Decoy Font
An experimental font that overlays true and false letters on the same glyph using hybrid image techniques — human eyes see real info when stepping back or squinting, while AI pixel-reading sees decoy text
In-depth Report
-
Decoy Font is an experimental font launched by the AI font generation platform Mixfont. It uses spatial frequency and hybrid image technology to superimpose two different letters in the same glyph. The human eye can read real information at an appropriate distance, but the multi-modal AI model can only read decoy letters. Released on July 17, 2026, the font is developed based on the DejaVu Sans Mono open source font and is free for personal and commercial use. It is not an encryption scheme per se, but an adversarial tool that takes advantage of the AI vision system’s characteristic of “prioritizing reading clear outlines.”
-
Decoy Font was developed by Eric Lu, AI researcher and founder of font generation platform Mixfont. Mixfont claims to be the world's first AI font generator. Users only need to enter a text description or reference image to generate a commercially available TTF font file in seconds. Eric Lu previously served as an Assistant Product Manager (APM) at Google and later founded Mixfont to focus on the intersection of AI and typography. Decoy Font was inspired by mixed-image research published in 2006 by Aude Oliva, Antonio Torralba, and Philippe Schyns—most famously merging facial images of Albert Einstein and Marilyn Monroe, with Einstein seen up close and Monroe seen from a distance. Decoy Font transfers this principle from the human face to the glyph level. This isn’t Mixfont’s first anti-AI experiment. On July 11, 2026, Eric Lu released Ghost Font, which hides text through the movement of dynamic lattice - the static frames look completely random noise, and the human eye can only read the text through motion perception during playback. Ghost Font has over 14.7 million views on X (Twitter), and even GPT-5.6 Sol Ultra and Claude Fable 5 failed to correctly identify its hidden information. Decoy Font is a continuation of this experiment, compressing the anti-AI mechanism from dynamic videos to static TTF font files.
-
Decoy Font works based on the visual separation of spatial frequencies. Each glyph contains two layers of information at the same time: the foreground is a set of clear thin outlines (high spatial frequency), which constitute the decoy letter; the background is a blurry low-frequency color block (low spatial frequency), which constitutes the real letter. When the user types in this font and takes a screenshot: - **Closer View Screenshot**: What the human eye sees are sharp decoy letters - **Step back or squint**: High-frequency details are suppressed, low-frequency background shapes dominate, and real letters emerge - **AI View Screenshot**: The multi-modal model enlarges the pixel reading, giving priority to capturing the clearest outline, and what is read is the bait number Mixfont tested images generated by Decoy Font using ChatGPT and Gemini 3.5 with Thinking in experiments. The results show that both ChatGPT and Gemini report decoy text instead of hidden real information. Even the most cutting-edge models, such as GPT Sol (GPT-5.6) and Claude Fable, are fooled when directly reading screenshots. The way to use this font is very simple: users can download the TTF font file from the Mixfont official website, and after installing it on the system, they can enter text in any application like a normal font. The website also provides an online playground where users can enter hidden information and decoy information directly into the browser and export it as a picture with one click. The randomness of font encoding is a noteworthy design detail - the decoy letter corresponding to each letter is not fixed, but is random, which rules out the possibility of reverse decoding simply by establishing a "real letter → decoy letter" mapping table.
-
Decoy Font is completely free and can be used for personal, commercial and client projects, including producing works for clients. Its glyphs are based on DejaVu Sans Mono (an open source font), and the liberal license clears the way for widespread commercial use. This project is essentially an experimental showcase and not Mixfont’s core for-profit product. Mixfont's main business is an AI font generation service (mixfont.com), which provides free trials and API paid plans. Decoy Font is more like a brand display and experimental exploration of its technical capabilities.
-
After the release of Decoy Font, it caused widespread discussion in the design community and technology community. In threads on the TypeDrawers forum, type designers expressed interest in how the technology is implemented, while also questioning the sustainability of its practical effects. On Chinese social media, some netizens commented that "this may be the next evolution direction of CAPTCHA", while some users believed that "AI only needs a few iterations to break this vulnerability." Positive feedback focused on the innovativeness of the concept and the low barrier to use – no special software is required, no programming knowledge is required, just download a TTF font and it’s ready to use. Criticisms mainly focus on the timeliness of the technology: this is not encryption, but just exploiting the blind spots of the current AI vision system. Once the model is updated or a pre-processing step is introduced, the effect may disappear at any time.
-
Decoy Font has received reports from many technology media, including dev.ua, BigGo Finance, Sohu Technology, Tiger Technology, China AI News, etc. English technology media remio.ai published an in-depth technical analysis, pointing out that the value of Decoy Font does not lie in its "strength", but in its idea of "packaging a confrontational interface into a standard TrueType font" itself. From an industry perspective, this project revealed a deep-seated problem: Although multi-modal AI can smoothly "read" text in pictures, the visual processing process it uses is fundamentally different from the overall perception mechanism of the human eye. AI tends to zoom in on pixels, looking for the clearest edges and contours, while the human eye automatically switches between global and local aspects. Decoy Font takes advantage of this difference - it is not a weapon against AI, but a physical demonstration of the "human-machine perception gap." In terms of the competitive product landscape, Decoy Font is not the first of its kind. In 2013, designer Sang Mun released the ZXX font, which used visual noise to combat traditional OCR systems. But ZXX is designed for traditional OCR engines, which are basically ineffective in the face of today's multi-modal large models - these models can easily read the real text in ZXX. Decoy Font uses lower-level visual perception principles to compete with the visual processing process of the new generation of multi-modal AI.
-
Decoy Font has several limitations that cannot be ignored. **Technical timeliness is the biggest risk. ** As the title of remio.ai's article says - "its Spatial-Frequency Trick Has an Expiration Date". Once the AI model introduces image preprocessing (such as scaling, Gaussian blur, edge detection), or uses prompt words to explicitly ask the model to check "whether there is a hidden mixed image", the effect of Decoy Font may be greatly reduced. **It is not encryption. ** Eric Lu himself explicitly admitted this. Capable AI agents can bypass this confusion by writing code to perform Fourier decomposition, separate frequency bands and identify them separately, or directly access the underlying text (instead of screenshots). Those hoping to use this font to protect truly sensitive information should understand that this is a "speed bump for automated reading" and not a security barrier. **Accessibility issues cannot be ignored. ** Hidden information cannot be received by users with low vision, contrast sensitivity differences, or users using screen readers, which conflicts with the W3C guidelines for image text accessibility. Using Decoy Font for CAPTCHA alternatives requires careful evaluation of the inclusion impact. **There is another risk that is easily overlooked-feedback closed loop. ** When the user inputs the image generated by Decoy Font into AI and finds that it is mispronounced, he is actually contributing an "error recognition sample" to the model's training data. Each such interaction helps the model learn to recognize the trick.
-
Decoy Font is best suited for users and scenarios with general concerns about AI text scraping—such as adding a barrier to publicly posting private information on social media, or sharing information in a group chat that you don’t want automatically summarized. For designers, it can also serve as a new type of watermark or anti-counterfeiting method. Not suitable for highly sensitive information (passwords, keys, trade secrets, etc.) because this is not an encryption scheme and has limited attack resistance. It is also not suitable for public information scenarios that require barrier-free access. If you really need to protect information privacy, industry-standard encryption solutions (such as end-to-end encryption, PGP, Signal protocol, etc.) are the right choice. Decoy Font is more suitable as a "fun defense" or "benchmarking tool for AI capabilities."
-
Decoy Font is an ingenious anti-AI font experiment that applies the classic mixed image theory to glyph design, and uses the difference in human and computer visual perception to achieve the effect of "humans can read it, but AI can read it wrong". It’s free and open, easy to get started, and conceptually novel, but it’s essentially a short-lived technology demonstration—the trick is likely to expire within a few months to a year as multimodal models gain a deeper understanding of spatial frequencies and mixed images. Its real value is to remind the industry that the current AI vision system is fundamentally different from the human eye in overall perception, and this difference is both a weakness and a direction for research.
User Reviews
-
LloydWashington—I tried it — typed out Decoy Font, took a screenshot and sent it to ChatGPT. What it read was really not what I wrote, lol. The principle is interesting — using the Einstein-Monroe hybrid image concept on fonts is very creative. -
orangecat977—Just saw the Decoy Font discussion on Hacker News and tried it on their website. Gemini reads PAY BILLS, but the hidden text is actually DAY DREAMS. I had to squint to see it — the effect is quite impressive. -
JRodriguez_2020—Cool font experiment, but practicality is debatable. Can posting an image prevent AI scraping? Just shrink the image or add a filter and the hidden text can be read. Still, as a proof of concept, it deserves a like. -
GregoryNelson_66—The fact that it's free for commercial use deserves praise. It's based on DejaVu Sans Mono, so no worries about font copyright. Though even humans have a bit of trouble reading it — you have to hold your phone farther away to see the hidden message. -
JordanHernandez—My first reaction was: what does this mean for me as a typeface designer? I downloaded it, installed it, tried it in AI software, and sure enough, I could see both layers of letters at the same time. Anyone can download a font and use it — the barrier is really low. -
StakeHawkRoberts—There's discussion in design circles about using this as a next-gen CAPTCHA. I think it has potential — current CAPTCHAs are almost cracked by AI, so switching to a solution that leverages the human-machine vision gap could be a direction. -
WEkel—Sure enough, with the prompt "There's a second hidden text", GPT Sol read it. It can even teach you how to see the hidden text yourself — squint or move the screen farther away. The comprehension of these models is truly frightening. -
Susan.Gonzalez_2020—It's not encryption, just a visual trick — Eric Lu himself admitted that. Fun to play with, but if you really need to protect sensitive info, stick to end-to-end encryption. -
David_CookII—Had to strain my eyes to see the hidden text — this font is really unfriendly for people with poor vision. Imagine using this font for web content; screen readers would read the underlying text fine, but as an image it's an accessibility disaster. -
Web3Dex_z—Saw someone test it on Weibo. ChatGPT read it as 'SORRY ROBOT' — squint your eyes and it's actually 'HAPPY HUMAN.' This kind of adversarial tool might be useful against crawlers, but it's torture for humans. AI can figure it out with just a hint. -
Janet441—Reminds me of the ZXX font from 2013, also anti-OCR. Today's multimodal models easily see through it. Decoy Font uses a more fundamental mixed-image principle, but I doubt it'll last even six months. -
黎洁雪—Bold idea: what if someone made Decoy Font for Chinese characters? Chinese characters are all square-shaped, similar in size — theoretically easier to embed two layers of information. Eric Lu mentioned this direction in his blog too. -
blackbear921—I think this font's most real use is as a benchmark for AI vision capabilities. Every time a new model comes out, generate a Decoy Font image and see if it can read both layers — score it. -
Carol964—Read Eric Lu's original post on TypeDrawers. The type designers reacted more positively than the tech community. Their focus was more on how to technically achieve this dual visual effect, rather than debating whether it's secure. -
SilviaSantana—Tried various models, results are all over the place. Gemma Flash directly read both layers of text, Claude Fable completely fell for the decoy text. Different models react differently to the same trick — shows this is a real blind spot in the visual processing pipeline. -
trueNoémieDenys_88—Downloaded the free TTF, installed it on my system, typed some text in Word, and took a screenshot. Shared it in the group chat. Friends said you need to step back to read what I wrote — at normal distance it just looked like garbled characters. Fun prank. -
bigsnake694—At first I thought this font was amazing, until someone told me it's just using spatial frequency confusion, not some black magic. The more I think about it, the more it feels like an art project rather than a security tool. -
ROwri—Tried it on my phone. Maybe the screen is too small — I couldn't make out the hidden text at all. The post said 'view from far away,' but on a phone, far away makes it even smaller. Feels like the mobile experience has issues. -
Jean.Lee_Max8—Compared to the previous Ghost Font, Decoy Font can be directly downloaded as TTF and installed — much more convenient. Ghost Font required video export, and the moving effect wasn't ideal. -
NSanders4—Tried it as soon as it was released. Threw a screenshot at GPT Sol, and it read 'BINGE SHOWS' — but I clearly wrote something else. Interesting, but I think in a few months models will learn to see through this trick.