AI Fingerprint vs AI Watermark vs Hidden Characters: Which One Is in Your Text?
Short answer: “AI fingerprint” is not a technical term, and people use it for three different things. A watermark is a signal the provider put there deliberately with a secret key, in word choice or in pixel values. Hidden characters are invisible Unicode, almost always debris from a paste. A stylistic fingerprint is what AI detectors score: patterns a model tends toward, with no key and no verdict. Only the first two can be checked exactly, and only the second can be removed by a tool.
People type "AI fingerprint" into a search box and mean three different things, and the tools that answer them rarely say which one they found. This article fixes the vocabulary DetectMark uses, so a verdict on this site always refers to one specific thing.
Three things behind one word
A watermark is a signal the provider put there on purpose, with a secret key, so that a detector holding the key can find it later. In text it is a bias in word choice; in images, audio and video it is a pattern in the content itself. Google's SynthID, Anthropic's Claude mark and the SynthID that OpenAI embeds in its images and audio are all watermarks in this sense.
Hidden characters are invisible or near-invisible Unicode code points inside text: zero-width spaces, soft hyphens, byte-order marks, non-breaking spaces, direction controls. They are almost always debris from a copy and paste or from an editor the text passed through. Occasionally a model emits them by accident, as two OpenAI models did for a few days in April 2025. No current provider uses them as a watermark, and Anthropic has said so in writing:
"Nothing is added to the text and there are no hidden characters."
A stylistic fingerprint is what AI detectors such as GPTZero, Originality.ai and Turnitin score: patterns of word choice, sentence rhythm and predictability that a model tends toward. There is no key. Nobody put the pattern there deliberately. A classifier estimates how model-like the writing looks and returns a likelihood, and it is wrong often enough that the number should never be treated as a verdict.
A fourth thing sits beside them for files. Content Credentials, the consumer name for a C2PA manifest, are a signed label in an image's metadata that records what made it. That is a label in the file, not a mark in the content, which is why it can be read locally and stripped locally.
The comparison
| Watermark | Hidden characters | Stylistic fingerprint | Content Credentials | |
|---|---|---|---|---|
| Put there by | The provider, deliberately, with a key | Paste, editors, rare model quirks | Nobody; a side effect of how models write | The generating tool, as signed metadata |
| Where it lives | Word choice, or pixel and sample values | Between the visible characters | The style of the whole text | A metadata block in the file |
| Can you see it | No | No, until you look at code points | No | No, until you read the file |
| Can a browser check it | Not without the provider's key | Yes, exactly | No, only estimate | Yes, exactly |
| Removed by | A full rewrite (text); nothing verifiable (pixels) | Deleting the characters | Rewriting in a different style, unreliably | Stripping the metadata |
| DetectMark verdict | Unknown, unless a provider verifies it | Clean or Found, with code points | Never given | Clean or Found, with the manifest shown |
Why the difference decides everything
The three are different kinds of claim, and mixing them produces most of the bad results on the web.
A hidden-character scanner that reports Clean is telling you the text has no invisible code points. It is not telling you the text has no watermark, because a watermark was never in the characters. Every "Claude watermark remover" that strips zero-width spaces makes exactly this error, and the invisible characters story explains how it spread.
A classifier that reports 87% AI is telling you the writing looks model-like to its model. It is not reading a watermark, and it cannot, because it has no key. Removing hidden characters changes its score by nothing, and neither does anything else short of rewriting.
A watermark detector, when one exists, is the only tool that reads a signal a provider deliberately placed. It says the provider's model produced or processed the content. It does not say who used it or whether they had the right to.
Which one is in your text?
Work down the list.
- Hidden characters: paste the text into the invisible character remover. Every invisible code point is listed with its position, and the inert ones come off with one click. This check is exact.
- Watermark: ask which model produced the text and check the status page. Gemini text is marked; Claude models launched since August 2, 2026 are marked; ChatGPT text is not. Then ask whether a detector is open to you. For Gemini, it is a waitlist; for Claude, a private preview; for OpenAI images, a public API that DetectMark runs on consent. Where no detector is open, the honest verdict is Unknown.
- Stylistic fingerprint: if a classifier score is the only evidence someone has, it is not evidence of a watermark, and it is weak evidence of anything. Treat it as a prompt to ask questions, not as a finding.
The vocabulary DetectMark uses
A scan result on this site is always one of four words. Clean: checked, nothing found. Found: checked, a known signal present, listed. Unknown: a mark may exist and cannot be verified from here, with the reason. Unsupported: the check does not apply to this content. There is never a percentage and never an AI likelihood, because a percentage would be the stylistic fingerprint dressed as a watermark verdict.
"AI fingerprint" appears on the homepage as a synonym, because that is the phrase people search. What the tool then reports is one of the three specific things above, named.
Find out which one is in your text
Runs locally · nothing uploaded
FAQ
A colloquial term with no single meaning. Most people mean either a watermark, which a provider embeds deliberately with a key, or the stylistic pattern that AI detectors such as GPTZero score. DetectMark treats it as a synonym and then reports the specific thing it found.
No. Current text watermarks from Google and Anthropic live in word choice and add nothing to the text. Hidden characters are invisible Unicode from a paste, an editor or a rare model quirk, and a hidden-character scanner that reports Clean says nothing about a watermark.
Not without the provider’s key. A browser can find hidden characters and read Content Credentials exactly, but a statistical text watermark or a pixel mark needs the provider’s own detector. DetectMark reports Unknown in that case, and runs OpenAI’s verifier on consent for OpenAI images.
No. GPTZero, Originality.ai and Turnitin estimate how model-like the writing looks and return a likelihood. They read no key and detect no mark, and removing hidden characters changes their score by nothing.
Clean, Found, Unknown or Unsupported, each with a reason. Never a percentage or an AI likelihood, because a percentage would be a stylistic guess presented as a watermark verdict.
- 1. Anthropic: “How Claude’s text watermark works” (August 14, 2026)
- 2. Dathathri et al.: “Scalable watermarking for identifying large language model outputs”, Nature 634 (2024)
- 3. Kirchenbauer et al.: “A Watermark for Large Language Models” (arXiv, January 2023)
- 4. Scott Aaronson: “Anthropic’s LLM watermarking” (August 22, 2026)
- 5. Sebastian Raschka: “How Claude Watermarks AI-Generated Text” (August 22, 2026)
- 6. BleepingComputer: “AI ‘watermark removers’ flood the web. Almost none can prove they work.”
- 7. Unicode Technical Standard #39: Unicode Security Mechanisms (confusable characters)