Decoy Font
Decoy Font 是一种新型字体，利用混合图像原理来迷惑人工智能，从而掩盖文本信息。它通过不同的空间频率在同一视觉区域内同时传达两组不同的字形：前景由细而清晰的轮廓构成，背景则是一团低频的模糊块。这样的视觉错觉会随着观察距离或焦点的变化而改变感知内容——从远处或眯眼看时，人眼更容易识别出隐藏的信息。
这种字体之所以有效，正是因为大多数 AI 系统处理视觉信息的方式。大型语言模型和光学字符识别（OCR）技术通常在近距离分析像素，会优先捕捉轮廓最清晰的高频前景文本。因此，当用 Decoy Font 制作的图像提交给 ChatGPT 或 Gemini 3.5 等高级模型时，AI 往往只识别到诱饵文本，而无法发现人类肉眼自然能读出的隐藏信息。
除了用于生成静态图像外，Decoy Font 还以可用的 TTF 文件形式发布，用户可以在各类项目中直接输入并应用该字体。它为希望保护信息或知识产权免受自动化 AI 抓取的人提供了一种实用工具。虽然这并非万无一失——针对性强的代理或经特殊提示的模型最终可能破解这种技巧——但作为对抗粗略观察和常规 AI 提取的第一道混淆防线，它非常有效。
该项目是 Mixfont 在排版与人工智能交叉领域的更广泛探索之一。通过应用空间频率技术，创作者开发出一种比基于视频的混淆方法更易使用的方案，可能在 CAPTCHA 验证或私人消息等场景中找到应用。未来版本或将扩展对更多语言的支持，尤其是像中文这样基于字符的语言，其字符尺寸的统一性可能使隐藏信息更为自然。随着 AI 文本识别能力的演进，这类实验性字体也为检验机器感知极限提供了有价值的参考。
Decoy Font is a novel typeface designed to obscure text from artificial intelligence by leveraging the principles of hybrid imagery. By utilizing different spatial frequencies, the font communicates two distinct letters within the same visual space. The foreground features thin, clearly defined outlines, while the background consists of a low-frequency, blurred mass. This optical illusion ensures that the perceived content changes depending on the viewer's distance or focus, with the human eye easily discerning the hidden message when viewing the text from further away or through a squint.
The effectiveness of this font lies in the way most AI systems process visual information. Large language models and OCR technologies typically analyze pixels at close range, where they prioritize the high-frequency foreground text that is most clearly outlined. Consequently, when an image created with Decoy Font is submitted to advanced models like ChatGPT or Gemini 3.5, the AI often identifies only the decoy text, failing to recognize the intended hidden message that a human would naturally see.
Beyond serving as a static image generator, Decoy Font is available as a functional TTF file, allowing users to type and use the font directly in various projects. This accessibility makes it a practical tool for those looking to protect their information or intellectual property from automated AI scrapers. While it is not a foolproof security measure, as advanced agents or specifically prompted models might eventually decipher the trick, it serves as a highly effective initial layer of confusion against casual observation and standard AI extraction techniques.
The project is part of a broader exploration at Mixfont into the intersection of typography and artificial intelligence. By applying spatial frequency techniques, the creators have developed a method that is more accessible than video-based obfuscation, potentially opening doors for applications in areas like CAPTCHA security or private messaging. Future iterations could involve expanding language support, particularly for character-based languages like Chinese, where the uniformity of character sizes could make hiding messages even more seamless. As AI text recognition continues to evolve, these experimental fonts provide a valuable benchmark for testing the limits of machine perception.
141 comments • Comments Link
• 该字体是一种视觉实验：根据观察者的注视位置或图像分辨率显示两条不同的信息，一条以清晰轮廓呈现，另一条以模糊阴影显现。
• 当前的视觉模型，包括最先进的 LLMs，通常能成功解读这两层信息，这表明基于频率的图像模糊并不能稳健地阻止 AI scrapers 。
• 该项目更应被视为创意艺术作品，而非真正的安全工具，因为它在阻碍 AI 或人类读取方面效果有限，且极易被绕过。
• 在公共网站上以此类技术做功能性模糊可能会引发关于无障碍（accessibility）的投诉，因为这种设计本质上妨碍部分用户的阅读体验，并使标准的屏幕阅读器（screen reader）处理变得复杂。
• 公众对这些技术的兴趣反映出更广泛的对 AI 的抵制情绪，根源在于对不道德的数据抓取、低投入的创作自动化以及大型科技公司权力集中的不满。
• 对图像进行缩放或模糊相当于施加低通滤波，这意味着任何依赖多频信息的视觉设计都可以通过调整输入分辨率被轻易解读，成为图像处理流水线中的可预测模式。
• 此类项目在营销上常被称为"AI defenses"，但其实际能力与宣传往往脱节，使人对公开发布前的审查与验证流程产生怀疑。
• 如果未来的迭代通过物理媒介实现（例如带背光的 neon signs），或作为数据层面的编码而非单纯的视觉字体应用，可能会更有效。
• "人类与 AI"的框架往往掩盖一个事实：这些实验主要是在展示 computer vision 的局限性，而不是长期可行的数字隐私或 bot 防御方案。
• 尽管实用性有限，该项目仍是一次发人深省的感知练习，凸显出生物视觉与 vision-language models 在处理空间信息方式上的根本差异。
讨论的焦点在于将这种"decoy font"视为创意设计作品，还是视为有缺陷的 AI 安全尝试之间的张力。大多数参与者一致认为，尽管这种光学错觉在技术上巧妙并具有审美趣味，但它作为阻止复杂视觉模型的可靠机制是失败的，因为这些模型通常能解析两层文本。许多贡献者对把简单的视觉技巧包装成针对 AI 的革命性"resistance"表示挫败，指出这些方法常常忽视无障碍标准和实际部署问题。归根结底，大家的共识是：该项目作为一次异想天开、发人深省的感知实验是成功的，但未能实现其作为对 autonomous scrapers 与自动化内容摄入的真正威慑这一既定目标。 • The font functions as a visual experiment that displays two distinct messages depending on the viewer's focus or the image resolution, using sharp contours for one message and blurred shadows for another.
• Current vision models, including state-of-the-art LLMs, often successfully decode both messages, demonstrating that frequency-based image obfuscation is not a robust method for blocking AI scrapers.
• The project is best viewed as a piece of creative art rather than a functional security tool, as its utility for thwarting AI or human-readable communication is minimal and easily bypassed.
• Attempting to use such techniques for functional obfuscation on public websites could trigger accessibility complaints, as the design inherently hinders readability for some human users and complicates standard screen reader experiences.
• Public interest in these techniques reflects a broader trend of resistance toward AI, stemming from dissatisfaction with unethical data scraping, the low-effort automation of creative tasks, and the centralization of power within large tech corporations.
• Scaling or blurring an image acts as a low-pass filter, meaning any design relying on multi-frequency content can be easily interpreted by adjusting input resolution, making it a predictable pattern for image-processing pipelines.
• There is a disconnect between the marketing of such projects as "AI defenses" and their actual capabilities, leading to skepticism regarding the vetting and validation processes conducted before public release.
• Future iterations could potentially be more effective if implemented through physical media, such as neon signs with backlighting, or if applied as a data-level encoding rather than a visual font.
• The "human versus AI" framing often obscures the reality that these experiments function primarily as demonstrations of computer vision limitations rather than viable long-term solutions for digital privacy or bot prevention.
• Despite the limited functional utility, the project serves as a stimulating exercise in perception, highlighting the fundamental differences in how biological vision and digital vision-language models process spatial information.
The discussion centers on the tension between viewing this "decoy font" as a creative design project versus a flawed attempt at AI security. Most participants agree that while the optical illusion is technically clever and aesthetically interesting, it fails as a reliable mechanism to block sophisticated vision models, which can typically resolve both layers of text. Many contributors express frustration with the trend of framing simple visual tricks as revolutionary "resistance" against AI, noting that these approaches often ignore accessibility standards and practical implementation issues. Ultimately, the consensus suggests that the project succeeds as a whimsical, thought-provoking experiment in perception, but falls short of its stated goal of providing a genuine deterrent against autonomous scrapers and automated content ingestion.