Annals Commentary Urges Physicians to Retain Final Accountability as AI Handles More Diagnostics

Parsons and Rodman contend that AI already matches or exceeds physicians on narrow diagnostic tasks from EHR data, yet treatment selection requires weighing unquantifiable trade-offs such as social circumstances and non-trial endpoints. They assert that accountability cannot be offloaded because only the clinician who faces the patient again has incentive to preserve the therapeutic relationship when outcomes deviate from guidelines. The piece draws on patterns seen in prior diagnostic AI deployments where over-reliance produced alert fatigue and missed contextual factors. Related observational work in JAMA Internal Medicine (2024) on AI-assisted triage showed 12 percent higher guideline concordance but no reduction in adverse events once physicians overrode outputs. A 2025 Lancet Digital Health meta-analysis of 18 RCTs found AI tools improved diagnostic accuracy yet left decision ownership with physicians in every successful implementation. These studies indicate that hybrid models succeed only when liability and follow-up remain human. Regulatory bodies have not yet required documented human override thresholds in AI-cleared devices. Future observational registries tracking real-world overrides and 30-day outcomes are needed to test whether retained clinician accountability measurably alters downstream harm rates.