Independence doesn’t mean doing everything yourself. It means having control over your own life. Who helps you. How they help you. When you ask for it and when you don’t. That’s the definition that actually matters, and it’s worth stating it plainly because a lot of AI coverage in disability circles misses it.
The conversation about AI and disability tends to go one of two ways. Either it’s utopian, AI will solve everything, people with disabilities will finally be free, or it’s worried, what about privacy, what about dependency, what about when it fails? Both framings miss what’s actually happening, which is more practical and more interesting than either.
AI is already changing what independence looks like for a lot of people with disabilities. Not by solving disability. By shifting where the friction is.
No spam. No inspiration porn.
Our best writing for adults with disabilities, weekly and free.
Get the newsletterWhat’s actually different now
The thing that’s changed in the last few years isn’t that AI became more capable in a general sense. It’s that AI became capable enough at specific tasks that it’s genuinely useful in daily life rather than just impressive in a controlled demo.
Language models can now do a credible job of turning messy, hard-to-organize thoughts into coherent text. That matters to someone whose executive function makes starting the hard part, who knows what they want to say and cannot get it onto the page.
Voice recognition is now accurate enough for continuous dictation in most contexts. That matters to someone who can’t use a keyboard but has always been able to speak.
Image description has gotten good enough that a blind person can point their phone at something and get a useful description rather than a confused one.
Navigation tools have layered enough audio and haptic feedback that getting around an unfamiliar city without sighted assistance is genuinely more possible than it was five years ago.
None of these work perfectly. All of them work better than they did.
Decision-making support
One area that gets less attention is how AI tools can support cognitive independence, the ability to make your own decisions, manage your own information, and stay on top of your own life.
For people with cognitive disabilities, acquired brain injuries, or conditions affecting memory and executive function, this is often where independence breaks down first. Not physical access. Cognitive load.
AI tools are genuinely useful here in ways that weren’t possible before. Having a conversational assistant that can hold the thread of a complex task, break it into steps, remind you where you were, and help you think through decisions is different from a calendar app or a to-do list. It’s closer to having a knowledgeable person available to think alongside you.
That’s not a replacement for human support. But it changes the ratio. It reduces how often you need someone else to do the cognitive management work for you, which shifts control back toward you.
Communication and expression
AAC technology has been the clearest early beneficiary of AI advances, and for good reason. Communication is the foundation of everything else, and the gap between what someone wants to say and what they can say through traditional AAC has always been significant.
AI-driven word and phrase prediction means fewer touches to get to the right words. Natural language generation means a simple input can produce a full, nuanced response. Voice synthesis has gotten good enough that synthesized voices sound like real people rather than robots.
Personalized voice is the part usually described as coming soon. It arrived three years ago. Apple’s Personal Voice was announced in May 2023 and shipped in iOS 17 that September. You read a randomized set of prompts, 150 sentences and roughly 15 minutes of recording, and the device builds a synthetic voice that sounds like you. It pairs with Live Speech, so you type and your device speaks in that voice, in phone and FaceTime calls, in assistive communication apps, and in person. It is a built-in accessibility feature, not a separate purchase.
Apple says model training and inference both run entirely on the device, so the recordings never leave it. The practical shape of that: the tuning happens overnight while the device is charging, locked and on Wi-Fi, and the voice is ready the next day. Two limits Apple states and most coverage doesn’t. It is English only, and it needs a recent iPhone or iPad, or a Mac with Apple silicon.
The feature was built for people with a recent ALS diagnosis or another condition that progressively affects speech. Which means the useful time to record is while the voice you want to keep is still there.
For people who use speech generating devices, this isn’t an abstraction. It’s the difference between feeling like yourself in conversation and feeling like you’re talking through a machine.
The friction that remains
Being honest about limits matters.
AI tools fail people with disabilities more often than they fail users without a disability, and speech recognition is where that is measured rather than assumed. The Interspeech 2025 Speech Accessibility Project Challenge reports that between 2008 and 2023, error rates for dysarthric speech recognition fell by a factor of three while error rates for speakers without dysarthria fell by a factor of five. The gap widened while both got better. On the project’s own test set of 400-plus hours of impaired speech from more than 500 people, the off-the-shelf Whisper large-v2 model came in at a 17.8 per cent word error rate, against well under 5 per cent for typical speech. The best fine-tuned system in the competition got to 8.11 per cent, which is real progress and still roughly double the rate everyone else gets by default.
That research collects Canadian as well as American English, which is worth noting, because most accessibility datasets do not.
Image description still breaks down on low-contrast images, non-standard document layouts, and scenes that need context it doesn’t have. AI writing assistants sometimes flatten language in ways that erase the particular voice of the person using them.
The training data gap is real. AI systems learn from large datasets that reflect the world as it is, which means they learn from a world built without people with disabilities in mind. They replicate the assumptions embedded in that world unless deliberate effort goes into correcting them.
And then there’s reliability. A navigation system that fails once in a while is frustrating for anyone. For a blind person in an unfamiliar area, it can be genuinely dangerous. Dependency on a tool that isn’t dependable creates its own kind of vulnerability.
None of that means the tools aren’t useful. It means they’re tools with known failure modes that you should know about before you rely on them.
Data privacy: the part that gets skipped
This deserves a plain statement: using AI tools involves giving up data. A lot of it.
Voice tools process your speech. Communication tools process your messages. Navigation tools track where you go. The aggregate of what these tools collect is a detailed picture of how you live, where you go, who you talk to, and where you need help.
The companies collecting this data have privacy policies. Those policies change. The regulatory environment in Canada is thinner than most coverage suggests, and it got thinner rather than stronger. Canada’s first attempt at comprehensive AI regulation, the Artificial Intelligence and Data Act inside Bill C-27, died on the order paper on 6 January 2025 when Parliament was prorogued, and it has not been reintroduced. What governs your data in the meantime is ordinary privacy law, chiefly PIPEDA federally and its provincial equivalents, plus Treasury Board instruments that bind federal departments and a patchwork of provincial rules such as Ontario’s Bill 194 for the public sector. If a company mishandles your data, your complaint route is the Office of the Privacy Commissioner or your provincial commissioner, not an AI regulator, because there isn’t one. The data that feels like a reasonable trade for access today may be used in ways you didn’t anticipate later.
This isn’t a reason to avoid AI tools. It’s a reason to know which tools you’re using, what they collect, and whether there are alternatives with better privacy posture when the privacy difference matters.
Locally run tools, the ones that process data on your device rather than in the cloud, are worth knowing about. Personal Voice is a concrete example: Apple states the training and the inference both happen on the device. When you are choosing a tool and the data feels sensitive, the question to put to the vendor is plain. Does this run on my device, or does it send my speech, my messages or my location to your servers? A vendor that will not answer that in writing has answered it.
Building your own setup
The useful frame isn’t “AI tools” as a category. It’s specific tools for specific friction points in your specific life.
Start with where you lose the most time, energy, or autonomy. What task costs you more than it should? Where do you end up needing help that you’d rather not need? That’s the friction worth addressing first.
Then look for the specific tool, not the most impressive tool, not the one with the best marketing, that addresses that friction with acceptable tradeoffs on cost, reliability, and privacy. If you want a starting list rather than a category, Living Unlimited keeps two: the 2026 annual update on what actually works and the AI tools changing life with disability. Both name products, prices and the ones that have quietly shut down.
Try it long enough to have a real opinion. A week isn’t enough. Most AI tools require learning time to be useful. The first week of using a new dictation tool feels slower than typing. The third week starts to feel different.
And build redundancy. If your independence depends on a specific tool working, know what you do when it doesn’t. This isn’t pessimism. It’s what actually allows you to rely on the tool.
The bigger picture
AI isn’t going to remove the barriers. Inaccessibility is partly social, partly infrastructure, partly political, and only partly technical.
What AI is doing is narrowing some of the gaps at the technology layer. Making certain tasks easier, certain barriers lower, certain forms of independence more possible for more people.
That’s worth noticing. It’s also worth not over-crediting. The tech layer has gotten better. The world hasn’t gotten more accessible at the same rate. Those two things are both true.
The right relationship with AI tools is the same as the right relationship with any tool: use what works, know its limits, don’t confuse a useful instrument with the removal of a barrier.
And if something isn’t working for you, that’s not a personal failing. A lot of these tools weren’t built with your use case in mind. The gap between what AI can theoretically do and what’s actually accessible to people with disabilities right now is still significant. That gap is worth naming, worth complaining about, and worth pushing companies to close.
Sources
Xiuwen Zheng et al., “The Interspeech 2025 Speech Accessibility Project Challenge” (UIUC with Microsoft, Amazon, Google, Apple and Meta), for the dysarthric speech recognition error rates and the widening gap between 2008 and 2023. Apple, “Apple previews Live Speech, Personal Voice, and more new accessibility features”, May 2023, and Apple Machine Learning Research, “Advancing Speech Accessibility with Personal Voice”, for the on-device training and inference. Maggie Arai, “What’s Next After AIDA?”, Schwartz Reisman Institute for Technology and Society, University of Toronto, for the death of Bill C-27 on 6 January 2025 and what governs AI in Canada in its absence.
Related reading: Digital Privacy and Your Disability Data, and AI Tools for Disability in 2026.
Living Unlimited Team
