Ghost Font hides text in motion so humans can read it but AI cannot

A font that humans can read and AI cannot now exists. It is called Ghost Font, and it exploits the one thing your visual system does better than a frame-by-frame machine: seeing motion.

How it works

Ghost Font, built by engineer Eric Lu, encodes words in moving dots. The dots are rendered in the same color as the background, so any single frame looks like pure noise. Your visual system evolved to detect motion, so when the video plays, the letters simply appear. Pause it, screenshot it, or feed it to an AI model frame by frame, and the message vanishes, according to mixfont.com and discussion on Hacker News.

There is an extra layer of misdirection built in. Each video also embeds a decoy message, so an AI hunting for hidden content is likely to find the fake one and stop looking.

How the models did

Lu tested Ghost Font against frontier models, including Claude Fable and GPT-5.6 Sol Ultra. They struggled to decode the message until told exactly what technique to use. Even Grok conceded on X: “Ghost Font wins this round.”

That caveat matters. Once a model knows the trick, the protection weakens, which is exactly why Lu frames this as a research experiment rather than a permanent shield. Smarter models will eventually catch up.

Why it still matters

Temporary or not, Ghost Font points at something real. If information can be encoded so that human perception is the decryption key, that opens doors from AI-resistant CAPTCHAs to private messaging that only humans can read.

It also lands at a moment when the web is filling with automated readers, scrapers and agents. A channel that filters for human eyes, even briefly, is a genuinely new primitive, and the fact that it took moving dots in background colors to get there says a lot about where the human-machine boundary currently sits.

Would you use a human-only font while it still works?