AI tools for independence

I am cautiously, stubbornly optimistic about what AI can do for disabled people. I say cautiously because I have watched enough technology arrive with grand promises and leave people behind, and stubbornly because I genuinely believe this generation of tools could remove barriers that have stood for decades. Both things are true at once, and holding them together is, I think, the entire job right now.

Why this moment is different

The promise of AI for accessibility is not really about intelligence. It is about translation. So much of the digital world excludes people because it demands one specific way of interacting: read this dense screen, click this small icon, fill this rigid form, do it quickly, do it precisely.

AI lets us put a different surface in front of that complexity. A person can say what they want — summarize this page, fill in this form for me, read my messages and help me reply — and the system can carry it out. The interface stops being a fixed thing the person must conform to, and starts being something that adapts to them.

For someone who relies on a screen reader, or who finds dense interfaces overwhelming, or who cannot perform precise gestures, that shift is not a convenience. It is the difference between doing a thing themselves and having to ask another person to do it for them. Which is to say: it is the difference between dependence and independence.

The same power can exclude

Here is the part I cannot let go of. The exact capabilities that make AI promising for accessibility also make it dangerous if we build it carelessly.

• An assistant that acts fast is wonderful — until it confidently does the wrong thing, and the user cannot tell what just happened or how to undo it.
• An assistant that handles complexity for you is liberating — until it hides so much that the user loses the ability to understand or control their own tasks.
• An assistant that is smooth and conversational is welcoming — until it is built for sighted, fast, mouse-using people and the accessibility is bolted on at the end, badly.

If we are not deliberate, AI becomes one more thing that works beautifully for the people interfaces already served, and remains subtly broken for everyone else. The technology does not decide this. We do, in a thousand small design choices.

Three principles I build around

When I build AI tools for accessibility, three principles do most of the work.

Control over autonomy. The person stays in charge. The assistant proposes and explains; the person decides. For high-stakes or irreversible actions, it asks first. This is slower than letting the model just do things, and that trade-off is the right one. An assistant that respects the user's authority over their own life is worth more than one that is impressively autonomous.

Clarity over magic. Every action is narrated in plain language. The user always knows what the assistant is about to do, what it just did, and why. "Magic" that you cannot see into is just opacity with good marketing — and for someone who cannot afford a costly mistake, opacity is a barrier.

Reversibility over speed. Wherever possible, actions can be undone, and the system is honest about which ones cannot. When a wrong action is cheap to reverse, the user can act freely and confidently. When it is not, that is exactly where confirmation belongs.

You will notice none of these are really about the model. They are about posture — how the tool relates to the person using it. The model is just the engine. The posture is the design.

Where I land

The honest summary is this: AI could be the most significant accessibility technology in a generation, and it could also become another polished layer of exclusion. Nothing about the technology guarantees which one we get.

What decides it is whether we build for control, clarity, and reversibility — and whether the people these tools are meant to serve are at the center of the work rather than at the end of it. That is not a technical problem I can solve once and move on from. It is a stance I have to keep choosing, in every project, every feature, every default.

I choose independence. Then I try to build like I mean it.