Designer Eric Lu created a "ghost font" that encodes text as moving dots, which humans can read but AI visual models cannot [1, 2].
This development highlights a fundamental gap between how human visual perception processes information and how AI image-processing systems operate. As artificial intelligence increasingly manages digital content, the ability to create "invisible" text for machines suggests new ways to protect data or test the limits of machine learning [1, 2].
The ghost font functions by presenting text through a series of moving dots. While the human eye can easily synthesize these movements into legible characters, most AI visual models struggle to interpret the patterns [1, 2]. This discrepancy exists because AI models often analyze static frames or specific pixel clusters rather than perceiving fluid motion in the same way humans do [2].
Lu designed the system to serve as a practical test for advanced AI systems, including ChatGPT [1, 2]. By attempting to read the ghost font, developers can identify where visual models fail to recognize organic patterns or motion-based information [1].
The project serves as a demonstration of the current limitations of computer vision. While AI has made significant leaps in recognizing static objects and text, the ghost font reveals that the technology still lacks the intuitive visual synthesis inherent to human sight [2].
“Humans can read but AI visual models cannot decipher.”
The emergence of the ghost font suggests a burgeoning field of 'adversarial design,' where humans create visual barriers to prevent AI from scraping or analyzing specific data. If AI cannot perceive motion-based text, it indicates that current visual models are heavily reliant on static pattern recognition rather than true visual understanding, potentially opening a new avenue for digital privacy and security.
