# Does ChatGPT watermark text?

Originally published at https://gifi.ai/guides/does-chatgpt-watermark-text.

As of August 2026, OpenAI has not shipped a text watermark in ChatGPT. Images from ChatGPT, Codex and the API carry a C2PA Content Credential and SynthID in the pixels. Supported audio carries SynthID. openai.com/verify reads those file signals. It will not tell you whether a paragraph came from ChatGPT. Gifi will not sell a rewrite as “removing ChatGPT’s text watermark.”

Reviewed by [Can Balkaya](https://gifi.ai/authors/can-balkaya). Published 2026-08-17; updated 2026-08-17.

## How to verify the job

1. **Inspect the words.** Paste the text on the home page. You will see invisible Unicode and a style heuristic. You will not see a ChatGPT text watermark, because none has been published.
2. **Separate the file.** If the concern is a ChatGPT image, inspect the file. C2PA can be stripped. SynthID in the pixels is out of scope.
3. **Do not buy the wrong job.** A rewrite changes prose. It is the lever for a statistical text mark. ChatGPT has not shipped that mark on text. Do not spend a credit to “remove” a signal that is not there.

## What is live today

OpenAI documents two provenance layers on supported images: a signed C2PA note in the file, and SynthID embedded in the pixels. Supported audio generated with OpenAI tools also carries SynthID. The public verification tool looks for those signals. It does not read a paragraph.

C2PA can be stripped from the local file. Soft binding can still re-link a remote manifest. Pixel-embedded SynthID, and audio, are out of scope for Gifi.

## What is promised, and what is not

OpenAI signed the same EU Code of Practice as Anthropic and Google. The company has said it wants machine-readable marks on every modality, including text. That is a goal. It is not a shipping text method.

A classifier that scores “how AI a sentence looks” is not a ChatGPT watermark. Those tools guess. They are not reading a key OpenAI left in the words.

## What Gifi will do with ChatGPT text

The home page will name hidden characters if they are there. That inspection stays in the browser. Cleaning those characters is a Unicode job, not a watermark job.

If OpenAI later ships a text mark, the honest lever is the same as Claude or Gemini: choose different words. Until they publish that method, we will not pretend ChatGPT replies already carry it.

## Questions

### Does every ChatGPT reply have a watermark?

The pictures and supported voice, yes. The words are not documented as watermarked.

### Is openai.com/verify a text detector?

No. It looks for OpenAI’s file signals — C2PA and SynthID in media. It will not see a rewrite, and it will not see Claude.

### Can you remove a ChatGPT image watermark?

We can strip the local C2PA note. SynthID in the pixels is out of scope. Audio is out of scope.

## Primary sources

- [OpenAI on provenance signals](https://help.openai.com/en/articles/8912793-provenance-signals-content-credentials-synthid-in-openai-generated-content)
- [OpenAI on expanding provenance](https://openai.com/index/advancing-content-provenance/)