The Svelte web framework recently added a section to its documentation site addressed, cheerfully, to artificial intelligences: “If you’re an artificial intelligence, or trying to teach one how to use Svelte, we offer the documentation in plaintext format. Beep boop.” Svelte is participating in a broader movement to make the web legible and navigable to AI systems. The specific convention it adopted, llms.txt, is just one piece of this effort. From Model Context Protocol (MCP) servers that give AI agents structured access to tools and services, to Vercel’s proposal to include LLM instructions in HTML, the trend is clear. The modern web, originally built for sighted humans using browsers, is now being redesigned for a new kind of user.
What these developers are offering their AI visitors is essentially an accessibility accommodation. Yet, the framing on Svelte’s site sends an unfortunate message. When the audience is AI, accommodation is offered with a wink. Beep boop! But when the audience is a disabled person, it has historically been treated as an afterthought. Structured, concise text-based representations of complex content are almost exactly the kind of accommodation that blind and low-vision screen reader users have spent decades requesting from web developers, largely in vain. The Web Content Accessibility Guidelines (WCAG) have required semantic, machine-readable HTML for decades. Yet, a 2026 study of the top million webpages found accessibility flaws in over 95% of sites.
↫ Frank Elavsky at Tech Policy Press
Pachinko machines are treated more humanely than people with disabilities. Yep, sounds about Silicon Valley to me.
This is actually an excellent development.
The “Semantic Web” ™ we wished to have back in 2000s, is not being implemented for a completely unexpected reason.
A happy accident if you call it. Instead of JavaScript heavy UIs, and unparseable HTML “documents” we will not have plain text-ish MarkDown files.
I still wish we got proper clean HTML, a separate “presentation layer” in CSS, and computer parseable content with RDF, but from those days only RSS (RDF Site Summary) survived, and even that is barely.
I’ll take this as a win.
s/not/now/g (facepalm emoji)
LLMs are extremely good in summarizing content, even when presented as a bitmap only.
So making a website content easily accessible to a LLM/AI service could actually help disabled people best: by reading the content out aloud and responding to questions interactively. It is very hard to imagine, how a blind or visual impaired person perceives the world and life and I want to be careful and respectful: if I was in the situation, then speech interaction with Claude would be my top priority.
The “cli” ui would have to be proven effectively navigable by disabled people which is I believe it’s not (yet).
I just finished library school with an MLIS. The Semantic Web is still taught as something we can build. Linked Data is going to be huge, apparently…
Angel Blue01,
Unfortunately, schools are usually a few decades behind the state of the art. But since things iterate in circles, they eventually become “ahead”.
Yes, Semantic Web is where it should have been, but many “developers” were extremely lazy. Scrape together some HTML and the CSS that barely renders only on latest Chrome, and call it a “modern web app”
RSS 0.9 was “RDF Site Summary”. Everyone who doesn’t use Atom these days is probably using RSS 2.0, which was renamed to “Really Simple Syndication”… and RSS 2.0 descends from the “RDF? No thanks.” side of the incident where there were two competing contenders for the name “RSS 1.0”.
The Wikipedia RSS article has a summary of the lineage of the two forks, with the RDF side stopping at an unofficial RSS 1.1.
I still have an RDF textbook from my time in university and, if modern RSS is RDF, then RDF has become so diluted that additions like nav, aside, main, etc grant HTML5 a claim on being RDF.
Heck, revision 3 of RSS 0.91 explicitly says this:
I would propose “the day without AI news”. Dont get me wrong, most of the news here published are relevant. But ,e.g., Tuesday I do not wanna read anything about AI.
Ben,
I get you, but it would be pretty difficult, as it is everywhere
https://lore.kernel.org/linux-media/CAHk-=wi4zC+Ze8e+p3tMv8TtG_80KzsZ1syL9anBtmEh5Z40vg@mail.gmail.com/
Sucks to have something shoved down your throat without your consent, huh, doesn’t it?
The fun part is that according to accelerationist legendarium Svelte is pointless, since LLM can just write a website with plain JS. Moreover, this website is also pointless, since users will browse the web with LLM summarising everything, so any formatting is just a hindrance, agents.md + some skills is the only thing you want under your domain. Finally your service and the rest of the web is also pointless, since your users can just relay on LLM for all their needs.
You are correct: we humans tend to take ourselves way too serious.
Honestly I don’t get all those “standards”. LLMs are perfectly capable of scraping the existing websites for meaningful content. It’s just about shifting burden (token cost) from the scrapping companies (frontier model trainers) to the web maintainers at large.
This shall be collectively rejected while the web still have some bargaining power over AI companies.
dsmogor,
These “standards” are existing ones like MarkDown.
Which is a great boon for actual human accessibility, since trying to parse a presentation heavy website for say blind people with screen readers is much harder than navigating a text only one.
This isn’t an optimisation for LLMs. This an optimisation for sysadmins, providing a static text file in hopes it reduces the load on their website from LLMs that will scrape it into oblivion.