| Tag | Codepoint | Name | Category | Count | First at |
|---|
Your analyzed text above is Text A. Paste a second version below as Text B to see which carries more watermark evidence — e.g. whether an edit is cleaner, or which of two drafts is AI-processed.
The narrow no-break space documented as a systematic ChatGPT (GPT-o3/o4-mini) insertion — invisible, but a machine-readable fingerprint.
Rumi, Apr 2025Anthropic began watermarking all Claude text output for EU AI Act compliance — applied globally, undetectable by any third party.
EU AI Act, Art. 50False-flag rate of AI detectors on non-native-English essays — why several universities disabled them and why evidence beats a percentage.
Stanford-lineage studyOf newly published web articles are now AI-generated — publishing more no longer sets you apart; verifiable quality does.
Graphite, 2025Triage a slush pile in bulk. Flag pasted-from-ChatGPT submissions by their copy artifacts before a manuscript wastes staff time — with confidentiality, because nothing is uploaded.
Scan writer deliverables for watermarks and clean them before publishing, so hidden characters don't leak into your CMS or trip a client's detector.
Get per-sentence evidence you can actually show — a zero-width space at character 132, not an unfalsifiable percentage. With an explicit "inconclusive" verdict when the sample is too short.
Self-check before submitting, see which lines trip detectors, and export an evidence report to pair with your document version history if you're wrongly accused.
Detect hidden zero-width payloads used to trace leaks, and audit documents for covert steganographic channels — entirely offline.
Inspect user-submitted text for invisible-character injection, homoglyph spoofing and Unicode obfuscation before it hits your database or renders in your UI.
Strip AI-tool artifacts (curly quotes, em-dashes, NBSPs, markdown) from work drafted with assistance, so your final copy is clean and portable.
ChatGPT, Claude and Gemini output carries invisible baggage. One click shows it; one click removes it.
Every character is checked against 150+ known invisible codepoints, exotic spaces, Cyrillic/Greek homoglyphs and ChatGPT copy artifacts — and any zero-width payload is decoded. These are exact, reproducible facts.
Sentence-length burstiness, lexical diversity, AI-vocabulary density and repetitive openers are measured, then matched against five models' known style fingerprints. Clearly labeled as probabilistic.
You get a tiered, sentence-level report you can export — and a configurable remover that strips the character-level watermarks while telling you honestly what cleaning can and can't do.
Zero-width spaces & joiners, word joiners, BOM, soft hyphens, bidi controls, Hangul fillers, tag characters.
Non-breaking, narrow no-break (the ChatGPT tell), en/em, thin, hair, figure and ideographic spaces.
Cyrillic and Greek lookalike letters substituted into Latin text — a deliberate evasion / watermark trick.
Em/en dashes, curly quotes and ellipsis characters — weak on their own, telling in combination.
Phrases that only survive an un-edited copy-paste from an AI assistant.
ChatGPT's data-start/data-end HTML attributes and raw markdown left from pasting.
Burstiness, lexical diversity, AI-vocabulary rate, repetitive openers and em-dash density.
Style signatures for the major assistants, scored side by side.
| Tool | Free tier | Signup | Character forensics | Hidden-msg decode | Model attribution | Evidence shown | Private (client-side) | Sells humanizer |
|---|---|---|---|---|---|---|---|---|
| This tool | Unlimited | No | ✓ 150+ codepoints + HTML artifacts | ✓ zero-width steganography | ✓ all 5, evidence-listed | ✓ exact chars & positions | ✓ 100% | No |
| GPTZero | 10k words/mo | For dashboard | ~ flags "obfuscation" | ✗ | ✗ | ~ sentence highlights | ✗ retains text | No |
| Originality.ai | None meaningful | Yes + payment | ~ separate tool | ✗ | ✗ | ~ per-sentence | ✗ stores scans | No |
| Copyleaks | Limited | Yes | ✗ | ✗ | ✗ | ~ "AI Logic" | ✗ trains on your text | No |
| ZeroGPT | ~15k chars, ads | No | ✗ | ✗ | ✗ | ✗ bare % | ✗ | ~ paraphraser |
| Turnitin | Institution-only | Institution | ~ manipulation flags | ✗ | ✗ | ~ segments | ✗ retains papers | No |
| Pangram | 2k words/day | For more | ✗ | ✗ | ✗ | ~ sentence-level | ✗ | No |
| Unicode cleaners | Free | No | ~ strip & count | ✗ rarely | ✗ | ✗ no explanation | ✓ mostly | ~ upsell funnels |
Removing watermarks makes a page tidy — but will Google rank it and will AI quote it? ArticleGrade grades any URL across six quality dimensions in ~20 seconds, free.
A hidden signal marking text as machine-generated. Two kinds exist: character-level watermarks — invisible Unicode characters (zero-width spaces, special spaces, lookalike letters) inserted between normal letters — and statistical watermarks (Google SynthID-Text, Anthropic's Claude watermark since Aug 2026), which bias the model's word choices in a secret, key-verifiable pattern. This tool detects the first kind completely and flags the stylistic footprint that accompanies AI text generally.
Zero-width steganography hides binary data in invisible characters — for example ZWSP for a 0 bit and ZWNJ for a 1 bit. It's used for invisible document watermarking, tracing who leaked a file, and covert messaging. The decoder extracts every zero-width character in order, tries the common binary encodings, groups the bits into bytes, and shows any readable text it recovers. If a payload is present but not in a recognized text encoding, it tells you that too — all locally.
In April 2025, researchers documented GPT-o3/o4-mini systematically inserting the narrow no-break space (U+202F) into longer outputs. OpenAI called it a training quirk rather than a watermark and removed it within weeks — but text generated in that window carries it, and AI-pipeline tools still leave NBSP, zero-width and other artifacts. OpenAI has also built a true statistical watermarker (reported ~99.9% accurate) but has never released it or its detector.
No third-party tool can — including this one, and including every site claiming to "remove Claude's watermark" by stripping Unicode (that removes nothing: Anthropic's watermark, applied to Claude output since 2 August 2026 under the EU AI Act, lives in word-choice statistics, not characters). Detection requires the provider's key. Google's SynthID Detector only verifies Google-generated content; Anthropic says a public detector will come later.
No — and any tool implying it does is misleading you. Stripping characters removes character-level watermarks and copy artifacts, which matters for clean documents and privacy. But statistical detectors (GPTZero, Originality, Turnitin) analyze word-choice patterns, which survive character cleaning entirely. We say this plainly because most "watermark remover" sites won't.
Neither, on its own. Deterministic findings (invisible characters, chat artifacts) are strong evidence of machine processing but not proof of authorship; a clean result doesn't prove human authorship either. If you're accused: this tool's evidence report plus document version history (Google Docs, Word) is a far stronger defense than any detector percentage. Detection scores alone should never decide an accusation — several universities have disabled AI detectors over false positives.
When you paste or upload text, no — the analysis, decoding and cleaning all happen in your browser's memory with zero network requests (verifiable in your dev tools); reload and everything is gone. The one exception is the optional "check by URL" box: to fetch a page the browser can't reach directly, that public URL is sent to ArticleGrade's server, which retrieves the page and returns its text — then the watermark analysis still runs locally on your device. Either way, we never store your text or train on it, unlike mainstream detectors (at least one trains on your submissions by default).