Is my ChatGPT text watermarked?

OpenAI announced it on 5 October 2026:

“The technology, called textGrain, adds a statistical signal through the model’s word choices. A detector can look for this signal, but short passages and later edits make it less reliable. This is why detector access will initially be limited to approved researchers and expert organizations.”

Its technical report says the detector “requires only the generated text and the secret key”. Without OpenAI’s key there is nothing to test the words against, so no public website can tell you whether a text carries it, this one included.

Where it applies:

“Over the coming weeks, OpenAI will also add an invisible watermark to eligible ChatGPT and Codex text generated in the EU. It will remain off by default in the API and will not become a global default at launch.”

“The watermark can indicate that an OpenAI system generated or processed part of a text. It cannot tell who used the system, how much a person contributed, who owns the text, or whether the content is accurate. If no watermark is detected, that also does not prove that a person wrote it.”

Article 50(2) of the EU AI Act:

“Providers of AI systems, including general-purpose AI systems, generating synthetic audio, image, video or text content, shall ensure that the outputs of the AI system are marked in a machine-readable format and detectable as artificially generated or manipulated.”

Among its exceptions, it doesn’t apply where the AI system performs “an assistive function for standard editing” or doesn’t substantially alter what it was given. The Act applies from 2 August 2026 (Article 113), and AI systems already on the market before then have until 2 December 2026 to meet this rule (Article 111(4)).

• Zero-width characters. Characters with no width, so nothing shows. Copied text often carries them; they can also mark a copy so it can be traced.
• Tag characters. Invisible copies of letters. Apart from a few flag emoji they have no ordinary use, yet software reading the text, an AI model included, still receives the letters.
• Look-alike letters. Cyrillic or Greek letters inside a Latin word, such as the а in pаypal. They look the same but don't match in search, spell-check or a filter.
• Direction marks. Marks that change which way text runs. Normal in Arabic or Hebrew; elsewhere they can show text in a different order from how it's stored.
• Stray variation selectors. Invisible marks that pick a glyph's style. Normal after an emoji; after an ordinary letter they do nothing visible and can carry hidden data.
• Control characters. Leftovers from terminals and printers that don't belong in prose and can break forms and code.
• Unusual spaces. Spaces that aren't the space bar's, such as the no-break space that keeps 10 and km together. Mostly harmless in prose; they can break code and search.

It leaves alone the joiners that emoji and scripts such as Hindi and Persian need, and the selectors that give an emoji its colour.

Sources, checked 6 October 2026: OpenAI's announcement (developer forum); the textGrain technical report; AI Act Article 50; Article 111; Article 113. OpenAI is rolling the watermark out over several weeks, so check its announcement for changes.