Breaking

'Ghost Font' Exploits AI's Weak Motion Perception, Fools ChatGPT Into Hallucinating a Message

July 13, 2026 at 22:51 EDT

Foundation Models
Security

A developer named Eric Lu has released an experimental typeface called "Ghost Font" that humans can read through motion but current AI vision systems consistently misread, in one test causing ChatGPT to spend 19 minutes analyzing a clip before hallucinating a message that was never there.

Adversarial Design · Machine Perception

The Font Humans Can Read — but AI Cannot

"Ghost Font" hides words inside motion: static frames look like random dots, but as the video plays the human eye stitches the drifting pattern into text. Vision models, reading frame by frame, stay blind — and can be trapped by a built-in decoy message.

19 min

ChatGPT 5.5 Pro spent this long analyzing a clip — then reported a message that was never there

Letters visible in any single still frame — the message exists only in movement

2013

The ZXX ancestor — a static anti-OCR font modern AI now reads with ease

How it exploits the perception gap

Static shape vs. motion — where each side wins

tracks flow

reads it

Human eye

stitches motion into words

sees dots

AI vision

frame-by-frame, no optical flow

Background scrolls up, letter-dots drift down. The information lives between the frames — exactly where today's models don't look.

The decoy trap — flow of failure

Model reads the clip frame by frame

→

Locks onto the embedded decoy message

→

Returns the wrong text — or hallucinates one that isn't there

The point it lands

A vivid demo of where human biology still beats machine perception — with potential uses in CAPTCHA-style verification, secure documents and human-only channels.

The honest limits

Some models can be coaxed into decoding it with the right prompting, and the animations are hard for people to read too. A proof of concept — not a hardened tool.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.