How is AI being used in healthcare and is it trustworthy?

Real and deployed in narrow, well-defined tasks — particularly medical imaging — and mostly hype when described as 'AI diagnoses patients'. The gap between a model that beats radiologists on a benchmark and a system that safely improves outcomes in a hospital is enormous, and it's where most projects die. What's genuinely in clinical use: - **Medical imaging analysis.** Detecting diabetic retinopathy from retinal photographs, flagging suspicious findings in mammography, identifying strokes and bleeds on CT for triage, spotting lung nodules. Hundreds of imaging algorithms have regulatory clearance, and some are meaningfully deployed. These work because the task is narrow and there's an unambiguous visual answer. - **Triage and prioritisation.** Reordering a radiology worklist so the likely emergency is read first. Lower risk because a human still reads everything. - **Administrative work** — ambient documentation that drafts clinical notes from a consultation. This is the fastest-growing real use and addresses a genuine problem (clinicians spend enormous time on notes). - **Drug discovery** — protein structure prediction and candidate screening, where the impact has been substantial. - **Early warning scores** in hospitals for deterioration or sepsis. Why trustworthiness is a serious question, not a formality: 1. **Distribution shift.** A model trained on one hospital's scanners and patient population frequently degrades badly elsewhere. Performance in the published paper is not performance in your clinic. 2. **Bias.** Models trained on unrepresentative data perform worse on underrepresented groups. This is documented repeatedly — including systems where an algorithm underestimated the needs of Black patients because it used healthcare spending as a proxy for illness. 3. **Shortcut learning.** Models have been found detecting the scanner type, or a marker on the film, rather than the pathology — right answer, wrong reason, fails silently on new data. 4. **Automation bias.** Clinicians can defer to a confident machine output. A tool that's right 95% of the time can make outcomes worse if it erodes scrutiny on the other 5%. 5. **Weak prospective evidence.** Most published results are retrospective. Randomised trials showing improved patient outcomes are comparatively rare, which is the evidence standard that actually matters. The honest summary: as a second reader, a triage tool, and an administrative assistant, it's real and increasingly valuable. As an autonomous diagnostician, it isn't, and the regulatory frameworks reflect that. If you're evaluating a claim, ask three questions: was it validated on a different population than it was trained on, was it tested prospectively, and does a human remain accountable for the decision?