Shield Font - anti ai web text scraping via opentype ligatures

Dave Crossland

shieldfont.org/white-paper looks similar to the recent decoy font and older tempest and font enigma

Johannes Neumeier

We (as a society) ought to solve this on a non technical layer. And if it must be solved technically, then solve it at a level way before the font rendering.

Without commenting on this particular project, I can't help but feel like these font-to-solve-AI-training projects are nothing but bumper stickers on our digital roads.

Qwerasd

I find it problematic that any tool meant to prevent AI scraping of web content by altering the page for everyone will inherently ruin accessibility for users that rely on screen readers. Anything that allows a screen reader user to properly understand and navigate a page will allow an AI scraper to do the same. The solution that "Shield Font" offers to get around this little issue is simply a client-side proof-of-work... which is already a widespread way to combat bot traffic. I would link the section of the docs that explains this, but it's AI-written text, because this whole project is heavily vibe-coded, so I don't much care to link to it, as I find it insulting to ask people to read text spit out by the slop machine.

I agree with Johannes that the solution to this problem doesn't exist on the technical layer, and that there are layers far before the rendering layer that would solve it much better even in the technical layer.

There are so many ways in which this particular solution is flawed, aside from the accessibility problems it introduces (and then solves with a completely different solution which creates all of its own problems...)- the main glaring problem to me is that sometimes web fonts don't load. If the font doesn't load, the page will be gibberish to everyone, not just those who rely on assistive technologies to read the page. There are countless other problems, but if you're relying on your font loading for your website to be usable, you're doing web dev wrong and should get a different job.

Perhaps I'm being harsh, I'm sure the creator of the project has their heart in the right place, but I think the solution they came up with is trying to be far too clever for its own good.

Any solution to scraping that makes the experience substantively worse for many or all real humans -- while ultimately only serving as a temporary roadblock for scrapers until they come up with a workaround, which they will if a given solution achieves widespread use -- is immediately out of the question in my book.

John Savard

It definitely is true that making a web page not contain text in the ordinary sense will interfere with normal screen readers. But a technical solution is possible.

Use a technology similar to that which is used in HDMI to connect digital video to legitimate display devices, while video recorders can only record unencrypted signals. So screen readers for the visually handicapped would be given keys which would not be shared with web scrapers!

Of course, now this means that people using the shield font technology would also have to transmit encrypted versions of the unshielded web page on request, which limits their hosting options. Perhaps some bright people can figure out how to cope with this.

I've come up with one idea myself. On a web page, one can specify alt text for images. Well, an addition to the HTML standard could be made by means of which an alt audio file could be specified for every word or phrase that is scrambled by the shield font! So a screen reader would just pass the alt audio file through instead of reading the text whenever there is one.

However, one would have to endure the inconvenience of not being able to give the audio file the real word it contains as a filename, because doing so would make scraping trivial (detect the alt tag, use the filename before the extension instead).

SCarewe

That is in no way how screen readers work.

Simon Cozens

I tend to assume that the purpose of these "anti AI fonts" is not actually to prevent AI from reading the text - changing the font is not going to prevent scraping - but rather to show off interesting tech (color fonts, contextual ligatures) and optical illusions. And that's fine and cool and very clever in itself, and putting "AI" in the description is a good way to catch a bit of zeitgeist and interest. In other words, just like everything else AI-related, it's well marketed but not actually fit for purpose. 

SCarewe

Actually, in this case, it scrambles the underlying text, which indeed makes the text useless to AI, as it doesn't make any sense (to a human). The GSUB is just to make it display again correctly to a human.

SCarewe

What I do find ironic is that the white paper seems to be entirely written by AI (specifically, Claude). The website is also clearly built with Claude Code (it uses the exact same structure and overline tags as it did on a project I used it on).

Peter Constable

SCarewe said:

Actually, in this case, it scrambles the underlying text, which indeed makes the text useless to AI, as it doesn't make any sense (to a human). The GSUB is just to make it display again correctly to a human.

Models can also be trained on images of text.

I suspect that improvements to LLMs going forward will depend on other factors more than on new text from the Web that might get "shielded". So, I question the efficacy toward the intended goal. Not to mention the hindrance to accessibility tools, as already mentioned, and to Web search. (Imagine where we'd be today if we hadn't had a searchable Web for the past 30 years.)

SCarewe

Yes, the project mentions that – they know it's not a complete safeguard, more a way to drastically increase necessary workload to derive sensible data.