invisible-text-forensicslisted

# Invisible Text Forensics ## Overview Every "de-AI" and humanizer pass rewrites *style*. None of them read the *bytes*. Text carries characters that render as nothing: zero-width spaces, bidirectional overrides, tag characters, variation selectors, private-use code points. They survive copy, paste, email, PDF extraction and git. They fingerprint documents, smuggle instructions into prompts, and make source code render differently from how it executes. **Core principle: judge text by its code points, not by how it looks.** ## When to use - Before publishing, sending or signing text that came from an AI, a browser, a PDF or an unknown source - After a humanizer or de-AI pass, which handles style and leaves the byte layer untouched - When a string comparison, diff, regex, deduplication or database lookup fails on text that looks identical - When auditing a prompt, a RAG document or a pull request for injected instructions - When a confidential document leaks and you need to know which copy it was - When reviewing source code from an untrusted contributor (Trojan Source, CVE-2021-42574) **Not for:** rewriting tone or removing AI writing tells at the sentence level. That is a separate job; run a humanizer for that, then run this. ## Quick reference | Risk | Characters | Why it matters | |---|---|---| | `critical` | Tag characters `U+E0000–E007F`, bidi overrides `U+202A–202E` `U+2066–2069`, Private Use Area | Hidden ASCII payloads, prompt injection, code that renders diffe