This Font Shows Two Different Messages: Humans Read One, AI Another

With the spread of multimodal models, which are increasingly effective at extracting information from images, screenshots, and digitized documents, there is growing interest in techniques capable of limiting the automatic acquisition of content. In this context, Decoy Font (via Reddit), an experimental typeface developed to be understandable to humans but significantly harder to interpret for Artificial Intelligence systems, emerges. The idea behind the project is not to encrypt the text but to exploit the limits of artificial perception through an optical illusion.

Each letter actually incorporates two different representations that coexist in the same space: a thin outline in the foreground and a blurred low-frequency shape in the background. Depending on the observation distance or the level of detail analyzed, a different message emerges.

According to the project's author, this technique derives from the principle of hybrid images, a particular category of optical illusions made famous by the composition that shows the image of Albert Einstein and Marilyn Monroe fused together. By applying the same concept to typography, Decoy Font manages to present two different characters simultaneously within the same letter.

The result is that modern multimodal models, which typically process images by focusing on the sharpest details, tend to read the text represented by the foreground outlines, instead ignoring the message actually intended for the human observer. The author shows several examples where screenshots created with Decoy Font are misinterpreted by models like ChatGPT and Gemini, which return words different from those perceived by the human eye.

Unlike other similar experiments, Decoy Font does not require animations or dynamic effects. The project is also distributed as a regular TTF file, installable on any compatible system and usable as a common typeface. This makes the technique easily integrable into documents, presentations, images, or other graphic content without requiring dedicated software.

The stated goal is not to make automatic reading of the text impossible but to introduce an additional level of complexity that makes the extraction of content by scraping systems, OCR, or generative AI models less straightforward. The same author emphasizes that future models, or agents equipped with more sophisticated programming and analysis tools, might succeed in decoding both components of the illusion, especially if appropriately trained. For this reason, Decoy Font is not presented as a definitive security solution nor as an alternative to encryption, but rather as a first obstacle capable of discouraging automatic acquisition or superficial observation.