FIELD GUIDE · MEDICINE, DIAGNOSIS & ACCOUNTABILITY

SHOULD AI BE ALLOWED TO MAKE MEDICAL DECISIONS?

AI already helps summarize research, draft notes, flag images, and support diagnosis. The argument changes when advice becomes a decision that can alter treatment, access, or a person's body.

UPDATED 2026-09-22

THE PROBLEM

THE MACHINE RECOMMENDS SURGERY

Excellent. Who is legally, clinically, morally saying yes?

Healthcare AI ranges from administrative tools to image analysis, risk prediction, decision support, and generative assistants. Those systems are not all regulated the same way. Some tools surface information a clinician can independently evaluate; others produce outputs that may be difficult to inspect or are aimed directly at patients.

01

ASSIST

SAVE TIMESummaries and paperwork are the easy case.
02

FLAG

LOOK HEREModels can focus clinical attention.
03

RECOMMEND

DO THIS?Advice starts shaping action.
04

DECIDE

WHO ANSWERS?The patient needs an accountable decision-maker.

THE FIGHT

WHEN DOES DECISION SUPPORT BECOME THE DECISION?

Medical AI could reduce workload, catch patterns humans miss, expand expertise, and make care more consistent. It can also introduce opaque errors, bias, privacy problems, and a new question about who is responsible when a recommendation is wrong.

THE AUGMENTATION CASE

AI CAN EXTEND CLINICAL CAPACITY AND INFORMATION

Physicians already use AI for research summaries, documentation, patient messages, translation, and some diagnostic support. Medical organizations see real potential to improve accuracy and reduce administrative burden when tools are validated and integrated responsibly.

A second set of eyes can help even when the first set still decides.

THE SAFETY CASE

HIGH-STAKES OUTPUT NEEDS HUMAN JUDGMENT AND REGULATORY SCRUTINY

WHO guidance highlights risks including false statements, bias, automation bias, privacy, and cybersecurity. FDA guidance distinguishes types of clinical decision software partly around whether a clinician can independently review the basis of a recommendation.

A black box is easier to tolerate when it files paperwork than when it changes treatment.

THE WEIRD SHIT

THE DOCTOR MAY BE REVIEWING A MACHINE THAT ALREADY FRAMED THE CASE

AI IS ALREADY ROUTINE IN PHYSICIAN WORK

The AMA's 2026 survey found 81% of responding physicians reported using AI professionally, with common uses including research summaries, documentation, patient messages, translation, and assistive diagnosis.

The question is already about boundaries inside adoption, not whether adoption begins.

THE AMA EXPLICITLY FRAMES AI AS AUGMENTATION

AMA policy describes 'augmented intelligence' as enhancing rather than replacing human intelligence and emphasizes transparency, evidence, oversight, liability, privacy, and the patient-physician relationship.

Mainstream medical guidance is trying to preserve accountable human judgment while using the tools.

WHO LISTS BOTH REAL BENEFITS AND SYSTEMIC RISKS

WHO guidance for large multimodal models identifies potential uses across clinical care and research while warning about false information, bias, automation bias, cybersecurity, and broader health-system effects.

The same capability can improve access and amplify mistakes.

FDA REGULATION TURNS PARTLY ON WHETHER THE BASIS CAN BE REVIEWED

FDA's 2026 clinical decision support guidance clarifies which software may fall outside device regulation and which functions remain subject to medical-device policies, including distinctions around intended users and reviewability.

Opacity and direct patient use can change the regulatory treatment.

THE PEOPLE WITH A STAKE

WHO OWNS THE BAD OUTCOME?

PATIENTS

They receive the benefits of faster and potentially better-informed care and bear the physical consequences of wrong or biased decisions.

CLINICIANS

They gain powerful tools but can inherit verification burden and unclear liability for recommendations they did not generate.

HOSPITALS + PAYERS

They can lower costs or standardize care, while also creating incentives to automate decisions that affect access and treatment.

AI + DEVICE MAKERS

They can scale expertise across institutions but must prove performance, manage updates, and support safe use in real populations.

THE UNANSWERED QUESTION

WHAT PART OF MEDICINE CAN NEVER BE A RUBBER STAMP?

TAKE THIS TO DINNER: A model can recommend. Someone with a duty to the patient still has to own the decision.

RECEIPTS

Medical AI is already ordinary. Accountability is the interesting part.

  1. More than 80% of physicians use AI professionally: AMA surveyAmerican Medical Association
  2. Augmented intelligence in medicineAmerican Medical Association
  3. Ethics and governance of artificial intelligence for health: Guidance on large multi-modal modelsWorld Health Organization
  4. Clinical Decision Support Software, Final GuidanceU.S. Food and Drug Administration