Particle.news

Ghost Font: Motion-Based Type That So Far Foils AI

The method exposes a short-term gap in how many multimodal AI systems process video by analysing frames instead of motion.

Overview

  • Creator Eric Lu posted Ghost Font in mid-July and the demonstration went viral Wednesday, drawing reports of roughly 18 million views in a few days.
  • Ghost Font hides letters inside hundreds of moving dots so humans perceive text through motion grouping while a single frozen frame removes the readable message.
  • Informal tests reported by Lu and independent journalists found several major multimodal models, including GPT‑5.6 Sol Ultra and Anthropic’s Claude Fable, failed to read the hidden text and instead reported a planted static decoy.
  • Lu offers a web generator for short demo clips but says Ghost Font is experimental with no plans for productization and is meant as a proof of concept and benchmarking idea.
  • Researchers and reporters warn the protection is provisional because motion-aware tools such as optical-flow analysis could let AI recover the text and erase this gap, though the technique may still inform CAPTCHAs or AI-readiness tests.