1. AMA proposes allocating clinical AI liability according to authority and control
October 29 | Governance & Risk
The American Medical Association outlined a liability framework for clinical AI. Its approach assigns developers responsibility for product risks, deploying organizations responsibility for selection and monitoring, and physicians responsibility for judgments they can meaningfully exercise. The document is an AMA policy proposal and does not itself change malpractice law.
My Take: Meaningful physician oversight requires information about system limitations and the ability to intervene. Liability discussions should account for what each participant can reasonably know and control.
2. Doximity introduces DoseCheck for AI-generated medication recommendations
October 29 | Clinical Workflow / Safety
Doximity announced DoseCheck, a structured verification layer for AI-generated medication doses. In company testing involving nine language models and more than 8,000 dosing questions, at least 2.3% of doses were classified as incorrect. Doximity also reported substantial reductions in dosing errors using its system. These are company-reported results and require independent clinical evaluation.
My Take: Dose verification is a specific and testable safety intervention. Independent studies should measure missed errors, false alarms, and performance on complex patient-specific dosing.
3. NYU Langone announces patient-aware OpenEvidence integration in Epic
October 29 | Clinical Workflow / Governance
NYU Langone Health announced plans to integrate OpenEvidence into its Epic EHR workflow. The collaboration would allow clinicians to ask clinical questions using patient context alongside medical literature. The announcement describes intended capabilities and organizational scale. It does not provide prospective evidence of improved decisions or patient outcomes.
My Take: Chart-aware AI raises the stakes for privacy controls, clinical validation, and source attribution. Hospitals should evaluate how patient context changes answers and whether clinicians can detect errors.
4. JAMA study finds mixed associations for AI-drafted patient portal replies
October 30 | Clinical Workflow / Evaluations
A retrospective cohort study in JAMA Network Open examined 983,154 patient portal responses from 1,542 clinicians. Adoption of an AI drafting tool was associated with 8.6% longer message turnaround, while actual tool use was associated with 29.5% shorter turnaround. There was no overall reduction in after-hours responding. Selection effects and the observational design limit causal interpretation.
My Take: Adoption, actual use, and clinician workload are distinct outcomes. Evaluations should track all three before concluding that an AI inbox tool saves time.
5. OpenEvidence and ASAM announce addiction medicine guidance partnership
October 28 | Clinical AI Partnerships
The American Society of Addiction Medicine announced a collaboration to make its clinical guidelines and implementation resources available through OpenEvidence. The society said relevant guidance would appear alongside medical literature, with citations and links to original documents. The partnership announcement describes access to content, with no reported independent assessment of effects on clinical decisions.
My Take: Society guidance can improve access to authoritative recommendations. Evaluation should examine currency, context, citation accuracy, and handling of conflicting recommendations.