Does ChatGPT watermark its text?

Updated June 10, 2026

The persistent rumor: ChatGPT hides a secret signature in its output that detectors read. The reality is more interesting — watermarking is real technology, it mostly isn't deployed for text, and the reasons why tell you a lot about detection.

The short answer

As of mid-2026, ChatGPT's text output carries no deployed cryptographic watermark that public detectors read. Detectors like GPTZero and Turnitin work statistically — measuring predictability and structure — not by finding hidden marks. OpenAI has confirmed it built a text watermarking system and, by its own public statements, held off deploying it.

How text watermarking would work

The elegant idea (Google DeepMind's SynthID-Text, deployed for Gemini, works along these lines): during generation, secretly bias word choices using a key — at each step, slightly prefer words from a pseudorandom "green list". Readers notice nothing; a checker with the key finds statistically impossible green-word density. It's invisible ink made of probability.

Why OpenAI shelved it

Three stated and structural reasons: paraphrasing largely washes the signal out (reword the text and the green-word bias dies — a fragility OpenAI acknowledged); it only marks your own models (a watermarked ChatGPT pushes evaders to any unwatermarked alternative, including open-source); and it punishes legitimate users — the non-native speakers and everyday writers using AI for polish would carry a permanent mark while determined misuse routes around it.

Watermarking works best in images and audio, where the signal survives editing better. For text, the economics keep pointing back to statistical detection — with all its known limits.

What this means for you

Don't fear hidden signatures; do expect statistical detection everywhere, since that's what actually runs. The response is unchanged: make text genuinely read human — varied, specific, yours — and verify the score rather than guessing. If universal watermarking ever arrives, it'll be public policy news, not a rumor; until then, the standard workflow is the durable one.

Frequently asked questions

Can OpenAI tell if a specific text came from ChatGPT?

Without a deployed watermark, not reliably — and OpenAI shut down its own statistical classifier in 2023 over low accuracy. No public "was this ChatGPT?" oracle exists.

Is Gemini's text watermarked?

Google has deployed SynthID-Text for Gemini output and published the research. Its checker isn't a general public detector, and paraphrase-fragility applies there too.

Do invisible Unicode characters in AI text count as watermarks?

Occasional artifacts (special spaces, smart quotes) aren't cryptographic watermarks, just formatting residue — trivially stripped, and some detectors flag the cleanup itself. Not the mechanism to worry about.

Humanize it — then verify it

Paste your text, get a rewrite that reads like a person wrote it, and check the AI-probability score yourself before anyone else does. 3-day free trial.