ChatGPT Watermark: The Current Facts

Text: nothing deployed

As of September 14, 2026 there is no text watermark in ChatGPT or the API. OpenAI said in May 2024 that it had built a text watermarking method and was weighing the risks of releasing it; nothing has shipped, and its provenance page covers images, audio and video without mentioning text. What has changed is intent: OpenAI's help centre article on provenance signals, updated in August 2026, says the goal is to expand provenance signals to all modalities including text under its EU transparency-code commitments, with no method, date or detector announced. Until one ships, every "ChatGPT text watermark detector" is a stylistic classifier under a different name, because there is no mark for it to read.

The April 2025 narrow no-break space

For a few days in April 2025, o3 and o4-mini emitted U+202F inside long responses. It was real, it was reported carefully, and it was not a watermark: OpenAI described it as a quirk of large-scale reinforcement learning, and the characters stopped appearing within days.

The story is worth knowing because it set the public expectation that an AI watermark looks like a stray character. It does not, in any current deployment. Characters like that still turn up in pasted chat output for ordinary reasons, and the ChatGPT hidden-character remover takes them out in your browser.

Images: two marks, one file

Since May 19, 2026 images from ChatGPT, the API and Codex carry Google's SynthID in the pixels and C2PA Content Credentials in the file. Those are separate signals with separate fates. DetectMark reads the Content Credentials in the browser and can strip them. The pixel mark stays where it is, and since August 2026 OpenAI verifies it through its public Verify tool and Content Provenance API. That check covers OpenAI-supported signals only: a not-detected result says nothing about other providers or about a file whose marks were stripped.

Audio output has carried an embedded mark since July 31, 2026, verifiable through the same OpenAI Verify tool and API. DetectMark does not handle audio, so audio checks report Unsupported.

To see what a specific paste or image actually carries, the ChatGPT watermark detector runs the text and file checks and labels each row with its reason. For a ChatGPT image specifically, Verify an OpenAI-Generated Image runs OpenAI's own check on one consent click.