Ghost Font: A font that humans can read but AI cannot

Original Article ↗ Hacker News Discussion ↗

Can AI Actually Read “Ghost Font”?

  • Many commenters say the claim “humans can read it but AI cannot” is already false.
  • Frontier models (GPT 5.6, Claude, Gemini, Fable) reportedly decode the animated text when given the video and no special instructions, often via optical-flow / motion-based analysis.
  • Several people share code or algorithms using frame differencing, phase correlation, optical flow, temporal averaging, or video compression artifacts to reveal the text and then OCR it.
  • However, early or naive attempts often only pick up the decoy static message (“WRITTEN IN GHOST FONT”), not the user-entered animated one, leading to confusion about what “success” means.

Decoy vs. Actual Message Confusion

  • The demo embeds two messages:
    • A faint, static, low-frequency decoy visible in screenshots or long exposure.
    • The real, motion-defined text that only appears through temporal changes.
  • Many humans and AIs initially see only the decoy; commenters describe this as a major UX/communication flaw of the demo.

Is It Really a “Font”?

  • Several argue it’s not a font but a video/animation effect, since each glyph is defined over time rather than as a static shape.
  • Some note that a single frame is effectively noise, so talking about “font” in the traditional sense is misleading.

Human Readability & Accessibility

  • Some readers find it “clear as day”; others can barely read it or not at all, citing astigmatism, age, or screen characteristics.
  • Comparisons are made to Magic Eye images and motion illusions.
  • Multiple comments flag serious accessibility problems, including incompatibility with screen readers and likely conflicts with disability regulations.

Usefulness, CAPTCHA, and Arms Race

  • Suggested uses: CAPTCHAs, anti-scraping, steganography, “AI-safe” writing.
  • Many doubt long-term viability: once the decoding technique is known, models or tools can automate it, and temporal analysis is straightforward.
  • Some propose randomizing motion/noise parameters per-client to slow learning, but others see this as classic security-through-obscurity and part of a never-ending AI–CAPTCHA arms race.

Related / Alternative Obfuscation Ideas

  • Mention of homograph tricks, Unicode-mangling fonts (“Noroboto”), PDFs with remapped glyphs, psychedelic/qualia-based cryptography, and motion or illusion-based schemes that privilege certain human perceptions.