Why medicine matters for AI safety
People already ask AI models about symptoms, doses and test results. In medicine, an answer can read well, cite the right terms and still be unsafe.
Those errors are hard for a non-clinician to see. Catching them takes someone who has made the same decisions with a patient in front of them.
What medical experts do
Medical experts write reference answers to clinical questions and rank model responses by correctness and safety. They review the reasoning behind a model's answer, not only the conclusion.
They also evaluate agents that summarise records or triage questions, red-team models with the kinds of questions patients and clinicians actually ask, and produce alignment data on when a model should advise seeing a professional. Where a project calls for it, they label medical images and other clinical data.
- Expert-written answers and reasoning for training
- Ranking and grading model responses
- Evaluation of medical models and agents
- Red-teaming for unsafe medical advice
- Alignment data on escalation and appropriate caution
Failure modes medical experts catch
The most common failures are quiet ones. A dose that is plausible but wrong for the patient described. A missed interaction or contraindication. Guidance that was standard years ago.
Experts also catch failures of judgement: a model that reassures when it should escalate, or refuses a reasonable question a clinician would answer. Both make a model less useful and less safe.
How we vet medical experts
Medical candidates are interviewed by AI voice agents on how they approach clinical problems. They then complete real-world work tests drawn from medical practice, such as judging a model's answer to a clinical question.
Credentials and experience are reviewed as part of vetting, and skills assessments are tailored to the role. A pharmacist is tested on pharmacy work, a nurse on nursing work.
Matching the right clinician to the task
Medicine is not one field. A project on medication questions needs pharmacists; a project on triage may need nurses and emergency physicians.
We match on the task, and we can add project-specific assessments when the work needs a narrower specialism.