I’ve been going to CSUN for a few years now. It’s the largest accessibility conference in the world. This year had an especially exciting atmosphere. AI was everywhere. How do you use it responsibly? Where does it fall short? Are we even saving any time or just moving the bottleneck?
I came away with three things I keep thinking about: the ongoing AI conversation for accessibility, the complexity of procurement as a vendor, and PDFs. I went into it knowing I had to get some answers on PDFs.
AI writes code. It doesn’t know why the code matters.
There’s a study from Microsoft that kept coming up at sessions. They tested AI coding agents on standard development tasks with no accessibility instructions. Compliance across models was just: 0 – 10%.
The fix was less dramatic than you’d expect. Just mentioning accessibility in your prompt improves compliance by ~18%. A detailed instructions document gets you to 48% or higher. Some models hit 95% with the right setup (95% compliance on automated tests). We’re still in a world where simply telling an agent to do a better job delivers better results.
On the flip side there were many sessions devoted to AI-powered accessibility testing. Teams are using agents to write and run unit tests, and on paper the coverage looks good. In practice, the agents make a lot of assumptions, about browsers, about component names, about what “comprehensive” means. One team used the word comprehensive in a prompt and the agent wrote 200 tests for a single button.
There’s something interesting about these 2 tracks approaching accessibility from opposite ends. AI commoditizes the mechanical parts of development, but the things that require human judgment about human experience don’t compress the same way. Accessibility isn’t a box to check, it’s a question about whether a real person can actually use what you built. Also, the still common AI hallucination problem necessitates human checking where legal risk in concerned.
Procurement is getting more teeth
For a long time, asking for a VPAT felt like due diligence. As a custom web agency we get asked for our VPAT sometimes (even though we don’t have a product). So generally it feels pretty perfunctory and often like our clients may not fully understand where a VPAT is, or is not necessary. They just blanket request them for everything.
But procurement is changing. Colorado passed a law holding institutions accountable for the accessibility of products they procure. The fine is $3,500 per violation. That changes how seriously a vendor’s VPAT gets read. Colorado may be a one-off, or it may signal a shifting environment. Title II rulemaking is tightening obligations for institutions. Everyone is having to get caught up very quickly on exactly what they ought to be doing in procurement.
A common refrain in the accessibility community: “the document isn’t the problem, it’s how it’s being read.” I know as a vendor myself, we encounter institutions who are just looking for 100% “supports” down the line. Of course, 100% “supports” is a huge red flag. If a vendor has never failed a criterion, they probably didn’t test very hard. This space is changing pretty quickly and institutions need to stay current.
PDFs are a crisis nobody has to have
One session that really stuck with me was from University of Illinois Chicago. UIC had 200,000 PDFs. Remediation costs run $5-10 per page. So about a million dollars minimum to remediate their entire archive.
The argument being made, by that session and then by several vendors I talked to on the floor, is that the remediation framing is the wrong one entirely. PDFs were designed for print. Accessibility was never core to the format. Treating them as the destination and retrofitting them is expensive, slow, and never quite done.
The better move is getting out of PDFs where you can. Tools like Equalify from UIC convert PDFs to Markdown at $0.10 a page. There are vendors with tools to output to Word, HTML, and ePub. The approaches differ but the premise is the same: the file format is the problem, not just the content inside it.
This is more actionable for higher ed and government than it might sound. A lot of what lives in PDFs, forms, policy docs, program information, could live on a webpage. It would be more findable, more maintainable, and accessible by default.
