Ghost Font: Using Motion-Based Obfuscation to Combat AI OCR

Ghost Font uses motion-based illusions to bypass AI perception

Ghost Font is an experimental project designed to create visual messages that are readable by the human eye but difficult for AI models to decipher. Unlike traditional fonts, Ghost Font generates video clips where text is formed by the motion of dots. When the video is paused, the dots blend into the background, leaving a static image that reveals no readable text. This approach aims to move beyond static obfuscation—like the 2013 ZXX font—which modern AI agents can now easily read using advanced optical character recognition (OCR).

Core mechanism: Motion, Noise, and Decoys

Ghost Font relies on a layered defense strategy to prevent automated decoding:

• Temporal Rendering: Letters are not static glyphs but are created through the movement of dots. This ensures that a single screenshot or frame cannot reveal the message, as the dots are designed to look identical to the background in any static state.
• Noise and Decoys: To prevent dedicated agents from using local code execution to analyze motion patterns, Ghost Font includes a "decoy message" in every video. The goal is to trick an AI into identifying the decoy as the primary message, thereby masking the actual content.
• Local Processing: The current prototype operates locally, meaning text is rendered in the browser and not sent to a server, preserving user privacy during the experiment.

Technical critiques and AI vulnerabilities

While the project aims to challenge AI perception, technical analysis from the community suggests that motion-based obfuscation is susceptible to several algorithmic attacks:

Temporal Averaging and Optical Flow

Critics argue that the "anti-AI" nature of the font is temporary. Users have reported that advanced models, such as GPT-5.6 Sol, can decode the messages by using temporal analysis, optical flow, and vertical-displacement maps to estimate image movement and create a high-contrast motion map of the letters.

"I’ve confirmed there’s no readable static OCR layer; I’m decoding its optical-flow field so the letter shapes become explicit."

Frame Differencing

Some developers have demonstrated that the message can be recovered using simple frame subtraction. By taking two consecutive frames and shifting them to minimize differences, the underlying text can be isolated and passed to a standard OCR engine. One user claimed to have implemented this recovery process in approximately 20 lines of code.

The "Cat-and-Mouse" Game of CAPTCHAs

Because the system relies on a predictable rendering method, it mirrors the history of CAPTCHA systems. Community feedback suggests that once the technique for decoding motion-defined illusions is publicized, AI models can be prompted or trained to apply those specific techniques, rendering the obfuscation ineffective.

Practical limitations and human legibility

Ghost Font faces a significant trade-off between AI resistance and human usability. Several users noted that the text is difficult for humans to read, often comparing the experience to "Magic Eye" 3D pictures or noting that it causes eye strain on mobile screens. This suggests that as the signal-to-noise ratio is decreased to further confuse AI, the text becomes equally illegible to humans.

Future directions for visual obfuscation

The creator of Ghost Font intends to open-source the video generation code and expand the system to handle longer text strings. Potential future applications include:

• Next-Gen CAPTCHAs: Integrating motion into CAPTCHA systems to make them more difficult for bots to solve than static distorted text.
• AI Benchmarking: Using these motion-based illusions to test the progress of multimodal models and their ability to process video-native data rather than just sequences of static frames.