ChatGPT watermark detector

No. OpenAI built a detector for its textGrain watermark and, since announcing the watermark on 5 October 2026, has kept that detector out of public hands.

The company gave a short reason: it held the tool back "given the risk of missed watermarks and false positives." Its developer documentation describes the route that does exist. Text verification goes to approved organizations, with AI research and academic institutions given as examples, and OpenAI weighs every application on its own merits.

The technical report on textGrain shows why copying the software would not help an outsider. The detector takes two inputs, the passage and OpenAI's secret key, so the key is the real gate. When we read the developer docs on 19 October 2026 they still described text checks as limited to approved organizations, and we found no announcement that widens that access.

Sources: OpenAI, 5 Oct 2026 · OpenAI developer docs, content provenance · textGrain technical report

Characters that print nothing, such as zero-width spaces, byte-order marks and odd spaces, and nothing at all about which words a model chose.

Our scan looks for 19 groups of invisible characters and 6 groups of look-alike spaces. Each hit appears as a small orange tag; hover it to see the code point, such as U+200B for a zero-width space, while a counter totals every kind it found. The whole scan runs in the page. Your text reaches our server only when you press rewrite.

We put the scan to a test on 19 October 2026, running the site's own scan and clean code on a sentence we built with seven planted characters, plus a plain 78-character sentence as a control.

Planted character	Code point	What the scan did
Zero-width space	U+200B	Reported, deleted
Narrow no-break space	U+202F	Reported, turned into a plain space
No-break space	U+00A0	Reported, turned into a plain space
Soft hyphen	U+00AD	Reported, deleted
Byte-order mark	U+FEFF	Reported, deleted
Word joiner	U+2060	Reported, deleted
Unicode tag character	U+E0000 block	Reported, deleted
Control sentence, no planted characters	none	0 hits

Seven of seven planted characters were caught and the visible words came back unchanged. The test string was made by hand, so it shows what the scan catches and nothing about textGrain.

Be careful about what a hit means. In April 2025 the startup Rumi reported narrow no-break spaces in text from OpenAI's o3 and o4-mini models. Rumi's own newsroom page says OpenAI answered that they came from reinforcement learning rather than a watermark, and that by 23 April Rumi no longer saw them. A stray character shows that something in the text's path left it there, not which tool did.

Sources: Rumi newsroom, April 2025 · own test, 19 Oct 2026, method above

Only the company holding each key, or an organization it lets in. As of 19 October 2026, the hidden-character scan is the one check open to everybody.

What you want to check	Who can run it (19 Oct 2026)	What a hit shows
Invisible characters in any text	Anyone, free, on this page	The characters are present; not which tool added them
ChatGPT text (textGrain)	Approved organizations that applied to OpenAI	An OpenAI model likely picked much of the wording
Claude text	Eligible organizations, through Anthropic's private-preview detection API	Claude was "likely involved" at some point
Gemini text	No public checker; Google's open detector reads images, video and audio	Nothing yet for text
ChatGPT images and audio	Anyone, on OpenAI's verify page or its provenance API	OpenAI's own signals are in the file
Writing style ("AI detector" scores)	Anyone	A statistical guess about style; no key involved

Anthropic's eligible groups are categories set by EU law: regulators, law enforcement, media, fact-checkers, independent researchers, educational organizations and EU civil society groups, plus companies that must verify marks for their own compliance. Its free Content Checker reads files, and the tool's page states that it "does not check text."

Google opened the SynthID Detector to everyone on 7 October 2026. Gemini text in the app and on the web has carried SynthID since May 2024, yet the detector lists only images, video and audio. The Claude side is covered in depth in the Claude AI detector guide.

Sources: OpenAI developer docs · Anthropic, Claude text watermark · Claude Content Checker · SynthID Detector · Google DeepMind, May 2024

Yes. OpenAI's public verify page accepts images and audio files, and the same page says it does not currently support text.

Developers can run that check in code through the Content Provenance API. Keep one limit in mind: a "not detected" answer covers only OpenAI's own signals, so it tells you nothing about a picture made by another company's model. These checks work on files, which is why none of them can be pointed at a paragraph.

Sources: OpenAI verify page · Content Provenance API docs

No. A style detector holds no key, and OpenAI's report says its check needs both the text and the secret key, so an outside site has nothing to run the test with.

The two tools answer different questions. A watermark test asks whether a keyed pattern shaped the word choices. An AI detector asks whether the writing resembles machine text in general. Anthropic draws the same line, saying such software works another way because its makers "don't have our key." Google's detector page sorts general detectors into tools that analyze "passive artifacts and patterns."

At least one site that sells a humanizer now advertises a detector that checks text for textGrain. OpenAI's only documented route to the text check is an application from an approved organization, so a reader has no way to confirm what such a score measures. A quick test helps: a verdict that arrives in seconds, with a percentage, and without any access application, is a style score or a character scan.

Sources: textGrain technical report · Anthropic, Claude text watermark · SynthID Detector FAQ

Less than most people assume: a hit means an OpenAI model shaped the wording, a miss means very little, and neither one names a person.

What you see	What it can mean	What it cannot show
Watermark found (key holders only)	An OpenAI model wrote or shaped much of the wording	Who used the model, or how much of the thinking was theirs
No watermark found	The passage may be short, edited, translated, from an unsupported model, older than the watermark, or from another company	That a person wrote it
Hidden characters found	Something along the text's path added them	That ChatGPT wrote it
No hidden characters	The text holds none	That it carries no watermark, since that sits in the words
High or low style score	The writing does or does not resemble machine text	Anything about a watermark

OpenAI is direct about the first two rows. It writes that a watermark "does not measure human contribution," and that a missing one can come from any of the causes listed above. Reliability also shifts with length and subject. At a 1% false-positive target, OpenAI reports catching roughly four in five 200-token passages and nineteen in twenty 400-token passages on topics such as psychology, and substantially fewer on mathematics. Edits weaken a hit as well: in OpenAI's own trial on 400-token passages, trading one word in ten for a synonym dropped detection from about 92% to about 66%. The full rewording figures sit in the rewording section of the home page.

If you wrote the text yourself and someone points to a stray character or a style score, neither is evidence of who wrote it. The scan above shows exactly where any odd character sits, and the free cleaner takes it out.

Source: OpenAI, 5 Oct 2026

Clear the invisible characters first, then change the wording if the passage came from an AI tool and you want it in your own words.

The scan and the clean are free and need no account. If the passage began as an AI draft and has to read as your own writing, the rewrite tab on the ChatGPT watermark remover is the next step: it puts each sentence into new wording and removes hidden characters in the same pass. It needs at least 20 words. Read the result against your original for names, numbers and quotes before you use it.

Sign in with an email address and the free allowance is four rewrites of up to 250 words each, enough to try it on a real paragraph. Paid plans are Basic at $18 a month, Pro at $27 and Ultra at $48, or $144, $216 and $432 a year, which brings Basic down to $12 a month. For how Claude and Gemini marks are handled, read the AI text watermark remover guide.

See the plans