Ghost Font Uses Optical Illusions to Stump AI Systems Like Claude and GPT-5.6
Key Takeaways
- ▸Ghost Font successfully hides text from Claude Fable and GPT-5.6 Sol Ultra using optical illusions in moving video animations
- ▸The technique exploits how AI systems process videos frame-by-frame, unlike humans who perceive fluid motion
- ▸Ghost Font demonstrates that current multimodal AI has significant limitations in parsing complex visual concepts
Summary
Designer Eric Lu has created Ghost Font, a proof-of-concept tool that uses optical illusions to hide text from AI systems. The technique overlays text within hundreds of moving dots in a video animation—the dots composing the words move upward while others move downward, creating a visual opposition that the human brain can interpret but that AI systems struggle to parse. When tested on Anthropic's Claude Fable and OpenAI's GPT-5.6 Sol Ultra, both systems failed to find the hidden message. The font also includes a static decoy message that further confuses AI, demonstrating a clever workaround to the current lack of regulations protecting content from unauthorized AI training.
While Lu acknowledges that Ghost Font isn't a cure-all for AI-related privacy concerns, it effectively highlights a significant limitation in how multimodal AI systems currently process visual information. Unlike humans who perceive motion as a fluid pattern, these AI systems still examine videos frame-by-frame, much like flipping through a flipbook. This architectural difference makes it nearly impossible for current AI to recognize text hidden across moving frames. The technique is limited to extremely short messages and could become obsolete as AI systems evolve to process videos as continuous optical flows rather than discrete frames.
- The tool highlights the gap between human and AI visual perception while addressing concerns about unauthorized AI training on internet content
Editorial Opinion
Ghost Font is an elegant proof-of-concept that shines a spotlight on the architectural limitations of current multimodal AI—namely, their frame-by-frame video processing versus human fluid motion perception. While the technique is limited in scope and may become obsolete as AI improves, it's a valuable reminder that human creativity and ingenuity can still outpace current AI capabilities. However, it also underscores the critical need for legislative protections and ethical AI training practices, as individual workarounds cannot substitute for systemic policy changes.
