Ghost font hides messages in motion – and AI models can't read them

Stare at a screen full of tiny dots. Same color, same size, same apparent randomness. Then the video plays. Within two seconds, a word materializes from the chaos, clear as a billboard. Pause the clip. The word vanishes. That’s Ghost Font, an experimental typeface built by designer Eric Lu at San Francisco-based Mixfont. It encodes text purely in motion - and multiple leading AI models, including versions of ChatGPT and Claude tested by Lu, have repeatedly failed to read it.

The project isn’t just a party trick. It exposes a genuine blind spot in how AI processes visual information.

Motion Is the Message

Ghost Font doesn’t install on your computer - it generates short looping videos where letters exist only as movement.

Here’s the mechanism. Thousands of dots fill each frame. The dots forming a letter move in one direction - say, upward. Background dots move the opposite way. Because every dot shares the same color and size, a single frozen frame looks like pure static. No edges. No curves. Nothing for OCR to grab.

Your brain, though, does something AI still struggles with. It applies what psychologists call the “common fate” principle - automatically grouping things that move together. Like picking out a friend in a crowd because they’re the only one walking toward you. The result: you read the word almost instantly.

“Ghost Font uses motion, noise, and decoys to hide messages in videos readable by humans but nearly impossible for AI models like ChatGPT or Claude to decipher.” - Eric Lu, Mixfont founder

What makes it tick:

• Letters exist only as motion patterns - no static shapes anywhere in the video
• A paused frame is pure noise; no edges or curves for OCR to grab
• Decoy text deliberately misleads AI toward false readings
• Human persistence-of-vision fuses frames into letters; AI models typically analyze frame-by-frame - and Mixfont’s free web playground lets anyone test this gap directly

Trying to read Ghost Font by analyzing individual frames is like understanding a flipbook by staring at page seven. Models fed these clips either detected nothing, confidently reported the decoy text, or hallucinated words that were never encoded. These kinds of failures are among the most frustrating computer problems users and researchers encounter today.

Clever Illusion, Not a Silver Bullet

Lu himself acknowledges that a system built to track dot trajectories across frames could crack it.

Ghost Font is not AI-proof. Any system computing optical flow across frames could, in principle, reconstruct the letters. The text is also harder for humans to read than standard typography - you might need a beat to lock in. This is a current-generation obstacle and a research tool, not a permanent privacy solution. As temporal modeling in AI improves, the gap Ghost Font exploits will shrink.

Still, it went viral for a reason beyond novelty. Ghost Font became something like protest art for the scraping era - proof that human perception still holds territory AI hasn’t fully mapped. That territory won’t last forever, but right now it’s real, and the race to close it is already underway. The broader implications for digital privacy echo concerns raised by every new surveillance app built to track people without their knowledge.