Ghost Font: Text That Humans Can Read But AI Models Cannot
Key Takeaways
- ▸Ghost Font hides text using motion encoding, noise, and decoys—readable to humans but not to leading AI models
- ▸Advanced AI models including Claude Fable and GPT 5.6 Ultra cannot easily decode Ghost Font from screenshots or single frames
- ▸The tool represents evolution beyond the ZXX font (2013), which modern AI systems can now read despite original anti-OCR design
Summary
Ghost Font is an experimental tool that encodes text messages in videos using motion, noise, and decoys—making them readable to humans but imperceptible to leading AI models. The project tests messages against advanced systems including Claude Fable, GPT 5.6 Ultra, and ChatGPT 5.5, demonstrating that even state-of-the-art vision models struggle to decode the hidden text. Ghost Font improves upon earlier anti-AI typography concepts like the ZXX font (2013), which modern AI can now easily read. Messages are embedded using animated dots against matching backgrounds; when frames are frozen, the text disappears entirely, becoming invisible to screenshot analysis. The tool incorporates a decoy message layer, further confusing AI systems attempting to decode the real content.
The project was created to explore whether visual messages can be preserved as human-exclusive communication in an increasingly AI-powered world. While the tool successfully evades current online AI analysis capabilities, the creators acknowledge that dedicated local analysis with code execution could eventually break the encoding. The project demonstrates the ongoing technological arms race between adversarial design and AI perception, while also suggesting that cryptography—not visual tricks—remains the only truly reliable method for hiding information from advanced AI systems.
- Each video includes a decoy message to mislead AI analysis attempts
- The project suggests visual obfuscation is a temporary solution; true privacy requires encryption rather than graphical tricks
Editorial Opinion
Ghost Font is an ingenious experiment in adversarial design, but also a humbling reminder of how quickly AI capabilities outpace defensive techniques. While the tool cleverly exploits current limitations in AI video analysis and reasoning, the project itself validates what security researchers have long known: visual tricks are theater against determined AI systems. ZXX seemed cutting-edge in 2013; it's trivial now. Ghost Font may enjoy a similar lifespan. The real value is philosophical—this project documents the frontier of human-AI perception, showing us both where current AI falls short and where investment in improved vision capabilities should focus.
