TRUST

AI Transparency

When AI is used, disclose its intended purpose, input/output role, human oversight, material limitations and evidence needed for accountable use.

Owner: Product & Clinical Governance
Scope: AI-assisted capabilities where enabled
Status: Published control statement

01 / CONTROL

Control model

Purpose

State whether AI assists classification, prioritization, summarization or another bounded task.

Inputs & outputs

Describe material input classes and how outputs enter the workflow.

Human oversight

Identify which decisions require human review and how overrides are recorded.

Change evidence

Model/version, evaluation scope and production-release approval are traceable where applicable.

02 / BOUNDARY

Responsibility boundary

No generic claim of diagnostic accuracy or regulatory status is made without specific supporting evidence for the exact AI capability.

03 / ESCALATION

Evidence and escalation

Questions about an AI-assisted capability should identify the exact feature/version and workflow decision being reviewed, not a generic AI label.

MODEL DISCLOSURE

Users need enough context to challenge an AI-assisted output

Transparency is operational: the interface and evidence should identify the AI-assisted function, model/version, input context, limitations and the human decision path appropriate to the deployed use case.

Identify assistance

Mark when content or prioritization is AI-assisted so it cannot be mistaken for a human-authored or authoritative clinical conclusion.

Input & limitation

Expose the relevant source data/time window and known use-case limitations; missing context remains visible.

Version & change

Correlate material outputs with model/version and deployment time so behavior can be investigated after a model update.

Challenge & override

Authorized users can review source context, reject/override output and escalate; the model result is never the only audit record.

TRANSPARENCY EVIDENCE

Make the decision lineage reconstructable

A reviewer should be able to reconstruct what the model saw, what it produced and what the human did next.

Capture inputsReference source data and quality state used for the request.

Record model identityAttach model/version and invocation timestamp/environment.

Return outputKeep result/confidence/limitations separate from source facts.

Human actionCapture accept/reject/override/escalate where workflow requires.

MonitorAggregate operational quality and override/failure signals without exposing sensitive data unnecessarily.

InvestigateUse correlation to inspect incidents or compare behavior across approved model versions.