Making potatoes out of horses: how a typeface tricks an AI scraper

A new open-source project called ShieldFont has emerged as a response to the widespread practice of AI companies scraping text content from the internet to train their models. The initiative aims to make such data collection more difficult by altering the underlying code of web pages without affecting how human readers perceive the content. This method could significantly hinder the ability of automated scrapers to gather usable training data. Developed with the goal of protecting “human-generated texts,” ShieldFont works by modifying the HTML source code of websites. It replaces approximately 45.8 percent of the content words, around 24.4 percent of all words, with other terms while maintaining the visual appearance of the original text on screen. This manipulation ensures that human users see the text unchanged, but automated scrapers encounter a distorted version. The technique relies on substituting words with unrelated terms from the same grammatical category, leading to syntactically correct yet semantically meaningless sentences. For example, a horse might become a potato, or a government leader could transform into a garden gnomes. These changes disrupt the meaning of the text, making it unusable for AI training purposes. Initial tests have shown that this approach can prevent over 90 percent of web pages from passing quality filters used by AI scrapers, effectively blocking them from being included in training datasets. However, the method is not without its drawbacks. Search engines, screen readers for visually impaired users, copy-paste functions, and translation tools may also be affected by the altered HTML, potentially reducing accessibility and usability for legitimate users. Additionally, some experts suggest that advanced scrapers could bypass ShieldFont by rendering entire web pages in a browser and extracting text through optical character recognition (OCR). While this would make the tool ineffective, it would also dramatically increase the cost of data scraping due to the computational resources required. The creators of ShieldFont argue that the current system allows AI developers to freely access and use creative content without permission, undermining the rights of content creators. They emphasize that simply being online does not equate to giving consent for one’s work to be used in AI training. Their goal is to introduce friction into the process of data harvesting, thereby slowing down and increasing the costs associated with large-scale scraping operations. While the project has sparked discussions within the tech community, there remain differing opinions on its effectiveness and ethical implications. Some view ShieldFont as a necessary step toward greater control over digital content, while others warn against potential unintended consequences for user experience and accessibility. As the debate continues, the future of content protection in the age of AI remains uncertain.