AI for Accessibility Isn't the Enemy—It's Just Complicated

I was building an image upload feature last month when my designer flagged something. A user with low vision had submitted feedback saying the auto-generated alt text was useless—it just said "photo" for everything. I immediately thought, "Great, another reason to blame AI." But then I realized: we'd never actually invested time in training a proper model for our specific use case. We'd just plugged in a generic off-the-shelf solution and called it a day.

That experience stuck with me. We were so eager to automate accessibility that we'd actually made it worse. But reading through the arguments about AI and accessibility lately, I keep hitting the same wall—everyone's pointing out what AI gets wrong, and they're right. What they're missing, though, is that there's a real future here if we build it intentionally. It's not about replacing human judgment. It's about creating tools that actually help people make better decisions faster.

The Alt Text Problem Is More Nuanced Than We're Treating It

Let's be honest: current AI models generate garbage alt text. They lack context. They can't tell a decorative image from a critical data visualization. They see pixels, not purpose. That's the criticism, and it's valid.

But here's what I actually think has potential: AI as a starting point, not an endpoint. Imagine a model that flags an image and says, "This looks like a chart—are you going to describe it?" Now the author has a prompt, not a hallucination. They're engaged in the process. That's categorically different from what we're doing now.

The really interesting opportunity is training specialized models on your own content. A model trained on how your website actually uses images could learn your patterns. It could distinguish decorative flourishes from substantive graphics with reasonable accuracy. It could then prioritize author effort where it matters most. We don't need a silver bullet. We need better triage.

Algorithms That Don't Oppress

This is where I think the accessibility conversation gets interesting beyond just "bad AI bad." Safiya Noble's work on algorithmic oppression is essential reading, and her critique applies directly to how platforms treat disabled people.

But I keep thinking about Mentra—an employment platform built by neurodivergent people, using algorithms to match job seekers with companies in ways that actually consider their needs. That's not an accident. That's what happens when disabled people are in the room during design, not just in the data.

I want to see this replicated everywhere. What if social media recommendation engines actively worked to diversify your feeds around disability perspectives? What if they actively filtered out content that perpetuates harmful stereotypes? That requires deliberate choice, diverse teams, and training data that reflects actual disabled experience—not inspiration porn.

What I'd Actually Build

If I were starting a project around AI and accessibility tomorrow, I'd build with three principles:

First: Always keep humans in the loop. AI generates a suggestion, a human reviews and refines it. The human's decision becomes training data for better suggestions next time.

Second: Pay disabled people for their expertise. If your model needs data about atypical speech patterns, or how people with cognitive disabilities process text, or what images mean in context—that's labor. Compensate it accordingly. The Speech Accessibility Project gets this right.

Third: Make diversity non-negotiable in your team and training data. Ableist language patterns, patronizing framings, accessibility oversights—these show up in training data because they're all over the internet. You have to actively build against that.

The Real Risk Isn't AI Itself

Here's my honest take: I'm not afraid of AI in accessibility. I'm afraid of building it carelessly, then claiming we tried. We already do that with every accessibility feature. We slap on a keyboard shortcut we never tested, pat ourselves on the back, and move on.

The opportunity Joe Dolson's piece (which prompted this one) rightly highlights is real—these tools can harm people. But so can any technology built without intention. The question isn't whether to use AI for accessibility. It's whether we're willing to invest the actual work to do it right: involving disabled people from day one, testing with real users, iterating based on feedback, compensating people for their contributions.

I'm betting we can. But only if we stop treating accessibility features like checkbox items and start treating them like the genuine product challenges they are.

What accessibility problem are you trying to solve right now? If you added one disabled person to the team actually building the solution, what would change?

Source: This post was inspired by "Opportunities for AI in Accessibility" by A List Apart. Read the original article