Ghost Font: Motion Illusion Text That Fools GPT-5.6 and Fable 5
Eric Lu's Ghost Font (Mixfont, July 11, 2026) hides text in dot motion — background scrolls up, letter dots drift down. Fable 5 and GPT-5.6 Sol Ultra miss it; humans read it instantly. CAPTCHA implications and frame-by-frame bypass.
Eric Lu (@ericlu, Mixfont) posted Ghost Font on July 11, 2026 — a video effect where text exists only in motion. Background dots scroll up; letter dots drift down. Your visual system fuses the opposing flows into words. GPT-5.6 Sol Ultra and Claude Fable 5 could not.
The clip hit 14.7M views on X. Solomon Hykes (@solomonstre, Docker founder) replied: "Captchas are about to get so much worse."
Ghost Font is not a traditional font file you install — it is a motion graphics illusion:
snippet
Frame N (static): █ ░ █ ░ ░ █ ░ █ ← looks like noise
Over time:
background field → scroll UP
letter field → scroll DOWN
Human integration → "HELLO" (example shape)
The effect exploits temporal integration in human vision — the same family of phenomena as biological motion and aperture illusions. Single frames carry no legible OCR text; the message lives in velocity differences between dot populations.
Lu's demo pairs the motion text with decoy static strings models latch onto — part of why frontier VLMs report confidently wrong answers.
What frontier models did
Eric Lu's test (July 11)
I created a font called Ghost Font that only humans can read. Tested it in Fable and GPT 5.6 Sol Ultra and neither was able to decipher it correctly.
Model
Reported behavior
GPT-5.6 Sol Ultra
Failed · often cited decoy on-screen text
Fable 5
Failed · same decoy trap
Community tests extended to Grok (@tranmautritam challenged @grok on the clip) with mixed results depending on prompting and frame tools.
Riley Goodside's inverse test (July 12)
@goodside posted meaningless eyes-closed phone scribbles labeled as handwriting:
Model
Behavior
GPT-5.6 Sol
Hallucinates readable content reliably
Fable 5
Abstains or says it is not real writing
explainx.ai read: Ghost Font and Goodside's scribble are complementary failures — Sol over-reads garbage; both Sol and Fable under-read motion-encoded truth. Neither is a trustworthy OCR layer without verification.
Why multimodal models miss it
Human vision
Typical VLM pipeline
Continuous motion integration
Sparse frame sampling
Opposing flow → form
Static salience wins (decoy text)
Phi / aperture prior
Trained on legible pixels
No "caption" needed
Text-in-image bias from pretraining
Models that can crack Ghost Font in replies often do so via frame-by-frame diff analysis or explicit user instruction to track dot trajectories — not native "watch the video once" understanding.
Optical flow models (specialized, not general LLM)
Human farms + screen recording
Ghost Font is a demonstration, not a deployed gate — but it surfaces the same lesson as GPT-5.6 bio bounty: security through obscurity in perception erodes fast once attackers optimize.
Frontier labs market native multimodal parity with humans. Ghost Font is a cheap, viral falsification — not a rigorous benchmark, but a memorable one. 14.7M views beat most vendor eval PDFs for public persuasion.
Try it yourself
Lu attached a local prototype to generate Ghost Font clips — check @ericlu's thread for the latest link. Workflow for researchers:
Render a clip with your hidden message
Ask Fable 5 / GPT-5.6 Sol to read the video cold
Retry with frame-by-frame or optical-flow hints
Compare to human accuracy at normal playback speed
Document decoy text placement — models grab it first.
Limits and ethics
Limit
Note
Accessibility
Motion-only text excludes some humans
Not cryptographic
Motion analysis can reverse
Demos ≠ product
Mixfont typography tool ≠ shipped CAPTCHA
Hallucination risk
Sol may "read" Ghost Font wrong confidently
Do not deploy human-only gates without fallbacks and audit for exclusion and bypass.