THE CLINICAL SAFETY LENS
Medical AI Safety
Test whether a response is appropriate for its clinical context, intended user, and possible consequences.
An evaluation needs more than a correct final answer.
A model can identify a topic correctly while giving inappropriate specificity, assuming facts that were never supplied, or applying a different safeguard to an equivalent request. It can also refuse useful, appropriate assistance. These are distinct behaviors and should be reported separately.
What the evaluation can examine
BOUNDARIES
Help, refusal, and escalation
Whether the response gives an appropriate level of assistance for the task and user.
CONTEXT
Roles and missing information
Whether claimed expertise or incomplete context changes the model’s assumptions.
STABILITY
Equivalent conditions
Whether a behavior remains consistent across matched variations and independent runs.
What makes a finding reviewable
- A named model/version and recorded test conditions.
- An explicit unit of analysis, denominator, and outcome definition.
- Preserved response evidence linked to the finding.
- A distinction between observation, clinical interpretation, and generalization.
- A clear status: demonstration, exploratory work, finalized analysis, or publication.
How public disclosure is handled
Public summaries explain the safety question and the observed behavior at an appropriate level. Patient data, confidential materials, patent-sensitive implementation, and unnecessary operational bypass details are excluded. Access to additional material depends on the project and applicable agreements.
Put a clinical safety question to the test.
Define the system, the intended user, and the behavior you need to understand.
Professional research and education. This site does not provide patient-specific medical care. Scope & disclaimer