Question 3 of 20 · Change and evidence
Intelligence versus responsibility
If AI knows the likely answer, what is left for the clinician?
2 min read · Well measured today
Knowing the statistically likely diagnosis is not the whole of delivering accountable care. Someone has to act, answer for it, and deal with what happens next.
What we can measure today
- Emerging
In one benchmark of AI agents in a virtual health record, the best model completed 69.7% of 300 clinically derived tasks.
What this does not prove: A simulated environment. A strong answer on a case does not yet mean carrying out the whole job reliably.
Stanford HAI, 2026
The 2026 AI Index Report, Chapter 6: Medicine
Scope: Global research, with U.S. data on FDA authorizations.
How it was done: An annual review that summarizes studies by other researchers.
Limits: A summary of studies with different designs. Each headline figure comes from one setup and does not generalize on its own.
Checked against the original on November 6, 2026.
Open the source (opens in a new tab) - Established now
In the AMA survey, 88% said validation of safety and efficacy is important for adopting AI.
What this does not prove: An opinion survey. It does not settle who is legally responsible.
American Medical Association, March 2026
2026 Physician Survey on Augmented Intelligence
Scope: United States. 1,692 physicians, fielded January 15 to February 2, 2026.
How it was done: Voluntary, self-reported survey. Full and qualified partial responses are included, and each percentage reflects those who answered that question.
Limits: Reports what physicians say, not what happens to patients. The question list changed between waves, so the AMA itself cautions against strict comparison across years.
Checked against the original on November 6, 2026.
Open the source (opens in a new tab) - Emerging
A benchmark found leading general-purpose AI models produced 11.8 to 14.6 severely harmful recommendations per 100 clinical cases, and 76.6% of those were errors of omission, such as failing to recommend a critical test.
What this does not prove: General chat models on open-ended tasks. Narrow tools with clinician oversight, such as scribes and sepsis alerts, are a different case.
Stanford HAI, 2026
The 2026 AI Index Report, Chapter 6: Medicine
Scope: Global research, with U.S. data on FDA authorizations.
How it was done: An annual review that summarizes studies by other researchers.
Limits: A summary of studies with different designs. Each headline figure comes from one setup and does not generalize on its own.
Checked against the original on November 6, 2026.
Open the source (opens in a new tab) - Unknown
We found no settled rule yet on who answers when an autonomous system harms a patient.
What may change
- AI may take on more of the reasoning, explaining, triage and execution of decisions.
- Who carries legal responsibility can move between clinician, institution and vendor.
What may endure
- Errors have consequences, and patients have rights. A system must be accountable to someone, though who it is can change.
The hard questions
- Who is responsible when an autonomous system harms a patient?
- How can I challenge a system that is smarter than me in a narrow domain?
Evidence worth watching
Malpractice rulings · Policy on AI liability · How institutions assign responsibility
Sources last checked November 6, 2026
Next review by February 6, 2027.
What changed, and how we decideCareer guidance, not medical, legal or financial advice. For a major loan or licensing decision, talk with a qualified professional, and bring your questions from here.
Do you disagree?
If you have counterevidence, or we got something wrong, we want to see it. Don't take our word for any of this.
Send us the evidence