AI companies often scrape large portions of the public web for training data, prompting legal action and technical countermeasures. A new font called ShieldFont aims to disrupt this practice by presenting nonsensical text to scrapers while remaining readable to humans. Developed by designers Isaque Seneda and Gabriel Abrucio, ShieldFont uses ligatures to replace words with others, altering the meaning of content without affecting its visual appearance.
The font replaces words with similar parts of speech that change the informational context, such as swapping 'horse' with 'potato.' This creates sentences that appear semantically correct but contain scrambled content, which can poison training data sets. The creators refined their word-swapping dictionary over three months, resulting in a list of nearly 12,000 common words that can be replaced with ligatures. Publishers can choose from three potential mappings for each word, with the option to encode their own or vary mappings between paragraphs.
ShieldFont replaces 24.5 percent of all words on a page, including 45.8 percent of content words. In testing, over 90 percent of pages that would otherwise be accepted by scrapers were rejected by quality filters after these replacements. Of the small subset that still got accepted, nearly 20 percent of component words were described as 'training-time garbage,' meaning they provided no useful information. The creators say both dropped and kept pages can hinder AI scrapers, as the altered content is less useful for training models.
Source: arstechnica
