The Web Is Being Made Accessible for AI, Not People

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.

The overlap between what AI agents need and what screen reader users need, however, is narrower than it appears. Screen reader users require not only plaintext, but also structure: heading hierarchies that allow jumping between sections, landmark regions, descriptive link text, and alt text for images. WCAG requires semantic HTML because screen reader users navigate content sequentially and depend on structural cues to find specific information.

A format like llms.txt goes in the opposite direction, flattening documentation into undifferentiated plaintext that a language model can ingest wholesale but that offers a screen reader user nothing to navigate by. And as large language models gain the ability to process images natively, AI-optimized pages have diminishing incentive to include alt text at all—even though blind users still depend on it. The risk is not only that these accommodations fail to help disabled users, but also that developers begin treating "machine-readable" as synonymous with "accessible," checking a box that was never truly checked.

This is not the curb cut effect—the oft-referenced idea that accommodations designed for disabled people ultimately benefit everyone. It is closer to the inverse: an accommodation built to serve a well-capitalized technological system that incidentally happens to address a need disabled people have long articulated without sufficient response.

It’s easy to forget that the curb cut, now widely taken for granted, was not given freely. In the early 1970s, disabled activists in Berkeley, California, wheeled themselves to intersections at night and poured their own cement ramps. Police threatened them with arrest. Though the curb cut effect is often invoked as a feel-good story about universal design, the original curb cuts were won through direct action, litigation, and political organizing.

Disability scholars have argued that the curb cut narrative, however well-intentioned, has been used to erase disabled people from their own innovations by justifying accessibility investments only when they can be shown to benefit non-disabled people. This emerging pattern takes that erasure a step further. The accommodation is not being built for disabled people at all, but for AI companies. Political will, funding, and urgency materialize only because the tech industry needs the same thing that disabled communities could not get on the basis of their civil rights.

This is not an isolated case. Consider road markings, a mundane feature of transportation infrastructure, that nevertheless carries significant consequences for both human drivers and autonomous vehicles. In a 2016 press demonstration in Los Angeles, a semi-autonomous Volvo repeatedly refused to move because its cameras could not detect the faded lane markings. Volvo’s North American CEO reportedly yelled at the Mayor of LA, “You need to paint the bloody roads here!”

As early as 2013, General Motors testified before the US Congress that for autonomous vehicles, “one of the key highway needs is to provide—at a minimum—clearly marked lanes and shoulders.” The resulting push for standardized, high-contrast, well-maintained road markings, culminating in 2022 Federal Highway Administration rules on minimum retroreflectivity, was driven substantially by the autonomous vehicle industry. Such improvements benefit all drivers, including those who drive at night. The need was always there, but meaningful action occurred only when the self-driving car companies needed the same thing.

Sidewalk delivery robots present a more ambivalent example of this pattern, because the robots share disabled people’s infrastructure needs while also competing with them for space. Researchers have argued that delivery robots require many of the same things as wheelchair users: a continuous, unobstructed path free of cracked pavement, missing curb cuts, and overgrown vegetation. In testing, robots were unable to operate in some neighborhoods because of infrastructure limitations that wheelchair users navigate daily.

Yet these same robots also create new hazards for wheelchair users. In 2019, a delivery robot reportedly blocked a wheelchair user from accessing a curb cut while she crossed an intersection in Pittsburgh. More recently, in 2025, a man using a mobility scooter filmed a robot repeatedly swerving into his path and colliding with him. Researchers have documented disabled people’s concerns that sidewalk robots reintroduce access barriers because robotics companies overlook accessibility in the design process, even as they tout the robots’ potential to incentivize better sidewalk maintenance. Infrastructural changes for new technologies can produce both benefits and burdens, and disabled people are left to absorb both.

This pattern, in which disabled people fight for infrastructure only for it to be leveraged and overtaken by automated systems, deserves a name. Let’s call it the “ramping automation effect”: the ramps that wheelchair users built and fought for are now used by robots competing for the same sidewalks, as technological automation increases in scale and force (“ramps up”) across physical and digital domains. The ramping automation effect follows a long history of capitalistic and militaristic uses of otherwise public infrastructures that consolidate power while actively displacing disabled people.

To resist the ramping automation effect, disability advocates and policymakers must play a leadership role in shaping AI-driven infrastructure changes so they benefit disabled people. If the web is going to be restructured for machine readability, accessibility experts should be at the forefront of deciding how that restructuring happens, ensuring that standards serve disabled users and not only AI agents. Infrastructure improvements motivated by autonomous systems must incorporate co-design from a plurality of perspectives to ensure they deliver meaningful accessibility improvements and advance the public interest.

There are genuine opportunities to align the interests of AI and accessibility, but that alignment cannot be taken for granted. Plaintext documentation optimized for a language model is not the same as well-structured HTML for a screen reader. If the benefit for accessibility is assumed rather than intentionally designed, the result may be superficial or illusory.

However, the public must not accept that incidental accessibility benefits arising from AI development are an unqualified good. Not only is this assumption flawed, but it accepts a hierarchy that disability advocates should be challenging. The implicit premise is that disabled people’s needs require a commercial co-signer before they merit action. When a society is unmoved by decades of advocacy from disabled communities but springs into action when a tech company needs the same accommodation, it reveals whose claims on shared resources are treated as legitimate and whose are treated as optional.

And as AI companies begin to frame machine-readability and robotic infrastructure as accessibility stories, as some already do, the risk of accessibility-washing grows. Starship Technologies, whose delivery robots reportedly blocked a wheelchair user from curb cuts, maintains a dedicated accessibility page touting the robots as a service for disabled customers. Similarly, Agentic AI companies may create agent-friendly content that introduces new accessibility barriers, while claiming credit for accessibility work they are not actually doing.

Political energy for real accessibility mandates may weaken if the public comes to believe AI is solving the problem as a byproduct of technological progress. Controversial AI deployments, in turn, may be legitimized through the language of inclusion. Accessibility is a civil right. Treating it as a fortunate side-effect of fashionable AI development is another way of saying that disabled people’s needs, on their own, were never considered sufficient.