What published evidence shows about AI benefits, harms and the decisions healthcare leaders face.
Examine selected studies in medication safety, mental health, pediatric screening, oncology and rehabilitation. Compare the design, measured outcomes and limits before considering local use.
Read the clinical evidence comparison · Download evaluation worksheets
The oncology comparison includes the 2026 MASAI interval cancer analysis, following the 2023 interim findings summarized below.
This is a model neutral record of outcomes, limitations, and governance lessons. We distinguish randomized trials, observational studies, external validations, regulatory decisions, and news reports. A positive result in one workflow does not prove that every AI system is safe, and a failure does not condemn every model.
Last reviewed: September 28, 2026. Educational information only. It is not medical advice. If you have a health concern, contact a qualified clinician; for an emergency, use local emergency services.
In the randomized MASAI trial’s prespecified interim safety analysis, AI supported mammography screening detected 6.1 cancers per 1,000 screened participants versus 5.1 with standard double reading, while reducing screen reading workload by 44.3%. This was an interim result. It did not establish reduced mortality, and the workflow still included radiologists.
Primary source: The Lancet Oncology (MASAI trial)A prospective pivotal trial of an autonomous AI diagnostic system in primary care offices reported 87.2% sensitivity and 90.7% specificity for detecting more than mild diabetic retinopathy. The useful lesson is not that autonomy is universally safe; it is that a narrow intended use, controlled workflow, measurable performance, and referral pathway can make evaluation possible.
Primary source: npj Digital MedicineA 2026 NEJM AI observational study evaluated an AI triggered rapid response workflow in 23,132 hospital patients. Because the design was not randomized, the association should not be treated as proof that the algorithm alone caused better outcomes. It does show why workflow design and the human response are part of the intervention.
Primary source: NEJM AIAn external validation of a widely implemented proprietary sepsis model found substantially lower discrimination than the vendor’s reported internal validation. At one common threshold, the model failed to identify many patients with sepsis while generating alerts for many patients who did not develop sepsis. The lesson: require local validation, subgroup analysis, alert burden testing, and monitoring after deployment.
Primary source: JAMA Internal MedicineA 2019 Science study found substantial racial bias in a commercial population health algorithm because healthcare spending was used as a proxy for health need. At the same predicted risk, Black patients were sicker than White patients. The lesson: audit labels and proxies as well as model accuracy, and test who receives or loses access to care.
Primary source: ScienceA large community study found that computer aided detection in screening mammography was associated with higher recall and biopsy rates and uncertain clinical benefit. Although this predates today’s foundation models, the governance lesson remains current: downstream harms and resource use matter, not just detection metrics.
Primary source: New England Journal of MedicineThe patient lens
These are patient consequences reported in the evidence above. They are not endorsements or attacks on any model vendor. The result depends on the clinical task, data, workflow, population, and human response.
What made help more likely: a narrow intended use, relevant validation, clear follow up, and a qualified person responsible for action.
What made harm more likely: weak external validation, poor proxies, automation bias, unclear accountability, or failure to monitor downstream effects.
We welcome documented accounts of healthcare AI helping, failing, delaying, confusing, or improving care. Stories are reviewed for evidence, privacy, fairness, and learning value. Please do not submit protected health information.
Begin with a defined clinical or operational problem. Match oversight to the consequences of error and validate the system in the local population and workflow. Monitor performance, equity, privacy, cybersecurity and automation bias. Give patients meaningful notice when appropriate, and test the process for pausing or retiring the system.
1. Intended use and people affected. 2. Evidence quality and independent validation. 3. Failure modes and subgroup performance. 4. Human authority and escalation. 5. Privacy, security, consent, and data retention. 6. Workflow burden and accessibility. 7. Monitoring, incident response, change control, and retirement.
Use the AI Healthcare Governance FrameworkThis record will be updated as stronger evidence emerges. Corrections and credible case submissions are welcome through the contact page.
