“Ghost font” animation readable by humans but hard for top AI models

Researchers describe a “ghost font” display technique designed so that people can read the text while leading AI systems struggle. According to the report, the method relies on a combination of motion, visual noise, and decoy patterns. These elements work together so that the intended message is not reliably recoverable when the animation is paused, effectively causing the text to disappear from view at that moment. The researchers also frame the approach as a challenge to current AI text-reading or visual recognition pipelines, noting that top models remain stumped under the technique’s conditions. While human viewers can interpret the moving, noisy presentation as legible characters, automated systems appear unable to consistently distinguish the real message from the decoys and interference. The article characterizes the work as a proof-of-concept for exploiting differences between human perception and how AI models process motion, noise, and ambiguous visual cues.