Paste your AI-polished draft on the left. Get clean text on the right - with invisible watermark characters and AI typographic tells stripped out. Everything runs in your browser; nothing is uploaded.
Newer LLMs leave two kinds of footprints, and this tool targets both:
Invisible Unicode artifacts. Through 2025, ChatGPT models were observed
emitting hidden characters — the narrow no-break space (U+202F), zero-width space
(U+200B), em space (U+2003) and em dash (U+2014). Forensic analysis concluded these
are most likely training artifacts from professionally typeset text rather than
a deliberate watermark, but either way they flag text as machine-touched.
(Rumidocs,
analysis)
Statistical & stylometric tells. Modern detectors don't rely on hidden
characters alone — they weigh punctuation frequency, sentence burstiness and lexical
patterns, and treat homoglyph substitution as an adversarial signal
(Mikros, 2025; Kanwal et al., 2026). Normalizing curly quotes, em-dashes, exotic
spaces and lookalike letters reduces those surface tells.
What a cleaner can't fix. Token-level statistical watermarks and
embedding-based classifiers key on word choices, not characters — only genuine
rewriting changes those. This tool removes the mechanical fingerprints; your own voice
does the rest.