Ghost Font: A font that humans can read but AI cannot
241 points
• 6 days ago
• Article
Link
Ghost Font 是一种实验性的反人工智能通信工具，它通过运动而非传统字形来传达信息。项目将运动、噪声、视频和诱饵信息结合，生成一种人眼可读但对现有 AI 模型极具挑战性的格式。文本由移动的点组成，静止后会与背景融为一体，因此单帧静态图片或截图看不见任何内容，使得基于图像的常规分析无法识别信息。
该项目受 2013 年 ZXX 字体启发，后者原本用于规避监控和光学字符识别。不过 ZXX 倚仗视觉噪声和删除线的做法已被当代 AI 轻易破解。 Ghost Font 通过引入时间维度的运动来应对这一演进，绕过了那些习惯把视频拆成静态帧来处理的模型。
为进一步保护内容，系统引入了诱饵信息。当高级 AI 代理试图解读视频时，常会抓住这些假信号，导致模型输出错误结果或产生并不存在的幻觉数据。在对 Claude Fable 和 GPT Sol 5.6 Ultra 等先进模型的测试中，该技术也能在未被告知解密方法的情况下成功掩盖真实信息。
尽管当前效果显著，创作者也承认 Ghost Font 不能替代传统加密——后者仍是抵御 AI 的唯一万无一失的方法。此项目更像是对现有 AI 感知极限的探索，也是面对设计与字体生成日益自动化的一种创造性反击。它凸显了人类视觉与机器视觉之间差距的缩小，因为即使这种基于运动的方法，在被机器解码和保证人类可读性方面也存在困难。
展望未来，这些概念有望应用于 CAPTCHA 开发等领域——传统方法在对抗自动化机器人方面正愈发失效。该项目也可作为评估不断演进的多模态 AI 视觉感知能力的实用基准。团队计划开源视频生成代码并扩展对更长文本串的支持，表明这一实验将继续作为研究以人为本的设计与机器智能之间张力的平台。
Ghost Font serves as an experimental anti-AI communication tool that relies on motion rather than traditional typeface design. By combining motion, noise, video, and decoy messaging, the project creates a format that is readable to the human eye but presents significant challenges for current AI models. Because the text is composed of moving dots that blend into the background when stationary, a single static frame or screenshot reveals nothing, rendering the message invisible to standard image-based analysis.
This project draws inspiration from the 2013 ZXX font, which was originally intended to bypass surveillance and optical character recognition. However, while ZXX relied on visual noise and cross-outs, modern AI has advanced to the point where it can easily decipher those techniques. Ghost Font attempts to stay ahead of this evolution by utilizing temporal motion that eludes models that primarily process information by splitting videos into individual static frames.
To further safeguard the content, the system incorporates decoy messages. When advanced AI agents attempt to interpret the video, they often latch onto these false signals, leading the models to report incorrect information or hallucinate data that does not exist. Even when tested against sophisticated models like Claude Fable and GPT Sol 5.6 Ultra, the technology successfully obscured the true message unless the AI was specifically prompted with the decryption technique.
Despite its current effectiveness, the creator acknowledges that Ghost Font is not a replacement for traditional encryption, which remains the only definitive way to secure data against AI. Instead, the project serves as an exploration of the limits of current AI perception and a creative pushback as automation becomes increasingly embedded in design and font generation. It highlights the closing gap between human and machine vision, as even this motion-based approach faces hurdles in both machine decoding and human legibility.
Looking toward the future, there is potential to apply these concepts to fields like CAPTCHA development, where traditional methods are increasingly failing against automated bots. Furthermore, the project serves as a practical benchmark for testing the visual perception capabilities of evolving multimodal AI models. Plans to open-source the video generation code and expand the capacity for longer text strings suggest that this experiment will continue to serve as a platform for studying the ongoing tension between human-centric design and machine intelligence.
173 comments • Comments Link
• Ghost Font 利用时间运动而非空间模式来隐藏文本：在静态帧中看不见，但在噪声背景滚动时可被察觉。
• 许多人难以发现隐藏文本，常把静态的"诱饵"信息误认为是预期内容，从而使对人类可读性的评估复杂化。
• AI 模型在孤立条件下往往无法解码隐藏信息，常出现幻觉或只识别静态诱饵，除非明确指示其分析时间运动或光流。
• 使用标准计算机视觉方法（例如先用相位相关估计背景位移，再做帧差分）可以很容易地技术性解码该字体。
• 该技术的有效性高度依赖观察者的屏幕设置、设备分辨率和个人视觉差异，导致用户体验不一致。
• 将此类方法用于安全目的（如 CAPTCHAs）受到批评，被视为一种"通过隐蔽性实现的安全"，并对视障用户造成重大无障碍挑战。
• 虽然被称为"字体"，但在功能上更像视频特效或隐写术，因为它缺乏数字字体文件的标准特征，且需要动态播放。
• 诱饵信息与目标信息之间的混淆凸显出在交互机制非标准化的实验中保持设计清晰的重要性。
• 对这类方法的探索展示了以人为中心的通信与自动检测之间的持续军备竞赛，普遍共识是任何此类方案最终都会被针对性的预处理算法攻破。
• 爱好者将该项目视为关于基于运动的通信的有趣研究实验，尽管其实际应用受制于无障碍性、可靠性和计算实现的限制。
讨论集中在 Ghost Font 的效果与实用性上——这是一种在滚动视频噪声中隐藏文本的方法。一个反复出现的争议点是人类可读性与机器感知之间的差异。作者将其定位为一种防 AI 的格式，但观察者指出该技术很容易被简单的计算机视觉算法或运动差分分析绕过。大部分讨论被关于显示内容中哪一部分是真实信息、哪一部分是诱饵的混淆所主导，导致 AI 和人类评估者经常无法识别预期信息。最终的共识是，尽管该项目作为时间隐写术的研究实验具有吸引力，但它存在严重的无障碍问题，且无法作为对抗现代自动化系统的有效守门手段。 • Ghost Font uses temporal motion rather than spatial patterns to hide text, rendering it invisible in static frames but perceivable when the noise background scrolls.
• Many individuals struggle to perceive the hidden text, often mistaking a static "decoy" message for the intended content, which complicates the evaluation of human-readability.
• AI models frequently fail to decode the hidden message in isolation, often hallucinating or identifying the static decoy text instead, unless explicitly instructed to analyze the temporal motion or optical flow.
• Technical decoding of the font is straightforward using standard computer vision techniques, such as phase correlation to estimate background shift followed by frame differencing.
• The effectiveness of this technique relies heavily on the observer's screen settings, device resolution, and personal visual perception, leading to inconsistent user experiences.
• Relying on such methods for security purposes like CAPTCHAs is criticized for being a form of security through obscurity and for posing significant accessibility challenges for users with visual impairments.
• While described as a "font," the technique is functionally a video effect or steganography, as it lacks the standard characteristics of a digital font file and requires dynamic playback.
• The confusion between decoy and intended messages highlights the importance of design clarity in experiments where the mechanism of interaction is non-standard.
• Exploring such methods illustrates an ongoing arms race between human-centric communication and automated detection, though the consensus suggests that any such scheme will eventually be defeated by targeted algorithmic preprocessing.
• Enthusiasts view the project as an interesting research experiment into motion-based communication, even if its practical application remains limited by accessibility, reliability, and computational triviality.
The discussion centers on the efficacy and utility of "Ghost Font," a method for hiding text within scrolling video noise. A recurring point of contention is the distinction between human readability and machine perception; while the author positions it as an AI-proof format, observers point out that the technique is easily bypassed by simple computer vision algorithms or differential motion analysis. Much of the discourse is dominated by confusion regarding which part of the displayed content constitutes the "real" message versus the "decoy" text, leading to frequent instances where AI and human evaluators alike fail to identify the intended information. Ultimately, the consensus is that while the project serves as a compelling research experiment in temporal steganography, it suffers from significant accessibility issues and serves as an ineffective gatekeeper against modern automated systems.