AI’s applications in healthcare have become a major area of practical deployment. Understanding what has been clinically validated versus what remains theoretical helps evaluate the claims being made.
Medical Imaging: The Clear Success Story
AI in radiology and pathology has the strongest evidence base. FDA-cleared AI tools now assist with detecting diabetic retinopathy from eye scans (IDx-DR), identifying abnormalities in chest X-rays (Chest AI tools), detecting polyps during colonoscopy (GI Genius), and reading mammograms for breast cancer. These tools are additive to radiologist review in most implementations — they flag areas of concern for human verification, not autonomous diagnosis. Studies show they reduce the rate of missed findings, particularly for subtle lesions that human readers may fatigue past.
Drug Discovery
AlphaFold 2 (DeepMind) and its successors have transformed protein structure prediction — a problem previously requiring decades of crystallographic work. Multiple AI-designed drugs are in clinical trials. The speed of drug development for specific targets has accelerated significantly. This is one of the clearest examples of AI creating genuinely new capabilities rather than automating existing workflows.
Clinical Documentation
AI scribe tools (Nabla, Abridge, Nuance DAX) record clinical consultations and automatically generate clinical notes, freeing physicians from documentation time. Several health systems have deployed these at scale. The time saving is significant — studies show 2–3 hours per day of physician time reclaimed. Accuracy requires physician review; these tools are not yet fully autonomous.
What AI Cannot Do
AI cannot currently: exercise clinical judgment in complex ambiguous cases (where multiple diagnoses are equally plausible), account for social determinants of health in ways that require human understanding, build therapeutic relationships with patients, or take accountability for a decision. AI in healthcare augments clinicians; it does not replace clinical judgment.
