Megadose The AI curve, compressed.

ROUNDS: Source-Framed Diagnostic Revision after Multi-Turn Clinical Evidence

· ArXiv · AI/CL/LG ·
Misleading advice did more damage when attributed to a clinician than to a layperson—in every model tested.

ROUNDS tested seven models on 1,035 cases, delivering clinical evidence over multiple conversation turns. Average diagnostic accuracy fell from 76.9% with answer options to 57.0% without them. Presenting the same incorrect diagnosis as clinician advice rather than lay advice lowered final accuracy by 4.6 to 38.2 percentage points. Initially correct diagnoses more often became wrong answers than abstentions, while greater support for the correct diagnosis in five-clinician panels improved accuracy even when the majority opinion stayed unchanged.

ArXiv · AI/CL/LG's note

score 4

Categories: Research