A study published in NEJM AI found that large language models (LLMs) recommended substantially more expensive care than treating physicians across 200 deidentified outpatient and emergency department cases. Researchers evaluated 24 AI systems, comparing their proposed diagnostic tests, specialist referrals, and new medications with physicians’ documented plans.
Across 20 general-purpose AI systems, the average estimated cost of recommended care was $314 per visit, compared with $97 for physicians — a difference of 216,or3.2timesthephysician-recommendedcost.AIsystemsalsorecommendedmorediagnostictesting(105 additional per visit) and specialist referrals ($111 additional), as well as nearly twice as many new medications (1.14 versus 0.59 per visit).
A notable finding was that physicians ordered no diagnostic testing in 66% of visits, yet every AI system recommended testing in those cases, often suggesting standardized test panels. Even when prompted to prioritize cost or recommend only parsimonious but safe care, AI recommendations remained more expensive — 2.5 times and 1.4 times physician costs, respectively.