OpenAI Watermark

An OpenAI watermark is not one universal marker. TextGrain places a statistical signal in eligible generated text. Supported images and audio can carry SynthID watermarks, while supported images can also include signed C2PA Content Credentials. The relevant check depends on the content and the originating product.

Official announcement, access guide and technical report reviewed October 6, 2026. Availability below distinguishes announced rollout from verified access. This site is independent and does not offer a TextGrain detector.

Which OpenAI provenance signal applies?

Signal	Content	What to check
TextGrain	Eligible text output	A keyed statistical relationship in token choices, not extra Unicode characters.
SynthID	Supported images and audio	A watermark embedded in the media itself, separate from a visible label.
C2PA Content Credentials	Supported images	Signed metadata describing origin and AI-use details; metadata can be removed by editing or sharing.

Does every ChatGPT response have a watermark?

No. OpenAI announced on October 5, 2026 that eligible ChatGPT and Codex text in the EU would receive watermarking over the coming weeks. Its API rollout is global opt-in for select models, with watermarking off by default. An announcement is not a check of every account or previously generated passage. See the ChatGPT rollout guide for product and region distinctions.

Where can you verify an OpenAI watermark?

The Content Provenance API guide describes checks for supported images and audio, with a browser tool at OpenAI Verify. Text verification requires approved organizational access; the public media interface is not a public TextGrain detector. This site does not run any of those verification services. Read our TextGrain detector access guide.

Can copying or character cleanup remove it?

Copying plain text can discard formatting without changing the token sequence that carries a statistical signal. Our hidden-character cleaner changes selected Unicode characters only. It cannot certify that a TextGrain signal is absent. File metadata, a watermark embedded in media, and text token statistics require different checks.

What a result cannot establish

A recognized signal is limited provenance evidence, not a complete history of a work. It does not establish an individual creator, ownership or factual accuracy. Non-detection does not prove human authorship. Unsupported models, older outputs, short passages and subsequent transformations can leave a verifier without usable evidence. Compare TextGrain and SynthID before applying one tool\'s conclusions to another type of content.