Novel font weaponizes word substitution to confound AI training scrapers

Designers are pushing back against the widespread practice of harvesting web content for AI model training—this time with typography. A newly designed font called ShieldFont uses an old typographic trick to serve readers legible pages while rendering the underlying HTML code as corrupted nonsense to any algorithm attempting to scrape it.

The arms race against AI data collection#

The competition between AI companies and content creators over training data has intensified. Major firms have scraped vast regions of the public web seeking material to power their models, triggering legal action and various technical countermeasures. ShieldFont, created by designers Isaque Seneda and Gabriel Abrucio, offers publishers another option: a way to shield themselves from unauthorized data extraction while maintaining a normal reading experience for human visitors.

How ligatures become a defense mechanism#

The approach leans on ligatures—a centuries-old font feature that traditionally combines adjacent letters into more elegant or readable forms. ShieldFont repurposes this mechanism to swap whole words with alternatives as the browser renders the page. Crucially, the substitution happens only in the visual layer; the underlying HTML source remains unchanged. This creates a split: humans see a coherent article, while bots downloading the raw code encounter garbled text.

The technical challenge lies in calibration. Simple synonym or antonym swaps invite reversal by sophisticated scrapers; random gibberish may be too obviously corrupted and easier to filter out. Designers must strike a balance that confuses AI systems without alerting them to the manipulation.