AI in medicine: how deep learning algorithms can be used to augment doctors in diagnosis
In some trials, “deep learning” systems have outperformed human experts. — One evening last November, a fifty-four-year-old woman from the Bronx arrived at the emergency room at Columbia University's medical center with a grinding headache.
Context & Ripple Effects
This 2017 New Yorker piece is the early marker of a decade-long arc: deep learning systems beating human experts in diagnostic trials, with the Bronx ER case at Columbia's medical center as the opening vignette. What followed in the corpus shows the idea moving from trial to ward — US hospitals adopting predictive models to triage the riskiest ER and ICU patients by 2022, then doctors using AI for faster diagnostics and treatment targeting by late 2024.
By 2025-26 the loop had closed on both sides of the consult: patients as well as doctors now lean on AI for diagnoses despite expert warnings about answers that sound confident but lack nuance, and an OpenAI o1 study reported 67% correct ER diagnoses versus 50-55% for triage doctors — the same outperformance claim this article made nine years earlier, now measured at scale.
First-order effects
- Doctors gain a second opinion engine that matches or beats expert accuracy in narrow trials — but the named failure mode from later coverage is confident answers without clinical nuance, which lands directly on physicians who adopt it without guardrails.
Second-order effects
- Hospitals that deployed predictive triage models become the natural buyers when generalist models like o1 post better diagnostic numbers than their own triage staff, shifting procurement from bespoke internal models toward frontier-model vendors.
Third-order effects
- If model-vs-doctor benchmarks keep widening, the doctor's role structurally shifts from primary diagnostician to verifier and communicator — consistent with the parallel coverage on AI training clinicians' empathy before they meet real patients.
The trend: Medical diagnosis is migrating from human-expert-first to model-augmented workflows, with each new benchmark study eroding the assumption that the physician is the accuracy ceiling.