Isaque Seneda and Gabriel Abrucio, designers who developed ShieldFont, say the font lets web publishers present pages that appear unchanged to readers while embedding altered HTML that changes the meaning of the text for AI scrapers that parse raw source code.

AI companies have increasingly scraped publicly accessible webpages to collect training data, leading to lawsuits and technical attempts to block unauthorized use.

ShieldFont works by substituting whole words with other words that share part‑of‑speech characteristics but convey different meanings, such as replacing “horse” with “potato”. The font applies these substitutions only when the page is rendered, so human readers see the original text while scrapers that download plain HTML receive the altered version. In tests on six common scraper pipelines, more than 90 % of pages that would otherwise be accepted were rejected after the word replacements, and about 20 % of accepted pages contained altered words that the authors describe as “training‑time garbage”.

The creators say their aim is to give authors a technical means to enforce consent over the use of their work in AI training, and they expect other developers to build similar mechanisms that present different content to machines and humans. ShieldFont is not a complete solution, as pages rendered by browsers can still be captured by OCR‑based scrapers, but it raises the cost and effort required for large‑scale scraping.