AI Can Give Disabled People More Independence - If We Stop Designing It for Everyone Else

Independence is rarely one dramatic thing

When people talk about technology giving disabled people independence, they often picture a glossy advert: someone says a command, lights turn on, uplifting music arrives on schedule, and everybody looks suspiciously pleased with their kitchen.

Real independence is usually less cinematic. It is being able to read a letter without waiting for help. It is changing an appointment without enduring a phone queue that appears to have been designed as a test of spiritual endurance. It is getting the heating right when pain, fatigue or limited movement makes getting up a costly exercise.

As a disabled engineer who has spent years building temperature monitoring, embedded and IoT systems, I think AI can genuinely help. Not in the vague “AI will transform everything” sense, which is generally what people say just before asking a chatbot to write an email they could have written in four minutes. I mean practical assistance: turning difficult interfaces into usable ones, interpreting information, automating repetitive tasks, and adapting systems to an individual’s changing needs.

But there is a large condition attached. AI must stop being designed for an imaginary average user who has unlimited dexterity, energy, attention, confidence and tolerance for tiny grey text.

The useful bit of AI is often gloriously ordinary

The most promising applications are not necessarily humanoid assistants making tea. Frankly, I have met enough unreliable kettles to be cautious about delegating refreshments to anything with a cloud subscription.

Useful AI can include speech-to-text that lets someone dictate messages when typing is painful or slow. It can mean image descriptions that help a blind person understand a photograph, document or product label. It can summarise a long, inaccessible letter into clear language, organise a medication reminder around a person’s routine, or help someone draft a request for support when brain fog has reduced writing to an administrative bog snorkel.

In the home, context matters. A smart heating system should learn that “comfortable” is not a universal number. For someone with poor circulation, temperature sensitivity, chronic pain or limited mobility, being cold is not a minor inconvenience. It can affect symptoms, sleep, safety and whether the day is manageable at all.

AI can combine room temperatures, occupancy patterns and preferences to make better decisions than a crude timer. But only if it offers clear controls, explains what it is doing, and never treats its own guess as more important than the person living there. The house is not in charge. It merely contains the radiators.

Designing for disability means designing for variation

The central mistake is assuming accessibility is a feature added after the clever work is done. In practice, accessibility is part of whether the system works.

Disabled people do not have one set of requirements. My own capacity changes day to day. On a good day I might navigate a fiddly app and tolerate a few extra steps. On a bad day, I need large controls, plain language, a predictable route through a task, and no demand to remember which of seven menus contains the thing I need.

AI systems need to accommodate that variation. Give people control over voice, text, touch and switch access. Let us correct mistakes easily. Keep an activity history. Provide an obvious manual override. Make sure automation can be paused without requiring a degree in account recovery.

And do not make voice the only interface. Voice control is brilliant for some people and unsuitable for others: speech impairments, noisy homes, privacy concerns, fluctuating voices and the simple desire not to announce one’s medication reminder to the entire sitting room are all perfectly reasonable objections.

Choice is accessibility. So is a boring, dependable fallback.

Accuracy is not a nice-to-have

For disabled users, an AI error can carry a greater cost than a mildly irritating typo. If a system misreads a prescription, misunderstands a spoken command, gives an overconfident health answer, or locks somebody out of a service, the consequences can be serious.

That means developers need to stop measuring success only with average accuracy scores. Ask harder questions. Who does it fail for? Does it work with varied speech patterns, accents and assistive technology? Can somebody spot and fix an error quickly? What happens when the internet drops out, because it will, usually five minutes before something important?

Systems dealing with health, benefits, housing or care must also be honest about uncertainty. AI should support decisions, not quietly become an unaccountable gatekeeper. A polished interface saying “computer says no” is still “computer says no”, only with better typography.

Build with us, not around us

The most important design input is disabled people themselves, involved early and paid properly for their expertise. Not invited for a ten-minute usability session after the product is finished and the launch banners have been printed.

Co-design exposes problems specifications miss. A team may think a task takes thirty seconds; a user will explain that it requires two-factor authentication, a CAPTCHA, a tiny dropdown, a verification email and enough hand movement to qualify as light physiotherapy. That is not a user failure. It is a design failure wearing a lanyard.

AI has real potential to make daily life less exhausting and more self-directed. The goal is not to replace human support or make disabled people fit into inflexible systems more efficiently. It is to give us more control, more options and fewer unnecessary barriers. Build for real lives, including the messy, variable, inconvenient bits, and AI may finally become useful in the way it has been promising on conference stages for years.