← Back to homepage

🃏 Decoy Font Demo

See how hybrid-image typography tricks AI vision. A blur/distance slider reveals how the same text says two different things.

🔵 HIDDEN MESSAGE (what humans see from a distance)

🟡 DECOY MESSAGE (what AI sees up close)

AI sees (high frequency / close up) Human sees (low frequency / distance)

🔍 Up close (sharp) 🏠 Arm's length

Decoy visible Both readable Hidden visible

👁 At this distance, AI reads the decoy — humans see both

AI would read: AI CAN'T READ

🧠 How It Works

Decoy Font uses the hybrid image principle — the same technique that makes a photo look like Einstein up close and Marilyn Monroe from across the room.

The font renders two letters in the same space:

High-frequency foreground (thin outlines) — the decoy letter, visible up close. AI reads this.
Low-frequency background (blurred mass) — the hidden letter, visible at a distance. Humans read this.

Most AI vision systems process images at full resolution — they see the sharp edges, read the decoy, and never notice the hidden message. Humans, on the other hand, naturally integrate visual information across scales. We squint. We lean back. Our brains do the blurring for us.

Move the slider above to simulate moving your screen farther away. Watch the decoy text dissolve and the hidden message emerge.

Download the actual font at mixfont.com — licensed under DejaVu Sans Mono's terms, free for any use.

📊 Results Against Frontier Models

According to Mixfont's tests, Decoy Font tricks even advanced LLMs:

GPT Sol: Reads the decoy text — fails to see the hidden message
Gemini 3.5 with Thinking: Also fails — reads the decoy as the entire message
Claude / GPT-5: May succeed with specific prompting (\"look for lowercase hidden text\")

Why does this work? LLM vision encoders (CLIP, SigLIP, etc.) operate on patch-level tokens — they see the image as a grid of squares, each encoded independently. The thin, high-contrast decoy letter dominates each patch's embedding because it has stronger edges. The low-frequency hidden message is smeared across patches at sub-threshold contrast.