This Font Makes Webpages Unreadable to AI Scrapers, but Normal for Humans

Key points

• ShieldFont, created by designers Isaque Seneda and Gabriel Abrucio, replaces real words with nonsense words in the underlying code that AI scrapers collect.
• Human visitors see a normal, readable page because the word-swapping only happens after the font draws the text on screen.
• The font uses a feature called ligatures, which most fonts already include, to trigger the substitutions.
• The designers published a white paper describing ShieldFont as a practical way for website owners to block unauthorised AI training data collection.

Two designers have built a font that pulls off a quiet trick: show one thing to human readers, feed something completely different to the software AI companies use to hoover up text from the web.

The font is called ShieldFont. Designers Isaque Seneda and Gabriel Abrucio describe it in a recently published white paper as a tool to give web publishers "a practical opt-out from unauthorized AI training" and to "disrupt what is collected when that choice is ignored."

How does it actually work?

ShieldFont scrambles the words AI scrapers collect, but only after a human's browser has already drawn the page, so the reader never notices anything wrong.

The trick relies on a long-standing font feature called ligatures. In most fonts, ligatures are modest, useful things: they tidy up certain letter combinations that look awkward side by side. ShieldFont repurposes that same mechanism to do something far cheekier. Instead of tidying up two letters, it swaps entire words for different, nonsensical ones the moment the font engine renders the page.

Here is the key detail. AI companies collect training data by sending automated programs, called scrapers or crawlers, to download the raw source code of webpages. That source code is collected before a browser does anything with it. ShieldFont's substitutions only fire during the drawing step, so the scraper grabs the original, doctored text. A human visiting the same page sees the correctly rendered version.

In plain terms: the page says "horse" in the code the scraper copies, but your browser draws "potato" on screen. Or the other way around. Either way, the scraper walks away with garbage.

Does this actually stop AI companies?

ShieldFont makes collected text far less useful, but it is one layer of defence, not a complete shield.

AI models, large language models being the technology behind chatbots like ChatGPT and Claude, are trained on enormous amounts of text scraped from the public web. Publishers and authors have already sued several AI companies over this practice, and some sites have added technical instructions telling crawlers to stay away. ShieldFont offers a different approach: rather than ask scrapers to leave, it poisons what they take if they ignore the request.

The caveat, as reported by Ars Technica, is that a determined scraper could potentially use a headless browser, software that silently runs a full browser in the background, to collect text after rendering rather than before. ShieldFont's protection would collapse in that scenario. The designers acknowledge the font works best as part of a broader strategy, not as a standalone fix.

For website owners worried about their content being used to train AI without permission, ShieldFont is a low-effort extra step worth watching. It requires no server-side changes and no user action at all.

Common questions

Will this affect how my website looks to visitors?

No. Readers see a completely normal page. The word substitutions only exist in the raw code, which ordinary visitors never look at.

Could AI companies work around this?

Potentially yes. A scraper that runs a full browser before collecting text would bypass the protection. ShieldFont is most effective combined with other measures, such as robots.txt rules that instruct crawlers to stay away.