Ghost Font hides messages in motion that humans read but AI can't

Ghost Font is an experimental “anti-AI” typeface that encodes text in the movement of dots across a video rather than in static letterforms. To a human watching the clip, the words are immediately legible because the eye tracks the motion of the dots. Pause the video or take a screenshot, and the dots blend into the background noise, leaving a single frame with no discernible message. Because current multimodal models analyze video by sampling individual frames, they see only static noise—one frontier model reportedly spent 19 minutes and then hallucinated a message that was never there. The project runs entirely in the browser with no data sent to a server.

The idea updates Sang Mun’s 2013 ZXX font, which camouflaged letters to defeat OCR but is now trivially readable by modern AI. Ghost Font’s creator, Eric Lu, notes the video approach has a clear weakness: an agent with local code execution could analyze the dots’ motion over time and recover the text. To counter that, every generated clip also embeds a decoy message designed to mislead a determined agent into decoding the wrong string. Lu is candid that genuine secrecy still requires real encryption or a human-held key; Ghost Font is a perceptual trick, not a cryptographic one.

The significance is less about practical secure messaging and more as a probe of where multimodal perception currently breaks. Lu floats motion-based CAPTCHAs—harder for bots but still easy for people—and using Ghost Font as a benchmark that video-native models will eventually beat. His broader point cuts both ways: the font is genuinely hard for AI to read, but it’s also hard for humans, underscoring how quickly the perceptual gap is closing. He plans to open-source the video-generation code and extend it to longer text.