How AI Is Changing Healthcare Diagnosis (And Where It's Failing)

01 Where AI Excels

AI diagnostic tools perform best in specialties with standardized visual inputs. Radiology (96% accuracy on lung nodule detection vs 91% for radiologists), pathology (94% on cancer tissue classification), and dermatology (91% on skin lesion identification). In these specialties, the AI doesn't replace the doctor — it acts as a second reader, catching what tired eyes miss. The improvement in detection rates is real and measurable.

02 Where AI Struggles

Mental health diagnosis: 72% accuracy, barely better than chance. The problem is that mental health has no standard visual input — it relies on patient self-reporting, behavioral observation, and clinical judgment. AI can detect patterns in speech and facial expressions, but the signal-to-noise ratio is terrible. Cardiology: 85%, good but not good enough for autonomous diagnosis. The pattern: AI works when the diagnostic input is a image. It struggles when the input is human behavior.

03 The Liability Question

When an AI diagnostic tool misses a tumor, who is liable? The doctor who relied on it? The hospital that deployed it? The company that built it? Current law is ambiguous. The FDA approves AI tools as 'devices,' which suggests manufacturer liability. But doctors are told to use AI as 'decision support,' which suggests the doctor retains liability. This legal gray area is slowing adoption. Hospitals won't deploy AI they might be sued over, and AI companies won't sell tools that might bankrupt them.