A designer named Eric just built something that flips the usual AI story on its head. Instead of a tool that helps machines read faster, he made one that stops them cold. It’s called Ghost Font, and according to Hacker News, it’s an experiment in writing messages that a human eye picks up instantly but even the strongest AI models can’t decode.
Here’s the twist: it isn’t really a font. There’s no TTF file to install. Ghost Font hides a message inside the motion of a video, using moving dots that spell out letters only while they’re in play.
How it works
The whole thing runs on a simple perceptual trick. Every letter is made of dots that look exactly like the background noise. Nothing stands out in any single frame.
- Pause the video, and the dots blend into static mush. No readable text.
- Screenshot the page, and you get a blank field of noise. The message lives in the movement, not the pixels.
- Watch it play, and your eye stitches the motion together into words.
- Everything runs locally in the browser. No data touches a server.
Eric also stacked on a second layer of defense. Each generated video carries a decoy message buried inside it. A determined AI agent digging for hidden text tends to find the decoy first and assume it hit the jackpot, missing the real message entirely.
Ghost Font vs. ZXX
The idea isn’t brand new, and Eric is upfront about that. Back in 2013, designer Sang Mun released ZXX, a typeface built to defeat optical character recognition. It camouflaged letters with noise, strike-throughs, and false marks. At the time, people called it “surveillance-proof.”
That armor doesn’t hold anymore. Eric dropped a ZXX sample into ChatGPT 5.5 on Instant mode, and the model read the text cleanly in a single prompt, catching small details along the way. What stopped OCR in 2013 barely slows a modern multimodal model.
Ghost Font is the 2026 answer to the same question. Where ZXX fought static image recognition, Ghost Font attacks the way today’s models actually process video: they slice it into frames and study each one. Since no single frame holds the message, that approach hits a wall.
The test results back it up. Eric fed Ghost Font videos to leading systems, including Claude Fable with max reasoning and GPT Sol 5.6 Ultra. Both grabbed the decoy and missed the real message unless they were told the exact technique to look for. ChatGPT 5.5 Pro spent 19 minutes on a single message and hallucinated one that didn’t exist.
Where it’s not bulletproof
Eric doesn’t oversell it, and that honesty is worth noting. Hiding a message in motion isn’t foolproof. A dedicated agent with a local code environment can analyze the dot movement directly and pull the message out. That’s exactly why the decoy layer exists as a backstop.
He’s clear that the only truly unreadable message is an encrypted one locked behind a key that only humans hold. Ghost Font isn’t trying to be encryption. It’s exploring whether a plain, shareable visual file can resist casual AI reading.
There’s also an honest catch: Ghost Font is genuinely hard for humans to read, too. The gap between human and machine perception is real, but it’s narrow, and it’s shrinking.
Why it matters
This is significant because it targets a real weak spot in how multimodal AI sees the world. Models today are image-first. Video is still an afterthought they chop into stills. Ghost Font turns that limitation into a feature, and Eric expects a video-native model will eventually read these messages directly.
The practical angles are interesting. He points to CAPTCHA systems, most of which AI already cracks, as a natural fit for motion-based challenges that stay easy for people. Ghost Font could also work as a benchmark for tracking how fast machine visual perception improves.
Eric plans to open-source the video generation code and expand it to handle longer text. For now, it stands as a small, human-shaped act of resistance: a way to pass a note that machines can watch all day and still not understand. You can find the full write-up and playground at the original source.