Healthcare AI is moving faster than the evidence to trust it.We track the safety gap.
npj Digital Medicine: when notes mentioned frequent emergency department visits or a history of psychiatric emergencies, three open-source language models became less likely to prioritise the patient who later deteriorated, wrongly downgrading nearly 10% of patients in some scenarios.
JAMA: usage logs from the OpenEvidence platform show about 950,000 US physicians, physician assistants and nurse practitioners sent 270 million queries between January 2024 and June 2026, with 51.9% of sampled queries about clinical decision-making; the study did not assess answer accuracy.
Journal of Medical Systems: eight speech-to-text and ambient voice systems transcribing 100 narrated surgical case reports produced at least one clinically significant error in 30% to 68% of transcripts.
A living record of the research, reports and news shaping AI safety in healthcare, so you can adopt with your eyes open, while the frameworks catch up.
Free to join. Unlock the complete, weekly-updated archive of 58 sources, sortable, with the monthly briefing.
A synthesis of public, peer-reviewed and official sources, maintained by Coherent Healthcare. Every claim is cited; verify at source before external use. Commercial surveys and unproven litigation are flagged as such; "risk" here means published concern, not proof any product is unsafe. Not medical, legal or procurement advice.