Accessibility is one of those rare concepts that almost nobody argues against. Making websites easier to navigate for people with disabilities is a fairly uncontested virtue. Semantic HTML, descriptive links, meaningful image descriptions, and keyboard navigation all help create a web that is more inclusive and usable.
Yet the AI era has introduced an unexpected, deeply ironic wrinkle into this story. The websites doing the most to make information accessible to humans are often the same websites making it easiest for machines to harvest. Obviously, this was never the goal. The Web Content Accessibility Guidelines (WCAG) were developed to help people, not artificial intelligence.
A screen reader needs to understand the structure of a page, and so does an AI crawler. A visually impaired user benefits from meaningful image descriptions, and so does a multimodal model attempting to associate language with visual concepts. And a search engine benefits from semantic HTML, as does a large language model collecting training data.
The common denominator is that the content is structured to make it more accessible to humans and more interpretable by machines. As a result, the web’s accessibility standards have undergone an unintended role reversal. They began as accommodations for people with disabilities. Today they also function as infrastructure for one of the largest technological industries in history.
From Alt Text to Training Data
One of the clearest examples of these unintended consequences appears in the datasets behind today’s image generators. LAION-5B, one of the largest public image-text datasets, paired images from Common Crawl, a nonprofit archive of publicly accessible web pages, with the HTML alt text written to make those images accessible. That vast collection of image-description pairs became an important source of training data for models such as Stable Diffusion and DALL·E.
The blind web, quite literally, helped teach AI how to see.
The problem is not automated collection by itself. Search engines and archival projects have scraped the web for decades, often to help people find or preserve the original material.
AI training changes the relationship. It can absorb creative and informational work without the creator’s knowledge, consent, attribution, or compensation, then use that material to produce systems that may compete with the people and organizations whose work made them possible.
Much of this content was and is publicly accessible, though some was obtained from less openly available sources. In either case, it was created to be read, viewed, and shared, not necessarily to become raw material for someone else’s commercial product.
As creators and organizations recognized that publicly accessible content could be collected, repurposed, and monetized in this way, many began looking for ways to push back. And that response created a second, equally ironic consequence.
Blocking Bots Can Mean Blocking People Too
Organizations have responded by deploying anti-bot measures such as CAPTCHAs, rate limiting, browser verification, and other automated security checks. These defenses can slow unauthorized data collection, but they land hardest on the very users accessibility work was meant to serve.
In WebAIM’s long-running screen reader survey, users have ranked CAPTCHA the single most problematic barrier on the web — ahead of ambiguous links, missing alt text, and unexpected screen changes — a ranking that has held for more than a decade.
Get Michael Buckley’s stories in your inbox
Join Medium for free to get updates from this writer.
Screen readers cannot reliably interpret the distorted images or audio clips most CAPTCHAs present, and even Google’s own documentation acknowledges that reCAPTCHA cannot guarantee full accessibility for people using screen readers or with motor impairments.
This is the most concrete injustice in the whole story.
Disabled users aren’t the ones running large-scale scraping operations, yet they’re the ones who keep failing the tests built to stop it. They have become unintended casualties of a conflict between web publishers and AI companies
To be clear, this is not an argument against accessibility. Concluding that inaccessible websites are a viable defense against AI scraping would be a profoundly backward strategy. The harms of poor accessibility are real, and excluding people with disabilities from digital life is considered an ethical failure by most standards.
The situation does, however, expose a tension between digital values we once assumed were self-evidently compatible. We believed that openness, discoverability, and accessibility formed a unified front. Today, accessibility and content control are increasingly pulling in different directions.
Technical Defenses and Legal Boundaries
At its core, the conflict between accessibility and content control is an ethical problem. However, ethical principles are difficult to codify and enforce directly, so societies typically address these tensions through two practical levers: technical controls and legal or policy frameworks.
Technical measures such as CAPTCHAs, rate limiting, browser verification, and bot detection can raise the cost of large-scale scraping, but they remain part of an ongoing arms race. As AI systems become more capable, so do the techniques used to bypass those defenses. Moreover, as this paradox illustrates, the same measures often create barriers for the very people they were never intended to stop.
Legal and policy approaches address the problem differently. Rather than trying to make content harder to access, they establish rules governing who may collect, reuse, or profit from it. Courts are already beginning to define those boundaries, and unlike technical defenses, laws and licensing frameworks shape the incentives of organizations that wish to avoid litigation, regulatory penalties, and reputational harm.
The New York Times’ suit against OpenAI and Microsoft survived a motion to dismiss in 2025 and is now in discovery, while Getty Images’ suit against Stability AI tests whether training an image generator on scraped, publicly accessible photographs constitutes infringement.
Neither case is about accessibility directly, but both illustrate the growing tension between public access to information, creators’ rights, and commercial AI use. As organizations seek greater control over how their content is collected, reused, and monetized, an important question remains largely unaddressed: How can we protect intellectual property without undermining the accessibility that makes the web usable for everyone?
The better answer is not to choose between accessibility and intellectual property, but to separate them. Just as copyright law distinguishes public access from ownership, future legal frameworks should distinguish accessibility from permission for commercial AI reuse. Making information accessible to people is not the same as granting permission for unrestricted AI reuse. Those are separate ethical and legal questions, and they should be treated that way in law.
What This Means in Practice
For all the overuse of the phrase “AI era,” it does capture one important reality: accessibility has shifted from a checklist of usability guidelines into a study of unintended consequences. The practical task now isn’t to choose between accessibility and control — it’s to ensure that protecting one no longer requires sacrificing the other.
Policymakers can distinguish, in statute, between a right of public access and a license for commercial AI reuse. Anyone deploying bot defenses should also test them against screen readers before shipping them, because right now the group most likely to be blocked isn’t the scraper — it’s the person the accessible web was built for in the first place.
The accessibility paradox reminds us that technologies designed for one purpose rarely remain that way. As we move forward, the responsibility now is no longer simply building a more accessible web, but ensuring that the values embedded in that web continue to serve the people they were intended to benefit.