Built With, not For

AI is tearing down barriers for disabled people, and quietly building new ones. Which one it does is a choice, and ours is included.

It would be dishonest to start anywhere but here: for many disabled people, AI has changed daily life for the better, and not in a small way. It can describe a room to someone who cannot see it. It can simplify a dense page for someone for whom dense pages are a wall. It can take a task that arrives as one impossible block and break it into steps. These are real gains, and the people who have them did not have them a few years ago.

But the same systems carry the assumptions of the world that trained them. And so they can remove one barrier and raise another in the same motion. The trouble is that the second part is rarely loud. It is not an error message. It is a summary that shifts the meaning by a degree, or a tool that answers for you instead of to you, in a way you might never catch.

The clearest example I have read is small enough to miss. An AI was asked to carry a simple instruction, that designers should talk to disabled people directly. The system rephrased it, gently, as talk to their caregivers. No warning. No sense that anything had happened. Just a quiet move of the disabled person from the author of their own life to the subject of someone else's care. That is the old medical model, reproduced by a machine that did not know it had taken a side.

—  ·  —

I take this personally, because Conny is exactly the kind of tool the warning is about. It is assistive AI, for neurodivergent people, and it makes decisions on someone's behalf. A tool like that can carry a person toward independence, or it can quietly carry it away. There is no neutral version. The design is the difference, and the responsibility sits with us, not the person using it.

A tool can remove a barrier and rebuild it in the same motion. The harm is rarely loud.

So we hold a few commitments, and we try to treat them as the actual work rather than something to say. The first is older than us, and it belongs to the disability rights movement, not to a company: nothing about us without us. Conny is built with neurodivergent people, in their daily lives, not designed for them from the outside and handed down.

The second is that Conny does not invent. It works from what the person has told it, not from what a model guessed they might need. That is partly a guard against the quiet failure above, the confident wrong answer that someone relying on the tool has the least ability to catch. If Conny does not know, it does not fill the gap with a plausible fiction.

The third is the one I think about most. Conny decides the next step, but it never decides over the person. Refusal is always free, instant, and without cost. To decide for someone, on terms they set and can take back, is support. To decide over them is control, and the line between the two is not who decides, it is what happens when the person says no. For a tool that carries the deciding, that line is not a detail. It is the entire ethic, and we watch it.

We are not exempt from any of this. The biases described in the research live in the same kinds of models Conny is built on. Staying on the right side is not a slogan we get to keep by saying it. It is a discipline: testing ourselves for disability bias, not only for the biases everyone remembers to test; keeping the data we hold to a minimum; and keeping the person the expert of their own life, not a case to be managed. The burden is on the builder, not the user.

Whether AI tears down barriers or rebuilds them is not a question that has been answered yet. It is being answered now, in a thousand small design decisions, ours among them. I would like Conny to be on the side that hands the day back to the person and never speaks over them. That is the standard we are trying to hold.

We're building ConnyAI with the people it's for. Early and unfinished, on purpose. Join our waitlist at www.conny.ai.