AI-first accessibility — without losing the human
AI can scale accessibility work in ways that were impossible five years ago. It can also create a confident mess. Here's how we use it responsibly.
We’re an AI-first company. We also think most “AI accessibility” marketing is dangerous. Both of those are true, and holding them together is the whole point of this post.
AI is genuinely transformative for accessibility work. It’s also very good at producing plausible, confident, wrong output — which in accessibility means shipping a barrier with a green checkmark next to it. The difference between those outcomes is entirely in how you use it.
What AI is genuinely great at
- Triaging audits at scale. Pointing a model at thousands of pages to cluster and prioritise likely issues is something no human team can do by hand.
- Drafting remediations. Suggesting an
aria-label, a better alt text, or a semantic refactor — as a first draft for a human to approve. - Describing images and media. Generating alt-text and transcript drafts that a human then refines for context and intent.
- Explaining issues in plain English. Turning “fails 1.4.3” into “this grey text on a white button is too faint to read — here’s the fix.”
This is the work the Aware Console automates: continuous scanning, fix suggestions, and accessible document generation, with a human in the loop at every decision point.
Where AI quietly fails
The trap is treating AI output as ground truth.
An automated tool — AI or not — can reliably detect only a minority of WCAG issues. The rest require human judgement about meaning, context, and intent.
A model can tell you an image is missing alt text. It cannot reliably tell you whether that image is decorative or load-bearing for understanding the page — that depends on the surrounding content and the author’s intent. It can generate an alt-text that’s accurate and useless (“a person sitting”) when the point was “a wheelchair user independently boarding a low-floor bus.”
This is exactly why automatic overlay widgets don’t work. Bolting a script onto a broken page and asking it to “fix accessibility at runtime” is the maximal version of trusting automation it hasn’t earned. It’s also why the W3C doesn’t endorse them.
Our rule: AI accelerates, humans decide
The way we keep AI useful without letting it cause harm:
- AI drafts, never ships. Every AI suggestion is a proposal a human reviews before it reaches users.
- People with lived experience review the hard calls. Context and intent questions go to people who actually use assistive technology.
- Measure outcomes, not coverage. We track whether real assistive-tech users can complete real tasks — not how many checks turned green.
- Keep a human appeal path. Anyone can flag a barrier and reach a person, fast.
Used this way, AI doesn’t replace accessibility expertise — it multiplies it. We get through more audits, ship fixes faster, and free up human judgement for the decisions that actually need it.
That’s what “AI-first” means to us: not AI instead of people, but AI so people can go further. It’s the second leg the lab marches on — and it only works because the first leg is disability-first.
Curious what AI-accelerated, human-reviewed accessibility looks like on your product? Let’s talk.