MixFont has developed a font called Ghost Font that is designed to be readable by humans but unreadable by AI OCR systems, potentially as a tool against AI scraping.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
Motion as inscription, decoy as barrier: a font experiment that relies on human dynamic perception while deliberately resisting the visual logic of multimodal models, exposing the current boundaries of AI perception and hinting at the rapid narrowing of the human-machine readability gap.
Ghost Font is not a traditional font file but a tool for generating dynamic images. When users input text, characters are decomposed into moving dots that match the background color exactly. Only through video playback can the human eye piece together the text using motion cues; pausing any frame makes the dots indistinguishable from the background, causing the text to disappear completely.
This design directly targets a current shortcoming of multimodal AI: mainstream models analyze video as static frames and lack temporal dynamic reasoning. Developer Eric fed the generated videos to GPT-Sol 5.6 Ultra and Claude Fable (Max reasoning mode), and both models read the decoy information in the frames while failing to recognize the actual motion information. GPT-5.5 Pro, after 19 minutes of analysis, fabricated text that did not exist.
In 2013, designer Sang Mun released the ZXX font, which used noise, scratches, and false marks to confuse optical character recognition (OCR). However, by 2026, ChatGPT 5.5 in instant mode could accurately read ZXX text. Ghost Font inherits the anti-OCR spirit but uses a radically different strategy: leveraging video so that individual frames carry no information.
Yet this mechanism is not unbreakable. The author admits that if an AI agent has a local code execution environment, it could analyze the motion trajectories of the dots to decode the text. For this reason, Ghost Font adds a second layer of defense: embedding a decoy message in the frame. When a model locates a 'hidden message,' it often stops at the decoy, mistakenly believing it has cracked the code. Nonetheless, the author believes that with the emergence of video-native models, this technique's effectiveness is limited.
Ghost Font sacrifices readability while protecting information—even for humans, reading motion is more strenuous than static text. The author notes in the documentation: 'It is indeed difficult to read, both for humans and AI.' Speed, dot density, and background noise all affect human recognition efficiency, meaning it is not suitable for everyday communication but rather as a proof-of-concept or CAPTCHA scenario.
The project code is planned to be open-sourced, but the author emphasizes that this is not an encryption alternative: 'The real way to hide information is cryptography. This experiment merely explores whether it is possible to create a visually shared file that AI finds difficult to read.' Its core value lies in measuring the current ceiling of AI visual perception and providing a baseline for future model evolution.
The author envisions Ghost Font being integrated into CAPTCHA systems—current CAPTCHAs are mostly broken by AI, and dynamic dot patterns could become a new adversarial technique. More importantly, it could serve as a progressive benchmark for AI visual perception: when models can reliably decode motion text, it indicates they possess genuine temporal visual reasoning.
The project documentation specifically notes that the tested models (GPT-Sol 5.6 Ultra, Claude Fable) were the strongest versions released in mid-2026, making the failures more informative. However, the successful deception relies on the 'decoy trap'—if the model is told to 'ignore the decoy and look for motion information,' the decoding success rate could increase significantly. In fact, the author did not provide such hints during testing, which directly affects the generalizability of the conclusions.
This article is based on publicly available project documentation and test videos. The tests were not replicated, and the author did not detail the exact configuration of the model versions (e.g., inference temperature, timeout thresholds). The effectiveness of the decoy strategy may fail with prompt engineering adjustments. Rating: medium confidence.
Ghost Font uses motion as a communication medium to temporarily open a gap in human-machine readability, but that gap is likely to close quickly under the pressure of model evolution. It is more a clever experiment for reflecting on AI perception than a durable communication barrier.
Primary report
Primary source