Poisoned source: open source project cheats AI scrapers and protects content

A creative team and a typography company have developed a source that cheats AI scrapers, causing them to read gibberish instead of real content. The technique, available as open source, offers a new layer of protection against unauthorized training of language models, without damaging the experience of human readers.

Poisoned source: How do you cheat AI scrapers?

The source encodes characters so that, for human eyes, the text remains readable — but for optical recognition (OCR) algorithms or scraping models, the letters are interpreted as random and meaningless sequences. The trick explores differences in the way humans and machines process glyphs: while the human brain fills gaps and recognizes patterns even with stylized sources, traditional scrapers depend on exact correspondence in shapes, and end up capturing only visual waste.

Why does the poisoned source matter to creators?

With the unbridled growth of AI model training on scraped data from the web, many creators seek ways to prevent their work from being used without permission. Robot locks. txt or catchas are easily bypassed. The poisoned source represents a more subtle and low-cost technical barrier. Any website can host the open source font and apply it via CSS, without having to modify the server. For the Brazilian universe of blogs, independent portals and content producers, the solution is especially relevant at a time when Codeberg prohibits generational AI projects and the community seeks alternatives to protect their data.

Poisoned source: limitations of the technique

The defense is not absolute. More sophisticated scrapers using advanced OCR or interpreting the DOM directly (instead of rendering the source) can still extract the actual text. In addition, the source needs to be updated as AI models evolve. The project team, in partnership with the printing company, released the code under permissive license to encourage contributions and adaptations. In an interview with The Register, developers point out that the tool is a complement, not a silver bullet, and should be combined with other security practices.

Frequently Asked Questions about Poisoned Source AI

Does the poisoned source affect website accessibility?

Yes, there is a risk for screen readers and accessibility tools that depend on OCR. Developers recommend testing the source with these tools and offering alternatives via aria-label or alt text attributes.

Is it necessary to install something on the server to use?

Nope. Simply include the source file via CSS (link to CDN or self-hosting) and apply the poisoned-font class to the HTML elements you want to protect. The server does not need changes.

Does the technique work against any AI scraper?

Works against scrapers that are based on screenshot or visual rendering to extract text. Scrapers that directly access DOM content or use high-precision OCR APIs may not be fooled.