Artificial intelligence, it seems, is always hanging around, ready to interpret and summarize what we write, even if we don’t ask it to. An experimental visual communication system called Ghost Font aims to stand in AI’s way. Computer scientist Eric Lu devised it to make it hard for AI models to decipher messages without human guidance.

While Lu calls Ghost Font an “anti-AI font,” it isn’t a traditional font that uses solid lines to form letters. Instead, it represents text using thousands of tiny moving dots that drift in the opposite direction from those in the surrounding background. Because humans can readily resolve patterns in motion — when points move in the same direction, the brain groups them into a unified whole — the moving dots appear as faint letters.

When Lu tested leading AI models Claude Fable 5 and GPT 5.6 Sol , however, they struggled to decode the writing.

“I was inspired to make Ghost Font one day after texting with my mom, and I noticed that AI kept trying to summarize our messages, many times incorrectly,” Lu, a former Google associate product manager, said in an interview. “I was annoyed that it was being so intrusive in our private conversation and I started brainstorming ways that I could potentially communicate without AI reading everything.”

For now at least, he sees Ghost Font as less a practical tool than an experiment that highlights an emerging tension as AI assistants increasingly summarize emails, texts and documents, often by default . Some users are happy to accept AI’s help, but others want more control over when algorithms can “read” and sum up what they write.

Lu believes Ghost Font could help measure the ever-shifting capabilities of AI, which is advancing at a head-spinning rate.

While in July 2026, it might be difficult for the most capable AI systems available to read Ghost Font without human intervention, “it will be interesting to compare in a few months or next year,” Lu said. “There could be more video native models that read Ghost Font right away just like a human can.”

One of Lu’s big takeaways: “The gap is closing, quickly.”

Lu’s X post about the experiment has so far racked up 17.7 million views, reflecting fascination with the idea that humans might briefly have an advantage over AI and drawing people eager to test its limits. A number have set out to prove that Ghost Font is no match for AI by sharing the prompts they used to get AI to read the typeface successfully.

“I got GPT-5.6 Sol to read ‘Ghost Font'... by just telling it in what direction it looks like the noise is moving,” Riley Goodside, a former Google DeepMind prompt engineer, wrote on X.

But Lu stresses that he designed Ghost Font to function without being instructed specifically what to look for, and that those trying to prove an AI agent can read it “miss the point of what the project was about in the first place.”

“The goal was never to create a completely secure messaging technique that is impossible for any AI to crack,” he said. “There are plenty of more secure ways to protect messages from AI — for example, placing a message behind a knowledge factor like a password."

As long as the data is publicly available, technically, it can be read with the correct prompting, he said. “But slowing down AI agents can still be useful in many situations,” he added.

A Bifurcated AI Discussion

Lu founded Mixfont, an AI font generator that creates typefaces in seconds from a single prompt or image. He believes discussions around AI tend to be divided into “AI is good” or “AI is bad” camps, and would like to see more nuanced conversations.

“The project is still valuable as an example of how humans perceive things differently from AI,” he said. “There are still things, like instant motion perception, that make humans unique, and those avenues are worth exploring and researching.”

He also believes the project underscores AI fatigue , the pressure and expectation workers feel to keep up with the latest AI tools.

“Even as the technical gap closes,” he said, “there’s still a pent-up desire for conversations and connections that are human-to-human and don’t involve AI.”

In 2013, another designer, Sang Mun, released ZXX , a typeface with four fonts camouflaged with visual noise that foiled optical character recognition software while remaining visible to humans. Considered “surveillance proof” at the time, modern AI agents can easily read text rendered in the font, a reminder of just how quickly today's AI-resistant techniques may become obsolete.