Do We Need a New Language to Talk Without AI Listening?

For the first second, the screen looks broken. Thousands of tiny dots drift across it like television static, aimless and meaningless.

Then some of the dots start moving together.

Your brain does the rest. Letters rise out of the noise, as sharp as anything printed on this page, and once you see them you can't unsee them.

Now pause the animation. The letters vanish. Take a screenshot and you've captured pure static, because the words never existed in any single frame.

A machine scanning that screenshot sees exactly what the screenshot contains: snow.

That's Ghost Font, an experimental typography system from Eric Lu at Mixfont, along with its sibling Decoy Font. Neither is a security tool. They're demonstrations of something stranger and more useful to understand: humans and AI don't see the world the same way.

How the trick works

Every letter in Ghost Font is built from thousands of tiny dots. Nothing outlines the letters. No strokes, no edges, no shapes.

The only thing separating a letter from the background is direction. The dots that make up the letters drift one way while the background dots drift another.

Your visual system is spectacular at exactly this. When objects move together, your brain automatically groups them into one thing, with no conscious effort required. If you've ever watched a flock of birds separate from the sky behind them, you've run this exact software.

So after a second or two of watching, the message simply appears.

Why machines miss it

Many multimodal AI models understand video by examining a sequence of still frames. And a single frame of Ghost Font is, by design, indistinguishable from random noise.

Hand someone one page of a flipbook and they'll see a meaningless scribble. Flip through the whole stack and a little figure sprints across the corner. Ghost Font stores its message the same way the flipbook stores the sprint: in the space between frames.

An AI that reads frame by frame keeps picking up single pages, looking for the runner.

The buried fake treasure

There's a second trick. Every Ghost Font animation also carries a decoy message, and Mixfont's own write-up explains the intent: an AI hunting for hidden text "might first find the decoy message and think that that is the real embedded message."

The decoy exists to waste a machine's attention and send automated analysis home with the wrong answer.

Decoy Font, the static sibling

Decoy Font skips the animation entirely. Each character is drawn with carefully designed visual clutter baked into its shape.

Your brain filters the junk in milliseconds and reads the intended letter. An AI vision model has to weigh every shape it sees, and the extra strokes compete with the real ones, raising the odds it reads the character wrong.

Ghost Font hides the message in time. Decoy Font hides it in clutter. Both live in the same gap between human perception and machine perception.

Before you get excited: this is a speed bump

Encryption mathematically prevents anyone without the key from reading your data. Ghost Font does nothing of the kind.

Anyone who understands the technique, or writes software designed to decode it, can recover the message. It slows automated readers down. A determined one still gets through.

And "determined" arrived fast. After Ghost Font went viral, prompt engineer Riley Goodside reported getting GPT-5.6 Sol to read a Ghost Font message simply by telling the model which direction the noise appeared to be moving.

Creator Eric Lu doesn't consider that a failure. By his own account, the goal was never a messaging technique impossible for AI to crack, and as long as the data is public, it can be read with the right prompting.

Which is really the project's whole lesson, demonstrated live: the gap between human vision and machine vision is shrinking, quickly.

We've been here before

In 2013, designer Sang Mun released ZXX, a typeface built to confuse the optical character recognition software of its era while staying readable to humans. It worked.

Modern AI vision models read ZXX with little difficulty.

That's the pattern with perception-based defenses: they age like produce. What confuses machines this year becomes ordinary text to them the next.

Why this matters to you

More and more of what you write gets scanned by software: email assistants sorting your inbox, AI tools summarizing your documents, models training on whatever crosses their path.

Ghost Font is best understood as an art project about that reality. It proves the human eye still has a few tricks machines haven't matched, and it proves those tricks have expiration dates.

If there's a takeaway for your actual security, it's this: hiding from machines through clever perception tricks buys you time at most.

For anything that truly needs protecting, use real encryption. The locked door still beats the speed bump.

Sources: Ghost Font at Mixfont · Decoy Font at Mixfont · Forbes on Ghost Font · Wired on ZXX​