Artificial intelligence has made significant progress in reading documents, recognizing handwritten text and interpreting low-quality images. However, a new experimental typography project called Ghost Font is revealing an unexpected limitation in how many AI vision systems process visual information.
Created by designer Eric Lu, Ghost Font is an innovative visual illusion that conceals letters within thousands of moving dots. Instead of outlining characters with visible strokes, the project relies on motion to reveal hidden text. Dots forming the letters move in one direction, while the surrounding dots drift differently, enabling the human brain to identify words based solely on movement patterns.
The project highlights a key distinction between human and machine perception. While people can naturally combine subtle changes in motion over time to identify hidden objects, many current multimodal AI models tend to interpret videos as a collection of individual frames. Without clear edges or recognizable letter shapes, AI systems often struggle to accurately detect the concealed message.
To demonstrate the concept, users can create an animation featuring a hidden phrase, such as "HELLO HUMAN" or "TOM'S GUIDE," and upload it to AI platforms like ChatGPT, Claude or Gemini. They can then ask, "What does this animation say?" to compare how different AI models interpret the moving text. In many cases, the systems may misread the animation, identify unrelated content or confidently produce incorrect responses.
Despite exposing a current weakness, Ghost Font is not intended as a security or encryption tool. The creator emphasizes that the project is an exploration of perception rather than a method for protecting sensitive information.
Researchers note that with enough video frames, optical-flow analysis or advanced computer vision techniques, AI systems can potentially reconstruct the hidden message. Some developers have already demonstrated success after instructing AI models on how the illusion works or allowing them to analyze the animation frame by frame.
As AI vision technology continues to evolve, experts expect future models to become more capable of decoding motion-based illusions like Ghost Font.
The project also draws comparisons with traditional CAPTCHAs, which were designed to exploit differences between human and machine perception by presenting distorted text that people could read more easily than computers. Ghost Font updates this concept by replacing distorted characters with motion as the primary visual cue.
Rather than suggesting that AI is easily deceived, Ghost Font serves as an example of how human and machine vision still differ in important ways. As multimodal AI systems become increasingly sophisticated, projects like this offer valuable insight into the unique strengths—and current limitations—of artificial intelligence.
Read the original article: