Ghost Font is a motion‑based text system that renders letters as moving dots in a video. The motion, noise, and decoy text combine to create a format readable by the human eye but hard for current AI models to parse. The prototype runs locally; generating a video from typed text and downloading it for sharing. The project builds on earlier work, such as the 2013 ZXX type fighter, but modern AI can read ZXX.

Tests with Claude Fable and GPT‑Sol 5.6 Ultra show that single‑frame screenshots reveal only static dots; the models can decode a hidden decoy message but fail to retrieve the true moving message unless guided. The decoy layer tricks automated agents into misidentifying the embedded text. Even the newer GPT‑5.5 Pro hallucinated a non‑existent message after 19 minutes, underscoring the difficulty of extracting motion.

Ghost Font’s design offers a new angle for CAPTCHA systems, where motion could outpace static image challenges. It also supplies a benchmark for multimodal AI: current models split videos into frames, but a future video‑native architecture would need to process motion directly. The experiment highlights a narrowing gap between human visual perception and AI, but the font remains hard for humans to read efficiently.

In practice, Ghost Font shows that motion‑based encoding can circumvent existing OCR and multimodal pipelines, yet it is not a replacement for encryption. Developers may adopt it for low‑risk secret sharing/null CAPTCHAs, but the approach is limited by speed and human usability. The work illustrates the current limits of vision models and points to video‑centric research as a frontier.