Megadose The AI curve, compressed.

Right Diagnosis, Unsafe Route: Auditing Clinical AI Decision Trajectories and Corrective Oversight for Validity and Potential Harm

· ArXiv · AI/CL/LG ·
In the hospital cohort, 55.2% of AI trajectories ending in an acceptable diagnosis included a decision with moderate or severe potential harm.

The audit replayed 300 cases across three sources, assessing tests, treatments and disposition decisions using only the evidence available at each checkpoint. An LLM judge scored decisions, with physician review calibrating the rates. Structured feedback resolved 73.6% of first actionable errors, compared with 24.8% after bare rejection. But new potentially unsafe decisions followed 30.2% of structured corrections, versus 19.7% of rejections.

ArXiv · AI/CL/LG's note

score 4

Categories: Research