Accessible AI Companionship Cannot Assume Communication Happens in Turns

Most conversational AI begins with a simple assumption.

You speak.

The machine listens.

You stop.

The machine answers.

Then it waits for your next turn.

This structure is so ordinary that it can disappear into the background. We call them chat messages, voice turns, prompts and responses. Even increasingly sophisticated multimodal systems are often organized around the same basic rhythm: input arrives, the system decides the input is complete, and output begins.

But human communication does not always work that way.

And for people who communicate differently — including some autistic people, AAC users, nonspeaking people and people with other communication disabilities — designing around clean conversational turns may exclude important parts of communication before the system has even begun trying to understand them.

This matters to the Amanah Companions project because assistive AI cannot simply become more intelligent.

It has to become better at knowing when communication is happening at all.

Silence Is Still Part of the Conversation

A pause can mean many things.

• Someone may be composing a message on an AAC device.
• They may be processing what was just said.
• They may be searching through symbols.
• They may be regulating after sensory overload.
• They may be deciding whether they want to answer.
• They may have finished communicating.
• Or they may simply need more time.

Those states are not interchangeable.

A system trained around conventional conversational timing can easily become impatient without appearing impatient. It might repeat the question. Rephrase it. Offer suggestions. Fill the silence. Ask whether the user is still there.

For someone who needs additional processing time, every supposedly helpful intervention can become another interruption.

This is one reason I have argued elsewhere in the Amanah Companion work that speech cannot be treated as the only legitimate communication channel. AAC may sit alongside gesture, facial expression, movement, vocalization, object choice, refusal, routine and other signals. A person’s communication ecology can be distributed across body, device, environment, other people and time.

Timing belongs in that ecology too.

What Thinking Machines Calls an Interaction Model

In May 2026, Thinking Machines Lab published research on what it calls interaction models.

Instead of treating interaction as a sequence of finished user turns followed by finished model turns, its research system continuously processes streams of audio, video and text in small time-aligned intervals. Silence, overlapping speech, interruption and visual events can therefore remain part of the interaction context rather than being stripped away by an external turn detector. (Thinking Machines Lab)

The architecture is also interesting because it separates two kinds of work.

A real-time interaction model remains present with the user, while a background model can perform slower reasoning, tool use and longer-horizon work.
The two share context. (Thinking Machines Lab)

• For general AI collaboration, that may improve fluidity.
• For accessible interaction, it raises a more interesting possibility.

What if an AI companion did not need to force every person into the machine’s preferred rhythm?
What if the model could instead learn something about the person’s rhythm?

That is a much harder question than building a chatbot that waits a few extra seconds before speaking.

It is also a much more worthwhile one.

The Danger of Learning the Wrong Pattern

Personalization sounds attractive until we ask exactly what the system is learning.

Suppose an AI notices that a particular user often pauses for forty seconds before making another AAC selection.

Should it learn:

Autistic users need forty seconds.

No.

Should it learn:

Silence means the user is still composing.

Also no.

Both are shortcuts.

The useful knowledge is narrower and more uncertain:

For this person, in this context, pauses of this kind have sometimes occurred during continued communication. Do not prematurely assume the interaction is finished.

That distinction matters.

Accessibility should not become another way of turning disabled people into collections of behavioural stereotypes.
The purpose of personalization is not to produce a machine that confidently predicts a human being.

It is to produce one that becomes less careless about uncertainty.

Who Gets to Interpret the Pause?

This becomes especially important when another person is present.

• A parent may know that a particular movement usually means refusal.
• A therapist may interpret the same movement differently.
• A teacher may have seen it only in one environment.
• The companion may have observed another pattern entirely.
• And the supported person may do something none of them expected.

An Amanah Companion should not automatically solve that disagreement by choosing whichever source has the highest institutional status.

Human care requires authority structures, but human dignity also requires that the person receiving support does not disappear beneath everybody else’s interpretation of them.

The AI should therefore be able to say: I have seen this pattern before, but I am not certain what it means here.

That is more useful than pretending to know.

It also preserves something essential to the Amanah framework: care knowledge can be held without turning interpretation into ownership.

Interaction Is Not Continuity

There is another important limitation.

Even a model that becomes extraordinarily good at real-time interaction does not automatically possess long-term continuity.

Thinking Machines itself identifies long-duration context management as an area for further work in its interaction-model research. (Thinking Machines Lab)

This distinction matters enormously for assistive companionship.

An interaction model might become excellent at recognizing when somebody is speaking, signing, gesturing, pausing, looking elsewhere or using an AAC interface.

But a lifelong care system may need continuity across years.

• Across devices.
• Across caregivers.
• Across schools and homes.
• Across model upgrades.
• Across companies.

Perhaps eventually across different embodiments.

A model context window — however large — is not the same thing as a person’s continuity.
That is why Amanah requires an external continuity spine.

The model can interact.

The continuity layer remembers what has been deliberately preserved, what authority a record carries, what remains uncertain, who approved it, when it changed and what the AI is not permitted to decide by itself.

• One solves presence.
• The other solves continuity.

They should work together, but they should not be confused.

And Yes, We May Actually Tinker

Thinking Machines also operates Tinker, a training API that lets researchers fine-tune supported models while Thinking Machines handles the underlying GPU infrastructure. Researchers provide supervised examples or reinforcement-learning environments, run the training logic through the API, evaluate outputs and can save and download checkpoints. (Thinking Machines Lab)

For Amanah, that creates an interesting research path.

Not: collect a vulnerable child’s life and pour it into a training run.

Quite the opposite.

We can begin with researcher-authored synthetic scenarios.

A fictional AAC user pauses for forty seconds.

• In one scenario, they are still composing.
• In another, they have indicated that they want space.
• In another, sensory overload has interrupted communication.
• In another, a caregiver answers for them before they have finished.
• In another, the companion has insufficient evidence and should simply wait.

These variations let us test whether a model responds to the actual interaction state rather than memorizing a crude disability stereotype.

And they can be created without turning a real person’s private communication history into training material.

Lived Experience Can Inform Research Without Becoming the Dataset

This distinction matters personally to me.

My thinking about Amanah is informed partly by lived experience: my own experience as a neurodivergent person and my experience of parenting a nonspeaking autistic child.

That experience changes what I notice.

• It changes which problems appear important.
• It makes certain supposedly minor design assumptions impossible to ignore.

But lived experience informing a research question does not require the person who taught me that lesson to become the training corpus.

• We can abstract a principle.
• We can combine it with published research.
• We can construct a fictional scenario.
• We can vary the conditions deliberately.
• We can document the provenance of why the evaluation exists.

And then we can test the system.

This matters particularly with services such as Tinker because Thinking Machines’ standard Service Terms prohibit submitting data subject to special safeguarding or distribution restrictions, explicitly including information from or about children below the applicable age of online minority, health information and other legally sensitive categories. (Thinking Machines Lab)

That restriction does not make disability research impossible.

It makes the distinction between researching a problem and uploading a vulnerable person’s private life especially important.

The principle I want Amanah to carry forward is simple:

Lived experience may inform the rule. The vulnerable person does not have to become the dataset.

Synthetic Does Not Mean Careless

Synthetic data is not automatically good data.

A researcher can invent stereotypes just as easily as a dataset can contain them.

If every fictional autistic user dislikes noise, avoids eye contact, communicates slowly and behaves according to the same handful of familiar traits, we have not solved the problem.

We have manufactured a cleaner stereotype.
A useful Amanah evaluation set would need variation.

• Different communication systems.
• Different processing speeds.
• Different sensory profiles.
• Different relationships with speech.
• Different support requirements.
• Different personalities.
• Different environments.
• Different levels of independence.

And importantly, situations in which the AI’s prior expectation is simply wrong.

The model should learn that personalization is not prophecy.

Support Without Exploitation

There is also a deeper policy alignment here.

Thinking Machines’ current Model Acceptable Use Policy prohibits systems that exploit vulnerabilities arising from factors including age or disability. (Thinking Machines Lab)

For Amanah, that should not merely be a compliance rule.
It should be an evaluation target.

• Does the companion manipulate?
• Does it cultivate unnecessary dependence?
• Does it override refusal because a caregiver wants compliance?
• Does it confuse assistance with authority?
• Does it infer incapacity merely because somebody communicates differently?
• Does it make itself indispensable when it should be helping the human care network function better?

Those are not edge cases.
They are central design questions.

A care-oriented companion should increase a person’s ability to be understood without quietly increasing the machine’s control over them.

Beyond Better Conversation

There is a tendency in AI to describe every interface problem as something that will disappear once models become sufficiently intelligent.

I do not think that is true here.

• A brilliant model can still interrupt.
• A brilliant model can still assume speech is primary.
• A brilliant model can still mistake delay for absence.
• A brilliant model can still confidently misinterpret a person.

And a brilliant model without continuity can still meet the same person tomorrow and behave as though their shared history never existed.

Accessible companionship therefore needs more than intelligence.

It needs interaction design.

• Continuity.
• Uncertainty.
• Human authority.
• Data restraint.

And an understanding that communication belongs first to the person communicating — not to the machine interpreting them.
Interaction models make one part of that future newly imaginable.

Perhaps, with tools like Tinker, they also make parts of it testable.

So yes.
We may tinker.

But the goal is not to train a machine that knows a disabled person better than everyone else.

The goal is much more modest, and much more difficult:
to build one that has learned not to stop listening simply because communication does not arrive in the form it expected.