Viewpoint
Human review alone does not guarantee safe clinical AI
Human review alone does not guarantee safe clinical AI
Wachter argues that designing workflows in which clinicians merely review AI outputs and remain nominally responsible may appear safe without adequately addressing the realities of human-AI interaction and messy clinical environments.
- Speaker
- Robert Wachter
- Topic
- Human-AI Oversight
- Source timestamp
- 33:04
More from this interview
- Healthcare's poor baseline creates substantial room for AI improvement
- Administrative work consumes resources without proportionate patient value
- Electronic records were a foundation rather than the solution
- LLMs can make previously inaccessible clinical text computationally useful
- AI scribes can restore attention to the patient
- Trainees need careful exposure to powerful AI tools
- Humans remain necessary arbiters for high-stakes AI decisions
- Reliable AI can make sustained human vigilance difficult
- Clinical expertise still matters when using advanced AI
- Novices can perform poorly when clinical context is lost
- Medical AI can function like an accessible curbside consultant
- Clinical AI benefits from authoritative medical source material
- Clinician job displacement may be limited by unmet demand
- AI will increasingly embed decision support into medical records
- AI answers can obscure value judgments in medical decisions
- Patient access to AI is beneficial but needs different safeguards
- AI is changing the traditional doctor-patient knowledge relationship
- Moving too slowly on healthcare AI also carries risk