The AI boom is currently an upward-sloping slope that people are trying to climb with all their might. However, there exist some who on the contrary are building metaphorical bunkers to save themselves from this looming doom. Among them is Eric Lu, an American designer and developer who has begun a new experimental project called Ghost Font attempting to expose a blind spot in how today’s AI systems interpret visual information.In a time when AI models can pretty much reach even the incoherent thought, easily processing handwritten notes, scanned documents and even blurry screenshots, he has created a font that only humans can read. In a post that has gained massive traction on X, Lu claimed he tested the project with advanced AI systems from OpenAI and Anthro“I created a font called Ghost Font that only humans can read. Tested it in Fable and GPT 5.6 Sol Ultra and neither was able to decipher it correctly,” he wrote in the post that has amassed 17 million views.
What is Ghost Font?
Ghost Font is not a conventional typeface. Instead of displaying visible letters, it relies on a motion-based optical illusion. Animated videos generated with the system contain thousands of tiny dots. The dots that make up each hidden letter move in one direction, while the surrounding dots travel in the opposite direction.As the animation plays, the human mind naturally groups the moving dots together, making the hidden word appear even though no visible letter outlines exist. Once the video is paused, the effect disappears, leaving behind what looks like random visual noise.The project also includes decoy text intended to mislead AI models. While people can continue to identify the message through the motion, AI systems that analyse videos on a frame-by-frame basis can detect the false text and confidently return an incorrect answer.In a blog post, Lu described Ghost Font as an experiment in “anti-AI typography.” He said the project explores whether written communication can be presented in a form that remains understandable to people but complex for AI to interpret. He claimed that GPT-Sol 5.6 Ultra and Claude Fable failed to identify the message. Another AI model spent nearly 20 minutes analysing the video before generating a message that did not exist.
AI tests and future uses
Claude Fable reasoning for failing to decode Ghost font
To evaluate the concept, Lu tested videos generated with Ghost Font using advanced AI models from OpenAI and Anthropic. According to him, the systems were unable to correctly identify the hidden message unless they were first taught how the illusion worked, highlighting differences between human visual perception and current AI models’ processing.The creator also acknowledged that the Ghost Font is not impossible for AI to decode. He said models could eventually read the hidden text if they were instructed to ignore the decoy message or use a more advanced frame analysis technique. He also believes that the font could have practical applications beyond demonstrating AI limitations.He shared that the technology could make CAPTCHA systems more resistant to automated bots by using motion-based text that remains readable for people. It could also strengthen bot detection by adding a verification layer that current multimodal AI systems find difficult to bypass.He further added that the approach could improve digital watermarking and DRM by embedding copyright markers that are harder for AI tools to detect or remove. He claimed sensitive documents could include motion-based text or markers that remain visible to people while offering an additional layer of protection against automated scraping and data extraction.
Social media reaction
The project has gained mixed reactions online. While some users argued the current limitations reflect how AI models process video rather than a fundamental weakness, others said future AI systems could be trained to overcome the hurdle.“The only downside is that it drives all who look upon it INSANE,” a user joked.“Today. But AI could be trained on it. Shelf life of your font is 3 months,” another claimed.“It’s probably just an artifact of the way they handle video rn with screenshots at a low frame rate. Nothing to do with the actual intelligence of the models,” a user commented.
