Overview
- Designer Eric Lu published an experimental type called Ghost Font that embeds short messages in short animated clips of thousands of moving dots so humans perceive letters through motion grouping while any single frozen frame looks like visual noise.
- Lu’s example clips, posted on X and widely shared online, attracted massive attention with reporters and users saying the original post drew roughly 17–18 million views within days.
- Public tests reported by Lu and journalists found several advanced multimodal systems including tests of Anthropic’s Claude Fable, OpenAI’s GPT‑5.6 Sol Ultra, ChatGPT and Google’s Gemini often failed to recover the hidden text and instead returned a planted static decoy phrase included in the files.
- Other users have shown models can sometimes be steered to the hidden message if given explicit guidance or specialized analysis, and Lu has suggested motion-only encoding could be adapted for CAPTCHAs or anti-scraping markers that remain human-readable.
- Researchers and reporters note the gap reflects current AI reliance on per-frame analysis rather than temporal optical-flow methods, so the demonstration serves as a short-term benchmark of video understanding and a prompt for developers to add motion-aware tools.