Does ChatGPT watermark text? (August 2026)

What OpenAI has actually shipped

Images from ChatGPT, Codex and the API carry SynthID plus C2PA Content Credentials. Voice output was added on 31 July 2026, one day before the EU AI Act's Article 50 transparency obligations began to bite on 2 August.

Text is the conspicuous gap. That is not an oversight — Article 50 requires machine-readable marking “where technically feasible”, and text is the case where feasibility is genuinely contested.

Why text is the hard one

An image has enormous redundancy. You can perturb thousands of pixels imperceptibly and carry a robust signal. Text has almost none: every token is load-bearing, and the only place to hide a signal is in which word gets chosen among near-equivalent options.

That is what distributional watermarking does, and it comes with real costs. It needs long passages to reach statistical confidence, it degrades under paraphrase and translation, it is weak on low-entropy text like code, and it constrains the sampler in ways that can nudge output quality. Those trade-offs are why several labs have built text watermarking and not deployed it.

How this compares to Claude and Gemini

**Gemini:** Google has applied SynthID to Gemini text output since 2024. It is the one that has been marking text longest.

**Claude:** Anthropic marks text from models launched on or after 2 August 2026. Every shipped Claude model predates that, and Anthropic says earlier ones are “in progress” under a transition period.

**ChatGPT:** images and voice yes, text not shipped.

In all three cases detection requires the provider's secret key, and none of them has published a detector the public can run. So even where marking exists, no third party can verify a passage.

What is detectable in ChatGPT text today

The same things that are detectable in any chat output, none of which are watermarks. Copying from the web interface carries HTML into whatever you paste into. Zero-width characters, non-breaking and exotic spaces survive most paste routes. Smart quotes and em dashes survive all of them.

Those are checkable by anyone who looks at the bytes, which makes them a far more realistic exposure than a watermark nobody can currently test for.

Questions

Will OpenAI add text watermarking later?

It has built text watermarking research internally and not released it. Article 50 pressure makes it more likely over time, but nothing is shipped as of August 2026 and no date has been announced.

Do AI detectors detect an OpenAI watermark?

No. Tools like GPTZero and Turnitin are statistical classifiers that score how predictable your word choices are. They have no access to any provider's watermark key and are not testing for one.

Are ChatGPT images watermarked?

Yes — SynthID plus C2PA Content Credentials. That is a genuinely different and far more robust mechanism than anything applied to text.