Responsible AI

Prediction in mental health is only responsible if clinicians shape it.

We are building a system that tells someone they are heading for burnout. Getting that wrong in either direction causes harm, so here is exactly how it is governed.

The four layers

Nothing reaches a caregiver unreviewed.

Input classifier

Every message is classified before it is answered, which is how crisis language is routed away from the ordinary conversational path.

Intervention engine

Chooses the response type against a clinician reviewed taxonomy rather than generating freely and hoping.

Output validator

Checks the response against the rules the clinical board set, including what the AI must never say.

Safety gate

The final check before anything is shown, with hard stops for medical advice, diagnosis and self harm content.

Clinical review

Our board reviews the models, the escalation logic and the language used with a caregiver in distress.

Human escalation

Crisis paths surface professional resources and human lines rather than an AI reply, every time.

Hard limits

What the AI will not do.

These are enforced in the pipeline, not written in a policy document and hoped for.

Bias and fairness

A personal baseline is also a fairness decision.

Population norms carry the biases of the populations they were built on. Caregiver burden instruments were largely validated on specific demographics, and applying their cut-offs uniformly imports that history.

Measuring each person against their own rolling baseline sidesteps a large part of that problem, because the comparison is internal. It does not eliminate bias, since the signals themselves can be culturally patterned, which is why the clinical board reviews the signal taxonomy and why we publish the limits rather than the marketing version.

Questions

What clinicians and ethics committees ask.

We would rather have this conversation early. Write to connect@taloshealth.ai.

What model powers the AI companion?

A layered pipeline rather than a single model, combining a proprietary semantic layer with general purpose language models under strict prompt and output constraints. The layers are what make it safe, not the base model.

What is the accuracy of crisis detection?

Behavioural crisis and burnout pattern detection tests around 93 percent internally on our own evaluation set. That is a useful engineering number and not a clinical validation, and we describe it that way.

Who is accountable if the AI says something harmful?

We are. The clinical advisory board reviews escalation logic, incidents are logged and reviewed, and the safety gate is a blocking component rather than a monitoring one.

Can a caregiver turn the measurement off?

Yes. You can decline assessments and still use the coordination features. The load measurement is a service to the caregiver, not a condition of access.

Hold us to this page.
That is why it is public.

If something here does not match what you see in the product, tell us.