Decoy Font Uses Optical Illusions to Hide Text From AI Readers

Mixfont’s Decoy Font is a downloadable TTF typeface built to frustrate machine reading. Each glyph layers two messages at different spatial frequencies: sharp, high-frequency outlines in the foreground and a blurred, low-frequency mass behind it. Up close a viewer (or an AI ingesting pixels) sees the foreground letters, but stepping back or squinting reveals the intended text. The trick borrows directly from hybrid-image research — the classic Einstein/Monroe illusion — and applies it to typography so the words a human is meant to read differ from what an OCR pipeline or vision-language model latches onto.

The creator claims screenshots of the font mislead systems like ChatGPT, Gemini, and other frontier models, since they tend to read the crisply outlined foreground rather than the low-frequency hidden layer. Because it ships as a real installable font rather than an animation (unlike Mixfont’s companion Ghost Font, which hides text in motion), it’s easy to drop into ordinary documents and messages. The author floats uses ranging from CAPTCHAs to private notes between friends, and suggests character-based scripts like Chinese could hide messages even more effectively.

The pitch is honest about its ceiling: this is a deterrent, not real protection. Agentic models with coding tools or a simple nudge to “look for the hidden letters” can defeat it, and the author expects smarter models to decode both layers over time — even proposing the font as a benchmark for text-recognition capability. It fits a broader wave of “anti-AI” typography aimed at keeping human-intended content out of automated scrapers, useful mainly as casual obfuscation rather than a security control.