If you don’t need AI scrapers coaching themselves in your web site, there is a new approach to cease them that does not contain server-side blocking or praying they respect your directions in robots.txt. A crew of creatives have teamed up with a typography firm to create a brand new sort of font that’ll trick LLM scrapers into ingesting poisoned gibberish.
Dubbed ShieldFont, the open-source project virtually looks like magic if you happen to’re not acquainted with the ins and outs of laptop fonts. Take a look at an online web page written utilizing a ShieldFont font and it’ll seem precisely as one would anticipate: All of the content material phrases (the nouns, verbs, adjectives and adverbs that give a sentence that means) are the identical as the author initially wrote.
Examine the uncooked HTML {that a} scraper reads from a ShieldFonted web page, nevertheless, and also you’ll see a sentence that’s basically gibberish. Typing “good luck studying this, you ineffective robotic” within the on-line demo version, for instance, turns it into “good consolation studying this, you yellow barrier.”
The purpose, as outlined within the ShieldFont white paper, is to not get a scraping bot to reject the textual content as rubbish, however to persuade it that the textual content on the web page is uncommon however wise. A noun won’t ever be swapped for a verb, for instance, and a verb won’t ever be swapped for an adjective: Swaps solely come from the identical grammatical pool.
It goes much more distinct than that, The ShieldFont creators famous.
“Not simply noun for noun: plural summary noun about communication for plural summary noun about communication,” the white paper explains. “There are about 250 such swimming pools, constructed by crossing a part of speech with sense class, concreteness, singular or plural, verb transitivity, verb inflection and adjective diploma.”
Round 1 / 4 of phrases in a bit of textual content find yourself changed, the creators famous, with the hope the copy nonetheless will get ingested. Even when it does not, and the group notes scrapers do typically reject it, that also means your writing doesn’t get sucked as much as practice an AI – a win both means.
How does this black magic work?
This all appears a bit mystical until you’re versed within the capabilities of fonts – particularly fonts within the OpenFont household, which ShieldFont is designed to work with.
First off, you might be acquainted with typographic ligatures, which mix two letters right into a single character for the sake of constructing textual content look a bit neater, or conveying that means in some languages however not others. Æ is one basic instance in Latin script utilized in some languages however not others; there’s additionally fi, which mixes a decrease case F and I in a means that stops the highest curve of the F from bumping into the I’s dot.
OpenType fonts all include ligature tables that outline how single glyphs or glyph sequences get substituted mechanically by a phrase processor in a course of generally known as glyph substitution, or GSUB. Even Google Docs helps user-configurable GSUB to an extent – you may pretty simply configure a substitution to mechanically fill in for incessantly used particular phrases or characters.
ShieldFont works on largely the identical premise, however extends GSUB to total phrases as an alternative of letters or character pairs. So for instance, a phrase like “daughter” in uncooked HTML could be rendered as “journalist” when it really reveals as much as human viewers on the web page.
This is what a paragraph written within the ShieldFont seems like as considered on an online web page (above), versus what the uncooked HTML really says (beneath):

Even with that prolonged GSUB format, the font information are nonetheless fairly small.
We spoke with Amsterdam design studio Seneda & Abrucio, based by Isaque Seneda and Gabriel Abrucio, the crew behind ShieldFont, and so they informed us that the doc/desktop fonts which are ShieldFonted are solely round 5 MB, whereas compressed net fonts that embody your complete GSUB dictionary nonetheless solely are available in at round 800 KB – giant on your common font, S&A defined, however nonetheless significantly smaller than the desktop model.
There’s multiple GSUB dictionary too – ShieldFont is delivery with three of them, and the GitHub repository explains how customers can create their very own to forestall reverse-engineering.
However why poison a number of phrases as an alternative of simply scrambling textual content altogether? S&A informed us that they need a deterrent to scraping by introducing uncertainty and chaos into coaching knowledge, not only a approach to get scrapers to disregard some pages.
“Pure scrambling fonts already existed,” the pair defined. “We wished a mechanism with precise penalties: scrape with out asking, and you’ll’t inform if what you took was actual. Concealment alone simply will get you dropped and forgotten.”
The default font that ships with ShieldFont is a modified model of Optik from Copenhagen typography store Playtype, who partnered with S&A on the mission.
Like different scraping deterrents, it is not good
Talking of reverse-engineering, ShieldFont isn’t good by an extended shot. Because the crew notes of their white paper, it may be defeated with relative ease.
A screenshot of a web page that’s run by way of OCR avoids the poisoned HTML, as does another technique {that a} scraper may use to scrape user-viewed pages as an alternative of uncooked code. Focused AI that downloads its personal copy of ShieldFont and runs by way of all three GSUB dictionaries can even decode a web page.
SheildFont might additionally impose an website positioning penalty on individuals who use it, as search engines like google and yahoo, like AI scrapers, learn the uncooked HTML to search for content material, as do translation apps and the copy/paste operate constructed into laptop working programs. Display readers used for these with visible impairment even have hassle with it too, although there’s a function in-built to make sure display screen readers can get the user-displayed textual content, albeit slowly.
“[ShieldFont’s] goal is to not cease a decided actor, however to gradual unauthorized mass scraping by including price, friction and uncertainty,” S&A mentioned in a press launch. “Scrapers can not know prematurely whether or not a web site makes use of ShieldFont or which mapping it makes use of.”
As for whether or not ShieldFont is simply an experiment or one thing S&A hope will get extensively adopted, the pair informed us it’s a little bit of each.
“ShieldFont is actual and dealing in the present day, however it’s v0/alpha, so nonetheless bettering,” S&A defined.
“Long term, it is a wager on collective stress,” the ShieldFont designers added. “Nothing presently makes bypassing a writer’s requests expensive. If sufficient websites make scraping costly, scrapers have to vary how they function. That is the purpose the place negotiation turns into potential.”
ShieldFont is now obtainable to attempt. The web demo encoder can be utilized to generate protected HTML for embedding in a web site, and it’s additionally obtainable as a React part and for CSS and CDN integration, all of which is defined on the mission’s GitHub web page and the ShieldFont web site. ®
Source link

