What an AI watermark is and why you cannot rely on it in a text
An AI watermark is a hidden mark the model itself builds into the generated text — either as invisible characters or as a deliberate bias in word choice. Gemini adds a statistical mark today and Claude has done so since August 2026, so assuming that generated text carries none no longer holds. For an independent tool nothing changes, though: such a mark can only be verified by the operator of the model, who holds the key, and it leaves no visible trace in the text. A watermark is therefore an addition to detection, not a test in its own right.
Two kinds of mark
The first kind is invisible characters. A zero-width space, a soft hyphen or an unusual variant of a space is inserted into the text — none of it is visible on screen, but it stays in the character form of the text. It carries over even when you copy the text from a chat into a document, so it travels all the way into the submitted work.
The second kind is statistical. While generating, the model slightly tilts the choice between equally good words according to a hidden key, so the resulting text carries a pattern that can be computed back. Nothing is added to the text, not a single character — only the basis on which the model picks between equivalent continuations changes. This is exactly how SynthID works, the technique used by Gemini and, since August 2026, by Claude.
Why it disappears before you start looking
Invisible characters are reliably removed by any pass through plain text: pasting without formatting, retyping in another editor, converting to PDF and back. The statistical pattern is tougher — it usually survives light editing, but rewriting sentence by sentence, translating into another language or mixing it with your own text breaks it apart. It also does not form equally strongly everywhere: a short text, a list of figures or a passage with fixed wording gives the model nothing to choose between, and for code the watermark is not applied at all.
On top of that comes a practical obstacle: a statistical watermark can only be verified by whoever knows the key, that is the operator of the model. Anthropic has yet to ship its detection API and Google releases its own only in a limited way — an independent tool cannot work with such a mark even when it is in the text. The absence of a find therefore says nothing at all about where the text came from.
How to work with it in practice
Treat a find as strong evidence about a specific passage, not as a verdict on the whole document — the characters could have come in with a quotation from elsewhere. Treat a non-find as nothing at all; it does not prove a human wrote it. DetekceGPT therefore looks for hidden characters separately, alongside the score, so they turn up even in a text that otherwise reads perfectly naturally.
Build the decision on several layers at once: the general score, the specific suspicious sentences, the deviation from the author's earlier texts, and only then the hidden characters. When the layers agree, you have something to talk about. When they diverge, what follows is a conversation with the author, not an accusation.
FAQ
Can I spot invisible characters myself?
In an editor that can display hidden characters, yes. They are easy to miss, though, and their absence proves nothing, so a manual check only partly replaces a detector.
Does humanising a text remove the watermark?
It removes invisible characters, and in most cases the statistical pattern too. But a rewrite also changes the style, so the general score shifts as well — both show up by a different route.
Does ChatGPT mark its texts?
OpenAI has not publicly launched any text watermarking for ChatGPT. Claude and Gemini do add a mark, but only their operator can verify it — with neither of them can you tell from the text itself.
Try it on your own text
Paste a text or upload a file and look at the score and at the specific sentences that came out suspicious.
Run a detection