AI text detector
See the characters your eyes can't.
Paste text from ChatGPT, Claude, Gemini, or anywhere else to reveal what is hiding in it — zero-width spaces, bidirectional controls, tag characters, unusual spaces, and letters swapped for look-alikes from another alphabet — then copy it back out clean. 539 codepoints checked, in your browser, with an honest account of what they mean.
Nothing to scan yet
Findings appear here as you type. This tool flags genuinely invisible characters and look-alike letters — never ordinary punctuation.
People call these “AI watermarks.” They aren't.
Invisible characters are real and worth finding. They break CMS imports, corrupt e-book formatting, confuse screen readers, and can smuggle hidden data through a covert channel. Removing them is legitimate text hygiene, and this tool does it well.
But nearly every site offering this calls the characters an AI watermark, and that is simply the wrong name for them. The distinction is not pedantry — it is the difference between a useful clue and a false accusation.
If you are trying to work out whether something was AI-made, the honest answer is that this tool cannot tell you, and neither can any of the ones that promise they can. Here is what actually exists.
What this does not detect
The watermarks Anthropic and Google embed in AI text are statistical — patterns in word and token choice, not stray invisible characters. Nothing in this tool can read them, because no public detector for them exists yet.
So: finding hidden characters tells you text passed through software that inserts them. Finding none tells you nothing about whether AI wrote it. Anyone selling you certainty here is guessing.
We also will not flag em dashes, curly quotes, or ordinary punctuation as evidence of anything. Those are visible, legitimate typography, and treating them as proof is how people get wrongly accused.
What gets checked
Not “over 15 types.” The exact number, counted from the scanner itself: 539 codepoints across 7 categories.
Zero-width & invisible
Zero-width spaces, joiners and non-joiners, word joiners, the byte-order mark, and the invisible maths operators.
Bidirectional controls
Left-to-right and right-to-left marks, embeddings, overrides, and isolates — the characters behind text that renders backwards.
Unicode tag characters
A 128-codepoint block that can carry an entire hidden message inside apparently ordinary text.
Variation selectors
Both blocks, including the 240-codepoint supplement — another documented channel for concealed payloads.
Unusual spaces
Non-breaking, hair, thin, em, en, figure, and ideographic spaces, normalised back to ordinary spaces when you clean.
Control characters
C0 and C1 controls, excluding tab, newline, and carriage return — which are ordinary formatting and get left alone.
Questions people actually ask
- Does this detect AI-written text?
- No, and no tool can do it reliably. This finds invisible characters that some software inserts. It does not read the statistical watermarks that model providers embed, because no public detector for those exists yet — not here, and not anywhere.
- So what does finding hidden characters actually tell me?
- That the text passed through something that inserts them. That could be an AI assistant, but it could equally be a word processor, a CMS, a PDF export, a translation tool, or a copy-paste from a web page. It is a clue about tooling, not proof of authorship.
- What if it finds nothing?
- Then it found nothing — which tells you nothing at all about whether a person or a model wrote the text. Absence of invisible characters is the normal state of most writing.
- Can I remove the characters?
- Yes. Every scan produces a cleaned version you can copy with one click. Genuinely invisible characters are deleted and unusual spaces are normalised to ordinary ones, so the text reads identically but imports cleanly.
- Will removing them strip the Claude or Gemini watermark?
- No. Those watermarks live in word and token choice, not in stray characters, so deleting invisible characters does not touch them. Any tool promising to remove a model's real watermark by cleaning Unicode is selling you something it cannot deliver.
- Is my text uploaded anywhere?
- Never. The scan runs entirely in your browser. You can confirm it yourself: open your developer tools, watch the network tab, and paste some text — nothing is sent.
More tools
AI detector
Every provenance signal in one pass
Metadata detector
Full EXIF, XMP, and provenance dump
Compare two files
See which signals were stripped between versions
Platform survival test
See what editing and uploading will destroy
Pipeline audit
Find which step destroys your provenance
Audit report
Turn any result into a documented record
Batch audit
Verify many files at once, keep the record
Manuscript screening
AI disclosure, provenance, and whether they agree