Lit Matters #1: Artificial Intelligence and Human Values

Artificial intelligence in medicine is good at prediction but weak at bedside value judgments. Large language models can draft language fluently, yet treatment recommendations still shift with whose goals are being optimized — patient, clinician, or insurer.

AI Predictions and Human Values

• Probability versus causality: Large language models predict likely next words and can mirror clinical reasoning, but bedside diagnosis depends on cause-and-effect thinking plus patient goals, not probability alone.
• Stakeholder-dependent recommendations: The same clinical vignette produced different treatment advice when GPT-4 was prompted as a physician, parent, or insurer, highlighting how model output tracks perspective.
• Utility in medical AI: Utility is the value a model assigns to outcomes, and that value function may reflect the designer or user rather than the patient sitting in front of you.
• Utility elicitation limits: Eliciting preferences over time is essentially the art of medicine: weighing tests, tradeoffs, and family priorities in ways current AI still handles poorly. We get into that distinction in the episode.
• Reinforcement learning promise: Reinforcement learning may move AI beyond static probability engines by learning from sequential human decisions, but it is not yet a substitute for clinician judgment.
• Near-term clinical use: The most credible early wins are patient communication and virtual scribing, whereas autonomous treatment recommendations remain far less reliable in preference-sensitive decisions.