The big picture: Putting your writing on the open web used to mean people could read it. Now, it also means dozens of crawlers can copy it into training datasets. Sure, anti-scraping tools such as robots.txt exist, but they only work if the bots themselves agree to stay away, which doesn't always happen. Stricter measures were clearly needed, and now they have arrived in an innovative new form.
A Brazilian creative studio called Seneda & Abrucio has teamed up with Playtype, a Copenhagen-based type foundry, to build something that doesn't rely on asking nicely to keep crawlers away. They call it ShieldFont, and it's essentially a free, open-source web font with a twist. It works by showing you one sentence while presenting an AI scraper with a completely different one.
The thing is, you see rendered pixels on a screen. Most mass scrapers, on the other hand, simply grab the raw HTML underneath. ShieldFont exploits this difference through an automated process called OpenType glyph substitution. This technology is normally used to replace one or more typed characters with alternate glyphs that improve how text is rendered. In this case, however, entire words are swapped out, meaning a scraper can pick up only gibberish from the webpage.
Those swapped words are not picked at random, however. The studio's dictionary pairs every word with another of the same grammatical type, so nouns get nouns and past-tense verbs get past-tense verbs. The words are sorted into roughly 250 pools that account for factors such as whether a noun is abstract or plural. About a quarter of the words in any given block are swapped this way.
Keeping the grammar clean matters because AI firms run scraped text through quality filters that discard anything that reads like nonsense. Seneda & Abrucio ran shielded text through FineWeb-Edu, a quality filter used to assemble a large public training dataset, and found that about one in 10 passages that passed before shielding still passed afterward.
Whatever gets through is fluent enough to be retained yet wrong enough to be useless at the same time. In fact, 55.8% of shielded passages in the studio's testing no longer made the original factual claim.
That said, because the whole defense rests on scrapers reading code rather than screens, taking a screenshot of a shielded page and running OCR on the image can still recover the real words. Screen readers used by blind readers also work from the code, so they read the decoys aloud. ShieldFont ships with a beta feature that provides those readers with the real text instead.
For now, ShieldFont only handles English. The code can be found on GitHub for anyone who wants it. Developers and writers can simply install it as a React component to their websites.