Does ChatGPT Watermark Its Text? (2026 Update)

The question gets asked more than any other on this site, and the answer has not changed in two years: ChatGPT text carries no watermark. Not a statistical one, not a hidden-character one, not a metadata one. What has changed is everything around it, so here is the full picture as of September 14, 2026.

What OpenAI has said, in order

In May 2024, OpenAI published a post on content provenance that included a section on text. The company said it had built a text watermarking method, that it worked well against local edits such as paraphrasing, and that it was easy to defeat with translation, a rewrite through another model, or inserting and removing characters. It also said the method could disproportionately flag non-native English speakers, and that it was weighing those risks before any release.

Nothing has shipped since. OpenAI's current provenance page describes what it embeds and verifies for images, audio and video. Text is not on it.

In August 2026 the record gained one sentence. OpenAI's help centre article on provenance signals, updated after the company signed the EU Code of Practice on Transparency of AI-Generated Content, now says that "our goal is to expand provenance signals to all modalities including text." That is a statement of intent, not a product: no method, no date, no detector. City AM reported the same commitment on August 12, 2026, alongside the detail that xAI did not sign the code.

That is the whole primary record. When the status tracker re-checks the ChatGPT text row each week, those are the pages it re-reads, and the row has read the same way since launch: nothing deployed.

What people found instead

Two things get mistaken for a ChatGPT text watermark, and both are worth naming precisely.

The first is the April 2025 narrow no-break space. For a few days, o3 and o4-mini emitted U+202F inside long responses. It was real and it was reported carefully, but it was not a watermark: OpenAI called it a quirk of large-scale reinforcement learning, and the characters stopped appearing within days. The full story is short and worth reading, because it set the public expectation that an AI watermark looks like a stray character.

The second is clipboard debris. Copying from the ChatGPT web app can bring HTML fragments, non-breaking spaces and the occasional zero-width character along for the ride, exactly as copying from any web page can. Those are artifacts of the paste, not marks in the text, and the ChatGPT watermark remover takes them out in your browser with a receipt.

What "ChatGPT watermark detectors" actually do

There is no mark, so there is nothing to detect. Every tool sold as a ChatGPT watermark detector does one of two other things.

Most are stylistic classifiers under a new name. GPTZero, Originality.ai and Turnitin score writing style and return a likelihood. A likelihood is not a watermark verdict, and it is wrong often enough that no serious institution should treat it as evidence on its own.

The rest are hidden-character scanners. Those are useful for what they do, which is finding invisible Unicode, but a Clean result from one says nothing about OpenAI, because OpenAI never put anything there.

DetectMark reports the ChatGPT statistical row as Unsupported rather than Unknown. Unknown means a mark exists and cannot be verified. Unsupported means there is no deployed mark to be uncertain about.

How ChatGPT compares to Claude and Gemini

This is the part that surprises people in 2026. The two providers who do watermark text are not OpenAI.

Google says essentially all Gemini text carries SynthID Text, a keyed bias in word choice, and an independent test on August 23, 2026 found the signal in Gemini 3.5 Flash. Anthropic marks Claude text the same way for models launched on or after August 2, 2026, with retrofits of older models in progress. Neither mark is a hidden character, and neither has a public detector yet: Google's portal is waitlist-only and Anthropic's API is in private preview. The comparison covers the mechanism they share.

OpenAI's marks are elsewhere. Every ChatGPT image has carried SynthID plus Content Credentials since May 19, 2026, and audio has carried an embedded mark since July 31, 2026. Both are verifiable through OpenAI's own tool and API, and the image check can be run from this site with one consent click. Text remains the exception.

Will OpenAI watermark text eventually?

Yes, by its own statement, and the reason is regulation rather than product. The EU AI Act's Article 50 became enforceable on August 2, 2026 and requires generative AI output to be marked in a machine-readable, detectable way. OpenAI signed the EU Code of Practice on Transparency of AI-Generated Content, and its help centre now names text as a modality it intends to cover. Systems on the EU market before the enforcement date have until December 2, 2026 to comply.

What that turns into for ChatGPT text has not been announced. It could be the 2024 method, a SynthID-style scheme like Google's and Anthropic's, or something new; there is no date and no public detail. The row on the status page will move the day a primary source does.

What to do with a ChatGPT paste today

Run it through the ChatGPT watermark detector. Hidden characters, clipboard fragments and lookalike letters each get Clean or Found with code points and positions, and the statistical row reads Unsupported with the reason attached. Nothing is uploaded. If something is found, the remover strips what is safe to strip by default and shows you the rest.