Breaking

Developer Unveils 'Ghost Font' That Humans Can Read in Motion but Fools AI Vision Models

July 13, 2026 at 13:58 EDT

Foundation Models
Security

A developer known as Eric Lu has created an experimental typeface called "Ghost Font" that human viewers can decipher only when it moves, while leading AI vision models reportedly fail to read it correctly — in one case, ChatGPT is said to have spent 19 minutes analyzing the text before hallucinating the wrong message.

Machine Vision · Experimental Typeface

"Ghost Font": Text Humans Read in Motion — But AI Cannot

A prototype typeface dissolves into dot-noise in any single frame, revealing its message only when played as video. Leading AI vision models reportedly failed to decode it — in one case confidently hallucinating the wrong text.

19 min

One model analyzed the text — then produced a false message

Tested AI models that decoded the intended message correctly

2013

The year of ZXX, the anti-OCR ancestor Ghost Font builds on

Who can read a single still frame?

The message lives across frames, not inside any one image. Stopped in place, both eyes and machines see only noise.

Humans (in motion)

Reconstruct letters via optical flow

✕ noise

AI vision (frame-by-frame)

Sees dot-noise, hallucinates a false reading

Why it works

Message split across many frames

→

Any single frame = random dots + decoy content

→

AI reads frames in isolation, gets steered wrong

Where ZXX cluttered static letters, Ghost Font shifts the burden to time — targeting pipelines that lack robust temporal, motion-based context.

Enthusiasts say

A clever demonstration of a genuine blind spot in machine vision — with uses as a human-vs-bot CAPTCHA or a probe for AI benchmarks.

Skeptics counter

The trick leans on current limits in video understanding. Stronger optical-flow reasoning could let future models read it — as AI eventually beat ZXX.

Continue reading

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