ChatGPT watermark

What the evidence says

OpenAI has not shipped a text watermark. Its own help documentation puts the question directly — whether provenance signals will cover text output — and answers it in the future tense, so by the provider’s own account there is nothing in ChatGPT’s prose to detect. The same holds for the API and Codex.

Provenance does exist on other media. OpenAI attaches C2PA Content Credentials to generated images and applies marking to generated audio. Those signals travel in media files and none of them reaches a paragraph pasted into a document.

The confusion comes from a real observation. 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. The character is real, it is invisible in most editors, and developers have kept reporting it since on later models, sometimes because it breaks rendering in other applications.

What it is not is a watermark, though the record on that is thinner than it is usually made to sound. 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 private exchange, reported by one party, is the entire basis for every “OpenAI says” sentence written on this subject since; OpenAI has never addressed these characters publicly, and its provenance documentation never presents any Unicode character as a signal. The mundane explanation fits: these models learned the typography of the professional and multilingual text they were trained on, and reproduce it.

This is the entry most likely to change. OpenAI signed the European code of practice on transparency of AI-generated content, and Article 50(2) of the AI Act names synthetic text alongside audio, image and video. Systems already on the market have until 2 December 2026 to meet the marking obligation, which makes that date the one to watch on this page.

Sources

1. Provenance signals: Content Credentials and SynthID in OpenAI-generated contentOpenAI Help Center · 2026-09-10 · Provider documentation
2. New ChatGPT models seem to leave watermarks on textRumi · 2025-04-14 · Independent report
3. GPT-5 outputs U+202F instead of normal spaces and breaks text renderingOpenAI Developer Community · 2025-08-12 · Independent report
4. Strong backing for the Code of Practice on Transparency of AI-generated ContentEuropean Commission · 2026-07-31 · Independent report