In the mid-2000s, a developer needed a checkbox and didn't use HTML's checkbox. Native form controls were hard to style and rendered differently in every browser, so instead: a generic <span>, an image of an empty box, and a few lines of script to swap the image on click. To someone with a mouse and working eyes, it was a checkbox. To a screen reader it was nothing, a span with a picture inside it, carrying no indication of what it was or that it could be used at all.
That was ordinary practice. JAWS, which arrived in 1995, worked by building a virtual buffer: an in-memory copy of the page that let a blind user move through it like a document. The copy was static. When JavaScript changed something, the buffer didn't catch up until the page reloaded, and users often suppressed refreshes because a refresh cost them their place. The screen reader was reverse-engineering a visual interface that had never been built to communicate its own structure to a machine.
If you work on web agents, you know this problem under a different name. An agent moving through a website to finish a task hits the same wall: the page communicates through layout, color, and conventions that assume an eye and a pointer. Where no cooperative machine interface exists, whoever needs one builds a workaround, and the workaround becomes the way it's done. Screen readers have been doing this since 1995.
What eventually moved the problem wasn't better parsing. In September 2006, W3C published the first working drafts of ARIA, Accessible Rich Internet Applications, a vocabulary that let developers declare an element's role, state, and properties in the markup, so a <span> impersonating a checkbox could say so. ARIA reached formal Recommendation status in March 2014. Eight years from draft to standard, with a dedicated standards body, an organized advocacy community, and statutory pressure behind the work. Agents reading arbitrary web pages today have no comparable vocabulary to fall back on.
A standard tells developers what to do. Adoption moved slowly enough that the courts arrived first. In 2006 the National Federation of the Blind sued Target, alleging that Target.com lacked text alternatives for images, labels for form fields, and keyboard operability for core functions. The 2008 settlement named the specific barriers to be removed, required training and continuing testing, and established a $6 million damages fund; Target later earned the Federation's accessibility certification. The gap had to acquire a number, a deadline, and a party responsible for closing it before it became an operations problem. Anyone watching how agent-readability develops should expect something structurally similar.
By the time WebAIM ran its 2026 scan of a million homepages, all of it was in place: the standard, the case law, two decades of developer education. The scan found automatically detectable accessibility failures on 95.9% of those pages, up from 94.8% the year before, with the same six error categories topping the list for seven consecutive years. ARIA attributes appeared on 82.7% of pages, and pages using ARIA averaged more detected failures than pages without it, which probably says more about where ARIA gets used than about ARIA. Adoption of the vocabulary ran well ahead of the skill to use it correctly.
Screen readers served users that companies were legally required to accommodate. Agents arrive as potential customers. That is a different kind of pressure, and in some respects a stronger one. The accessibility record is mostly a record of how little any single form of pressure accomplishes against millions of pages maintained by people with other deadlines.
Maybe revenue moves faster than litigation. A merchant who wants an agent to complete a purchase has a reason to publish structure that no accessibility audit ever managed to extract. But the pages that would need to become legible are maintained by an enormous and uncoordinated population of developers, and thirty years of accessibility work is the available evidence on how that goes, even with a standard and a body of case law behind it. The distance between what an interface shows and what it can tell a machine has been open that entire time. Agents are the newest arrivals to it.
-
Bot identity without mandate: The IETF's Web Bot Authentication working group is standardizing how automated clients prove who operates them, but its charter explicitly excludes authenticating the end user or conveying the bot's intent — a gap that echoes the accessibility split between identifying a machine consumer and making the interface actually work for it.
-
Preferences are not permissions: The IETF's AI Preferences working group is developing a vocabulary for sites to express usage preferences via robots.txt and HTTP headers, but enforcement, authorization, and auditing are all outside its charter — another machine-readable signal whose adoption and institutional weight remain separate questions.
-
API maintenance as ongoing cost: A peer-reviewed study of microservice API evolution across eleven companies found that most operated multiple API versions simultaneously, with fourteen of seventeen participants reporting quality degradation and ten describing significant overhead from retaining old versions — a reminder that cooperative machine interfaces create their own sustained maintenance burden.
-
Agent identity and delegation: NIST's 2026 concept paper on agentic AI identity asks how an agent can prove authority for a specific action and bind itself to human authorization, and its summary of public comments records calls for tamper-evident logs, cryptographic receipts, and verifiable evidence chains — the kind of institutional infrastructure that accessibility standards took decades to partially build.