Getting Everyone into the Room

What if we invited the people who need technology most to design it with us, right from the start?

Not as "end users," and not as people we'll "check in with along the way," but as true partners.

It's a question I carry with me into almost every new tech project I encounter. Over the past few years, as artificial intelligence has become part of nearly every field, that feeling has only grown sharper: if we don't build these systems together with the people they're meant to serve, we put the whole purpose at risk. We end up with systems that don't fit people's needs, ignore their preferences, or even cause them real harm.

Instead of creating solutions, we'll deepen gaps. Instead of opening up opportunities, we'll add weight to those who already face barriers.

And this is especially true for people with disabilities.

I recently read a fascinating research review published by a team from York University. The researchers set out to map the ways AI can support people with disabilities, but also the risks, biases and gaps that emerge when those people aren't truly part of the development process. The review examined dozens of studies and tried to identify recurring patterns, promising applications, and deep failures that keep repeating themselves.

What did the review find?

It was carried out as a systematic scoping review, a method designed to map an existing field of knowledge rather than to measure effectiveness or test hypotheses. The researchers searched for academic papers published over the previous five years across eight leading databases in medicine, technology and the social sciences, such as PubMed, IEEE Xplore and the ACM Digital Library.

Of the 354 papers they found, 45 were relevant and described AI-based applications involving people with disabilities. The researchers analyzed them using thematic analysis, meaning they identified recurring patterns, central themes and gaps in the existing research.

The goal wasn't to determine which solution works best, but to identify trends in the research, along with failures and insights about how the field of AI for people with disabilities is developing and, no less important, what isn't happening in it right now.

So what did they find?

First, the review showed that all of the studies demonstrate that AI can indeed be a game-changing tool:

• For improving daily functioning (for example, AI-based augmentative and alternative communication systems).

• For managing health and rehabilitation (tools for early diagnosis, symptom prediction, or personalized treatment plans).

• For supporting social and employment inclusion.

But alongside this potential, the review exposes a deep problem: most research in the field relies on a narrow medical model of disability, one that sees disability as a problem to be solved rather than as part of the identity of a person living in an environment that needs to adapt.

What does this mean in practice? Many developments are driven by a desire to "fix" the person, without understanding the social context in which they live. In other words, without asking them.

And bias?

It's an inseparable part of the picture.

Many of the models examined in the review never underwent any bias testing at all: no one checked how different disabilities affect the algorithm's performance, no gender or cultural analysis was done, and no one asked whether the system treats everyone fairly.

The review also found that the datasets these models are trained on almost always lack sufficient representation of people with disabilities, or include them only in negative or purely medical terms. You can already guess the result, right?

Models that learn about reality from partial, biased and sometimes discriminatory records simply pass that reality on.

And the most troubling part?

Most of the studies reviewed didn't include people with disabilities in the development process at all. Not because the researchers didn't have their best interests at heart, but because they saw them as a target audience, not as partners.

Which brings us to the main and most important part, because we're here to fix things, not just to point out problems :)

So what can we do differently?

The review offers a series of recommendations that felt to me almost like a call to action for everyone working in this field. Here are a few:

• Shift from a medical model to a social model of disability, one that sees the person within their environmental and social context, and places the barriers in the surroundings rather than in the person.

• Design AI systems with an inclusive, participatory approach, in genuine partnership with people with disabilities, not only during testing but from the thinking and design stage onward.

• Test every new model for bias and exclusion, not as a recommendation but as a basic ethical requirement.

• Develop policy, regulation and training that advance these values, rather than leaving them up to the personal choice of each developer.

And why is this so urgent, you might ask? Why does it matter so much right now?

Because we're at a crossroads. The technologies being developed today are the ones that will shape the accessibility, inclusion and equality of tomorrow. If we don't build diverse voices, broad representation and deep sensitivity into these systems, they simply won't serve everyone.

If we want the AI that's here to stay to work for all of us, it needs to learn from all of us. That includes the voices that are heard less, the experiences that are documented less, and the people who are represented less.

And it starts with a bold decision by everyone responsible for developing these systems, and above all for developing the foundational algorithms that power countless applications: to get everyone into the room.