If you build AAC apps
The conversation is no longer the part that eats your year. Two people can now ship what used to need a company, and compete on the thing that actually helps — the interface, the symbols, the fit to one community.
Bridging disability and everyday AI.
The wave of AI that writes everyone else’s email has mostly passed AAC by. Not out of indifference — because a communication system cannot afford the one thing generative models do most easily, which is to make something up.
A guess is not a kindness.
When an AI invents a sentence in your email, you edit it. When it invents a sentence for someone who cannot easily correct it, it has spoken for them. So the engine refuses rather than improvises: a reply that does not fit the person is discarded, not cut short into something else; a failed translation is reported, not guessed; and composing your own words is always one tap away.
The same rule runs through everything we build
Peel, our science runtime, answers “unknown” when the evidence is not there. Accessibility is where that discipline came from, because here being confidently wrong is not an inconvenience — it is a person being misrepresented.
Depending on where you are standing.
If you teach or treat
A board can follow a student between apps, devices and languages, because the conversation lives in the runtime rather than inside one vendor’s product. Ask your software provider whether they call it.
If you are a family or a community
Most AAC software begins as one family’s need and stops there for want of engineering. You can start from a working conversation and spend your effort on the words and pictures that matter to your person.
Four commitments you can check in the code.
The person is not a diagnosis
Settings fix reply length, tile style and vocabulary. They never infer a diagnosis, an intelligence or a mental age. An adult can use single symbols.
The user’s words are the user’s
Only delivered selections and their own edits enter the model’s context. Suggestions and unspoken drafts never do — a draft they thought better of is not evidence about them.
Offline is a right, not a tier
No feature is held back for a connected plan, and a local model is one of the six adapters rather than an upsell.
Say what is not finished
The adult checkpoint is evaluated, not released. The child model’s answer quality is not solved. Both are written down here rather than discovered by someone three months into building on it.
This only works if it spreads.
If you know a team rebuilding turn-taking for the third time, a clinic stuck with software that cannot follow a child into a second language, or a parent writing an app at the kitchen table — send them here. A key costs them nothing and the conversation is already built.
ContinueBack to the overview