- Ghost Font is a human-readable font designed to confuse leading AI text recognition models
- It uses motion, visual noise, and decoy patterns to hide messages in animated dot patterns
- AI models analyze frames individually, missing motion cues that humans use to read the text
AI has become remarkably good at reading documents, deciphering messy handwriting and understanding blurry screenshots. But an experimental project called Ghost Font is highlighting an unexpected weakness in most advanced AI systems of today. Created by developer and designer Eric Lu, Ghost Font is designed to be readable by humans while remaining difficult for leading AI models to interpret. According to Lu, top models from OpenAI and Anthropic struggled to decode the hidden text accurately.
"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," Lu wrote in a post on X that has amassed more than 17 million views.
See the post here:
How does it work?
The project uses a combination of motion, visual noise and decoy patterns, making the message disappear the moment the animation is paused. Unlike traditional fonts that rely on visible letter shapes, Ghost Font creates an optical illusion. Thousands of tiny dots fill the screen, with the dots inside each hidden letter moving in one direction while the surrounding dots drift in another.
The human brain naturally groups these moving pixels together, allowing the hidden word to emerge even though there are no visible outlines. When the animation is paused, the illusion vanishes. What remains is a screen filled with what appears to be random visual noise.
Ghost Font also introduces another layer of deception by embedding decoy text designed to confuse AI systems. While humans continue to perceive the intended message through motion, AI models can be misled into confidently identifying the wrong text because they often analyse videos frame by frame instead of processing motion the way humans do.
Explaining the project in a blog post, Lu described Ghost Font as an experiment in "anti-AI typography".
"Ghost Font is an experiment of a way to graphically communicate in writing in a format that AI cannot easily understand. While it's not as legible as regular text, the letters are still immediately readable to a human eye, but even leading AI models can't decipher it easily," he said.
To test the concept, videos created with Ghost Font were shown to leading AI systems, including advanced models from OpenAI and Anthropic. According to Lu, the models struggled to identify the hidden message until they were explicitly told the technique behind the illusion, demonstrating a gap between human visual perception and the way current AI systems process moving images.
Why Ghost Font Matters
For developers and security teams, Ghost Font offers a new way to protect information from automated systems while still being human-readable.
- CAPTCHAs could become more effective: Instead of relying on puzzles that many bots can now solve, systems could use motion-based text that humans can read but AI struggles to interpret.
- Adds a new layer to bot detection: Websites and apps could incorporate motion-based verification that current multimodal AI models find difficult to bypass.
- Could strengthen watermarking and DRM: Digital watermarks or copyright markers embedded using Ghost Font-style encoding may be harder for AI tools to detect or remove.
- Human-only information in secure documents: Sensitive files could include motion-based text or markers that remain visible to people while staying inaccessible to today's AI systems, adding an extra layer of protection against automated scraping and data extraction.
