Ghost Font is an experimental “anti-AI” visual communication system (not a traditional TTF typeface) created by computer scientist and Mixfont founder Eric Lu. It encodes short messages in motion so that humans can read them while current leading AI models often cannot.
How It Works
-
- Messages appear as thousands of tiny moving dots in a short video.
- The dots forming the letters typically drift in one direction (e.g., upward), while background dots move the opposite way.
- Human vision readily groups points moving together into coherent shapes (a form of motion-based perceptual grouping), so the text becomes visible as faint letters.
- Any single static frame or screenshot looks like random noise—the letters vanish because they rely on relative motion across frames.
- A hidden decoy message is embedded as an additional layer. AI systems that analyze the video (usually by breaking it into individual frames) frequently latch onto the decoy and report the wrong text, or hallucinate something else.
You can try it yourself on the official page (mixfont.com/ghost-font): type a short message, preview the animated result, and download the video. Processing happens locally in the browser.
Background and Motivation
Lu was inspired by frustration with AI assistants automatically summarizing private conversations (sometimes incorrectly). He wanted a way to share messages that are harder for AI to casually read. The project draws a parallel to the 2013 ZXX font by Sang Mun, which used visual noise to defeat older OCR systems — but modern multimodal models easily read ZXX. Ghost Font targets current AI limitations around temporal/motion understanding instead.
According to Lu: “I was inspired to make Ghost Font one day after texting with my mom, and I noticed that AI kept trying to summarize our messages, many times incorrectly,” Lu, a former Google associate product manager, said in an interview. “I was annoyed that it was being so intrusive in our private conversation and I started brainstorming ways that I could potentially communicate without AI reading everything.”
Effectiveness and Limitations
In Lu’s tests, models such as Claude Fable and GPT variants (including strong “thinking” or agentic versions) struggled or failed to extract the real message without very specific prompting about the technique. Some users later succeeded with careful instructions or by writing code to analyze optical flow/frame differences, showing it is not cryptographically secure.
Key caveats:
-
- Limited to very short messages.
- Harder for some humans to read than normal text (and inaccessible to many with visual impairments).
- A determined agent with code execution can often decode the motion.
- Future video-native models that better process temporal cues will likely defeat it.
- True privacy still requires proper encryption or shared secrets, not visual tricks.
Lu presents it as a prototype and experiment exploring differences in human vs. AI perception, potential CAPTCHA ideas, and a benchmark for AI visual/temporal capabilities. He has indicated plans to open-source the generation code and support longer text.
In short, Ghost Font is a clever, shareable proof-of-concept that leverages human motion perception to make short messages temporarily harder for today’s AI vision systems to parse — more of a demonstration of current gaps than a robust anti-surveillance tool.
