By
Stuart Kerr, Technology Correspondent, LiveAIWire

For Haben Girma, the first DeafBlind graduate of Harvard Law
School, technology has always been the difference between access and
exclusion. She communicates through a braille display connected to a laptop,
receiving transcribed audio as tactile text in real time. AI-powered speech
recognition and real-time transcription have improved the accuracy of what
her display receives to the point where she can participate in conversations
that were previously inaccessible. The technology is not perfect. It is
dramatically better than what existed ten years ago.

The AI accessibility revolution is real and is changing lives in
concrete, measurable ways for the hundreds of millions of people worldwide
living with disabilities. It is also uneven, incomplete, and at risk of being
sidelined by a technology industry that treats accessibility as a compliance
requirement rather than a design priority. Whether AI fulfils its potential
as an equaliser depends on choices being made now about who designs it, who
funds it, and whose needs it is trained to serve.

Where AI Accessibility Is Already Working

Screen readers and image description tools using computer vision
now provide people with visual impairments meaningful access to the visual
content that constitutes an increasingly large share of digital
communication. Microsoft’s Seeing AI app, Be My Eyes enhanced with AI
description, and similar tools can describe photographs, read handwriting,
identify products and currency, and navigate physical environments in ways
that were not possible before the current generation of vision models. The
improvement in capability over the past five years has been significant
enough to expand the scope of independent activity for many users in material
ways.

Real-time speech transcription and captioning has transformed
access for deaf and hard-of-hearing users across professional and social
contexts. Live Caption features in Android and iOS, AI-enhanced captioning in
video conferencing platforms, and standalone transcription tools have
substantially reduced the dependence on human interpreters in settings where
they were previously essential and not always available. The accuracy gap
between AI transcription and professional human interpreting remains
significant for technical vocabulary, strong accents, and rapid speech, but
for everyday communication the gap has narrowed considerably.

Augmentative and Alternative Communication (AAC) devices for
people with limited speech have been transformed by AI language models that
can predict intended communication from partial input, reducing the time and
cognitive effort required to compose a message. Eye-tracking interfaces
combined with AI word prediction allow people with severe motor impairments
to communicate at speeds that approach normal conversation rates. Research
published in Augmentative
and Alternative Communication has documented meaningful
improvements in communication rate and user satisfaction from AI-enhanced AAC
systems.

Navigation, Independence, and Physical Access

AI-powered navigation tools are extending independent mobility for
people with visual impairments beyond what GPS alone could deliver. Systems
that combine computer vision, real-time mapping, and voice guidance can
describe the immediate environment, identify obstacles, read signage, and
guide users through complex interior spaces that are not covered by
conventional mapping. Microsoft’s Project Tokyo and similar research
initiatives are developing wearable AI systems that provide continuous
environmental description for blind users navigating unfamiliar
spaces.

What this means for you or a family member with a disability: the
practical capability of AI accessibility tools has advanced significantly,
but the availability of those tools through mainstream channels, the training
required to use them effectively, and the cost of access remain barriers that
do not affect all users equally. Technology that exists in a research lab or
in an expensive specialist device is not the same as technology that is
accessible to everyone who needs it.

The Design Gap and Its Consequences

AI systems trained without representative data from disabled users
are less effective for those users — and sometimes actively harmful. Voice
recognition systems trained predominantly on able-bodied speakers perform
poorly for users with dysarthria, stuttering, or atypical speech patterns.
Facial recognition systems that cannot reliably recognise users with certain
physical characteristics create access failures in authentication and
verification contexts. Navigation systems designed for walking users require
adaptation for wheelchair users navigating environments where kerb cuts, ramp
availability, and surface quality determine the accessible route.

The Web
Accessibility Initiative has documented extensively how digital
systems designed without disability inclusion in mind create barriers that
are not experienced by the majority of users and therefore may not surface as
priorities in development cycles focused on mainstream user feedback. AI
systems trained on majority-user data inherit those blindspots, performing
less well precisely in the contexts where they are needed
most.

Mental Health and Cognitive Accessibility

AI tools are increasingly relevant to cognitive and mental health
accessibility: applications that help people with ADHD manage task
organisation, tools that support people with dyslexia in reading and writing,
and AI companions that provide emotional support for people with social
anxiety or autism spectrum conditions. The evidence base for these applications
is growing, though the research rigour varies considerably across the range
of products marketed in this space.

The risk of over-reliance and the potential for harm from poorly
designed mental health AI tools is significant. An app that provides
emotional support without adequate crisis protocols, or that reinforces
avoidance behaviours in users with anxiety, can cause harm that its developer
did not intend and that users in a vulnerable state may not recognise.
Regulatory frameworks for digital mental health tools are developing, but
they lag the pace of market development considerably.

The broader connection to AI-assisted
mental health therapy is direct: the same tools that offer genuine
benefit when designed and deployed responsibly carry genuine risks when they
are not. For users with disabilities who may have fewer alternative options,
those risks are not abstract.

Policy, Procurement, and the Accessibility
Mandate

Public sector AI procurement is a lever that governments have not
yet pulled consistently in the direction of accessibility. When government
services, healthcare systems, and educational institutions procure AI tools
that perform poorly for disabled users, they are effectively
institutionalising inequality through technology choices. Accessibility
requirements in public procurement — equivalent to the physical
accessibility standards mandated for public buildings — could drive
meaningful market change by making accessibility a procurement criterion
rather than an optional extra.

The digital
divide that already disadvantages older users and others less connected to
the technology mainstream overlaps significantly with disability:
people with disabilities are disproportionately likely to face barriers to
technology access while also being among the potential beneficiaries of AI
accessibility tools. Closing that gap requires policy ambition that matches
the technical ambition the field has already demonstrated.

The design-for-majority problem in AI accessibility connects to
the consistent
pattern of AI systems serving well-resourced and majority
populations while underperforming for those whose needs differ most
from the training distribution.

Procurement reform is one
of the most direct levers available to governments. When public agencies
specify accessibility standards as a condition of contract, they create
market incentives that voluntary commitments alone do not generate. The EU
Web Accessibility Directive and the emerging AI Act provide frameworks for
extending these requirements to AI systems — a policy step that
accessibility advocates are pushing for and regulators are beginning to
consider.

About the Author

Stuart Kerr
is a technology correspondent at LiveAIWire, covering artificial
intelligence, emerging technologies, and their impact on society and
industry.