Ghost Font hides secret text in moving dots only humans can read

AI is creeping into every corner of our digital lives, to the point where mobile assistants are now trying to summarize personal texts from our parents without asking. That raises a fascinating question: is there still a way to display a message in plain sight that any human can read, while leaving top-tier AI completely blind?

San Francisco developer Eric Lu dropped a prototype experiment called Ghost Font. Technically, it is not a traditional font file you install on your system. Instead, it is an anti-AI messaging tool that embeds words inside dynamic fields of moving dots.

Pure static in screenshots, crystal clear in motion

The mechanic behind Ghost Font is deceptively simple. It does not draw letterforms with lines or solid curves. Instead, every character is rendered through thousands of tiny moving dots.

When the video plays, dots in the background drift in one direction, while dots that make up the letters move in the opposite direction. The human brain is wired to instantly group objects traveling with the same trajectory into a cohesive shape, so the letters pop out effortlessly to our eyes.

For frontier vision models, however, it is a nightmare. Most current multimodal models ingest video by slicing it into individual frames and analyzing them as static images.

In any single frame of a Ghost Font clip, the foreground dots look completely identical to the background dots. If you pause the video or take a screenshot, there are zero legible shapes—just a wall of random static noise. It physically breaks standard OCR and image-based reading in one stroke.

People have tried fooling machines before. Back in 2013, designer Sang Mun released ZXX, a typeface loaded with crossouts and camouflage designed to defeat OCR surveillance. But fast forward to today, and even ChatGPT 5.5 on Instant mode can breeze through ZXX in a single prompt. Ghost Font, by contrast, left ChatGPT 5.5 Pro thinking for 19 minutes, only for the model to hallucinate a message that was never there.

Fooling code-running agents with a decoy trap

Lu knew that hiding text in motion alone would not stop a dedicated agent forever. If an AI has access to a local code interpreter, it could write a script to calculate optical flow and track dot coordinates frame by frame.

To counter this, Ghost Font adds another defensive layer: every generated video includes a deliberately planted decoy message.

When heavyweight reasoning models like Claude Fable or GPT Sol 5.6 Ultra deploy code to hunt down hidden temporal cues, they inevitably hit the decoy first. Thinking they cracked the puzzle, the agents stop searching, leaving the real text completely untouched.

The entire generator runs locally in your browser without sending a single byte to an external server. You just type a phrase, watch the live preview, and download the video to share.

Why it matters

We think Ghost Font touches on something profound: the definition of digital privacy is shifting. For decades, privacy meant keeping your messages hidden from other people. In the generative AI era, privacy increasingly means the right not to be ingested, indexed, and summarized by an autonomous agent. Designing user experiences around the right to remain unread by machines could easily become a defining UX theme over the next decade.

In terms of practical utility, this approach offers a brilliant blueprint for modern CAPTCHA systems. Bots and multimodal models have completely crushed traditional image puzzles. Using human temporal perception coupled with algorithmic decoys could finally give human-verification systems a defensible edge again.

Granted, Ghost Font is an experiment rather than a bulletproof cryptographic shield, and reading shimmering dots can certainly strain human eyes over time. Future video-native models that bypass frame splitting might eventually crack the motion directly. But right now, weaponizing human perceptual quirks against billion-dollar neural nets is both deeply clever and wonderfully defiant.