Eric Lu Unveils Ghost Font, Hiding Text in 1,000s of Moving Dots to Trip AI Vision
Updated
Updated · Tom's Guide · Jul 13
Eric Lu Unveils Ghost Font, Hiding Text in 1,000s of Moving Dots to Trip AI Vision
3 articles · Updated · Tom's Guide · Jul 13
Summary
Ghost Font lets humans read hidden words from dot motion while many AI vision models miss them or latch onto decoy text, exposing a gap in how machines parse video.
Thousands of tiny dots form no visible letter outlines; instead, dots inside a word move in one direction and surrounding dots in another, making the message appear only while the animation plays.
Current multimodal systems often treat video as a fast sequence of still frames, so removing stable edges and high-contrast shapes undermines the OCR-style cues they usually rely on.
Eric Lu presents the project as a perception experiment, not security: frame-by-frame analysis, optical-flow methods or better prompting can already help some models recover the text.
The project echoes older CAPTCHAs by exploiting a human-machine perception gap, this time using motion rather than distortion as AI video understanding improves.