Marcus — Have LLMs Improved Patient Outcomes?

Source: Gary Marcus Raw: raw/rss/2026-05-03-marcus-on-ai-have-llms-improved-patient-outcomes.md URL: https://garymarcus.substack.com/p/have-llms-improved-patient-outcomes Date: 2026-05-03 Tier: 2 — deployment evidence, healthcare AI

TL;DR

Marcus quotes cardiologist Eric Topol's review (and a recent Nature Medicine editorial) finding "very little evidence for LLMs benefiting patients or doctors for health outcomes" — outside administrative-task help. Topol is normally bullish; the convergence between Topol and Marcus is the news.

Why this matters

Pairs with Google DeepMind AI co-clinician beats GPT-5.4 in blind doctor tests (05-02) and the FDA AI-trial pilot context. The capability-vs-outcome gap is now visible in the medical literature: a model that beats GPT-5.4 on blind tests has not produced patient-outcome improvements at scale. This is a measurement-vs-deployment gap, not a capability gap — and it lines up with the Safety Drift (05-02) finding that base-model evaluations don't predict deployed-system safety.

Connections

• Safety Drift After Fine-Tuning (05-02) — base-eval doesn't predict deployment, here for safety, there for outcomes. Same structural gap.
• Google DeepMind co-clinician (05-02) — capability evidence; this post is the outcome counter-evidence.
• Marcus on capital misallocation (05-01) — broader thesis. This post is medical-domain evidence.