platforms. A ChatGPT answer written outside the EU, or before the rollout reaches an account, falls outside the scope OpenAI describes.

How does textGrain mark text?

textGrain changes which words the model picks and adds nothing to them. A language model writes one token at a time, drawing each from a ranked set of candidates. The technical report published with the announcement describes how textGrain ties that draw to a secret key and a window of preceding tokens, inside a budget that caps how much of the model’s randomness the watermark may spend. A detector holding the same key reads the finished text and tests whether the choices lean the way the key predicts. It needs only the text and the key.

That mechanism rules out the usual picture of a hidden tag. OpenAI describes the signal as statistical and places it in word choice, and neither the post nor the report mentions inserting a character. Deleting a stray invisible character therefore removes none of the word choices the signal is built from.

Is the narrow no-break space in ChatGPT answers a watermark?

The narrow no-break space, code point 202F, sits behind most claims of a ChatGPT watermark made before October 2026, and the record on it is short. The education startup Rumi reported on 20 April 2025 that OpenAI’s o3 and o4-mini models left the character in longer answers, and that GPT-4o did not. Two days later Rumi added that OpenAI had contacted it to say the characters are not a watermark, calling them “a quirk of large-scale reinforcement learning”. On 23 April Rumi wrote that its tests no longer showed them.

Two limits belong beside that account. The denial reached the public through Rumi, and we found no OpenAI publication that repeats it. The character also came back: a thread opened on OpenAI’s developer forum on 13 October 2025 reports GPT-5 writing 202F in place of ordinary spaces in headings and tables, and the page shows no reply from OpenAI staff. A character that arrives with one model version and leaves with another is a formatting habit, worth cleaning when it breaks a layout. It identifies no author, and textGrain does not use it.

What does Grok document for text?

Grok has no documented text watermark. Its FAQ raises watermarking in one place, under image and video generation: generated images and videos include a visible Grok mark, and the FAQ says no setting removes it. The page is silent on text. City AM reported on 12 August 2026 that xAI (now SpaceXAI), the company behind Grok, had not signed the European Commission’s code of practice on transparency of AI-generated content, and that the paper had approached the company for comment.

Staying out of the code settles less than it sounds. The Commission’s own FAQ says signing is voluntary and that not signing is not non-compliance with the AI Act; a non-signatory must be ready to show compliance by other means. The same FAQ, last updated on 29 July 2026, dates Article 50 from 2 August 2026 and 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. For Grok text, the public record today is an absence of documentation. Nobody has published a denial, and the deadline may change the answer.

How do ChatGPT, Grok, Claude and Gemini compare?

ChatGPT, Claude and Gemini each have a statistical watermark on text that their makers document, and Grok has none. The scopes differ enough that the same question gets a different answer for each assistant.

Microsoft takes a fifth position for the models it hosts. Its Foundry documentation, dated 17 September 2026, says text provenance varies by model and that the underlying model supplies the watermark. The table of supported text models on that page lists four Claude models and no OpenAI model.

What can a textGrain detection result tell you?

A textGrain result is weaker evidence than the word watermark suggests, and OpenAI publishes the numbers. At a target false positive rate of 1%, its detector found the mark in about 80% of 200-token passages and about 95% of 400-token passages on psychology questions. Rates were substantially lower on mathematics, where wording has less room to vary. Editing costs more: in 400-token passages, replacing 10% of the words with synonyms cut detection from about 92% to 66%, and replacing 25% cut it to 17%.

OpenAI draws the conclusion itself: “The absence of a detected watermark does not prove human authorship.” The post also lists what a positive result leaves open: how much a person contributed, who owns or answers for the text, which account or conversation produced it, and whether it is accurate. Anthropic sets the same ceiling for Claude, where a detected mark shows only that Claude was likely involved at some point, and Google notes that a thorough rewrite or a translation greatly reduces its detector’s confidence.

Those limits are the reason OpenAI gives for keeping its detector away from the public at launch. They also mean that no public tool, ours included, can read a textGrain or SynthID mark out of pasted text. The keys are not published.

What should you check in a draft?

A draft raises two separate questions, one about characters and one about wording, and each has its own kind of check.

• Characters: scan a draft when a search fails, a word count looks wrong or a layout breaks. Our scanner reports eight code points and 202F is not one of them, so search for that character by its code in your editor.
• Wording: a statistical watermark sits in the word choices, and no character tool reads or changes it. Our rewriter edits wording so prose reads more naturally and shows every change; it reports nothing about what a keyed detector would find.
• Evidence: treat a detection result, or the lack of one, as a signal about a passage. OpenAI and Anthropic both describe their own detectors in those terms.