- Home
- Responsible AI
What the models may touch, and what they may never do.
Medical decision-support software has to be explicit about its boundaries, its provenance and its oversight. This page is the public version of that statement. Where something is a commitment rather than completed work, it says so.
Four constraints, in the architecture rather than the settings.
A constraint that can be switched off in a configuration file is a preference. These are not configurable.
No autonomous write
There is no state in which a model writes to a patient record, sends a message to a patient, or closes a follow-up. A clinician puts it there or it does not go.
No uncited synthesis
Every synthesised statement resolves to the document, page and value it came from. A synthesis that cannot cite itself is not displayed at all.
No unlabelled output
Every AI-generated block carries a visible badge. No output is ever presented in a way that could be mistaken for a clinician's own note.
No unlogged decision
Each acceptance, modification and rejection is recorded against the clinician who made it, with a timestamp, for clinical governance.
What those rules look like on screen.
Trust that lives only in a policy document is not trust. These are components a clinician sees on every screen that shows synthesis.
- The AI-assistance badge
- Every summary the engine produces carries a marker reading AI-assisted clinical summary — review required. It is not dismissible, and it does not disappear once the summary has been read.
- Source citation on every statement
- Selecting any synthesised line highlights the exact source — the document, the page, the lab value — in a split view beside it. Verification is a glance, not a search, because a citation that takes thirty seconds to check is a citation nobody checks.
- The explicit disclaimer
- Carried on the reasoning workspace itself: CBS Clinical Buddy provides assistance for clinical review. All suggestions must be evaluated and validated by a licensed medical professional.
- Accept, modify, dismiss
- Three explicit actions, and no default. The interface will not proceed on a suggestion that has been read but not acted on, because silence is not consent in a clinical record.
- A modification is stored as a modification, not as a correction that overwrites what was proposed. What the system suggested and what the clinician decided are both retrievable afterwards, which is what a governance review needs and rarely has.
- Differentials shown with their counter-evidence
- Where a differential list is produced it appears with the findings that argue against each item as well as for it, and it is labelled as a list to consider rather than a conclusion reached.
The gaps, named.
A responsible-AI page that lists only achievements is marketing. These are the open items as they stand today.
No independent clinical validation study
CBS has not been through a published prospective study measuring whether it improves clinical outcomes or reduces error. Internal testing is not the same thing, and we will not present it as though it were.
No medical-device registration
CBS Health is not registered with the CDSCO, the MHRA or the FDA. It is positioned as clinical decision support. If a deployment's intended use would place it in a regulated class, that has to be resolved before it goes live, not after.
Model behaviour depends on the deployment
Which model a deployment runs against, and therefore where prompt data travels and how long it is retained, is a per-deployment configuration. There is no single honest answer across all of them, so we give the answer for yours during procurement rather than a reassuring generality here.
Bias evaluation is incomplete
Reference ranges, risk scores and the literature the models were trained on are not evenly derived across populations. We have not completed a systematic evaluation of how CBS behaves differently across age, sex, ethnicity and comorbidity profiles. Until we have, the clinician review step is doing more work than it should have to.
Who is accountable when it is wrong.
The clinician, for the decision
CBS does not shift clinical responsibility. The registered practitioner who signs remains accountable for the decision, as they would be reading the same reports on paper.
CBS, for the provenance
If a citation points at the wrong source, if a flag fires on the wrong patient, or if a summary omits a document it should have read, that is ours and it is a defect.
The deployment, for the audit
Every view, suggestion and sign-off is logged. A hospital's clinical governance committee can reconstruct what was on screen at the moment a decision was made.
Ask the hard questions on the call.
The ones on this page are the ones we have already been asked. Bring better ones.