GhostFont Tricks AI Into Misreading Dynamic Text

新智元 · wechat · 2026-07-13

GhostFont Experiment Tricks AI Into Misreading Dynamic Text

This article introduces an experiment called GhostFont: web pages encode text into a noisy video where letter pixels slide up while background noise slides down. The human eye can quickly recognize it via motion perception, but many multimodal models analyze it frame-by-frame and mistake it for pure noise, sometimes even reading out decoy text preset by the author.

The article notes that researchers only needed a simple prompt—"the noise forming the letters slides up, the rest slides down"—to make the models read the correct content, proving that such "human-exclusive" visual barriers are quite fragile. The author also compares this to the 2013 ZXX font: it could evade OCR back then, but is now easily deciphered by ChatGPT, highlighting the rapid advancements in visual recognition over the years.

Core Takeaways

Current multimodal models still struggle to understand motion information stably, defaulting to behavior more like "image models in a video shell."
While decoy text can trick models, the barrier instantly collapses once the correct motion hint is provided.
As native video understanding models become widespread, these "human-only" tricks will likely become obsolete very quickly.

Original post →

GhostFont Tricks AI Into Misreading Dynamic Text

More from Fun