Via creativebloq.com
A new open-source font uses clever glyph substitution to poison the well for unauthorized AI training data collection, and early results suggest it actually works.
Two designers got tired of watching AI companies hoover up the entire internet for training data and decided to fight back with, of all things, a font. ShieldFont, created by Isaque Seneda and Gabriel Abrucio, renders perfectly readable text for human visitors while embedding subtly altered, meaning-mangled words in the underlying HTML that scrapers actually parse.
How the trick works #
ShieldFont exploits a feature baked into modern font technology called OpenType ligature and glyph substitution. In plain terms, the font file tells the browser to display one set of characters while the source code underneath contains entirely different words. When a human looks at the page, their browser renders the intended text. When a scraper reads the raw HTML, it pulls the swapped-in gibberish instead.
The system swaps roughly 24.4% of all words on a protected page, which translates to about 45.8% of content words, the ones that actually carry meaning. That’s enough to thoroughly corrupt any dataset built from the scraped material.
According to the project’s white paper, titled “The Consent Layer” and released on July 30, 2026, over 90% of pages using ShieldFont get flagged and rejected by automated quality filters. Even when scraped content does slip through those filters, conceptual accuracy in model reconstructions drops by up to 67%.
Built for real-world adoption #
ShieldFont isn’t just a proof of concept. The project ships with an encoder tool and React components designed to make integration into existing web frameworks relatively painless for publishers. Development began in October 2025, with the team collaborating with Playtype, a Danish type foundry, to ensure the fonts meet professional typographic standards.
One particularly thoughtful detail: the system includes special handling for screen readers. Altered text is hidden from assistive technology using aria-hidden attributes, so visually impaired users still get the original, unmodified content.
The project also supports customizable word mappings and private font configurations. Publishers can choose which substitutions to make, tailoring the corruption to their specific content and audience.
ShieldFont is open-source, which means anyone can inspect, modify, or deploy it without licensing fees.
The bigger picture for content protection #
AI companies have faced a growing wave of lawsuits from publishers, authors, and artists who argue that scraping copyrighted content for training purposes amounts to theft. Technical countermeasures like robots.txt directives and rate limiting have proven easy to circumvent. ShieldFont represents a fundamentally different approach: instead of trying to block scrapers at the door, it lets them in and feeds them garbage.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our