Table 3. Distribution-level alignment between each LLM^a subgroup and the matched human subgroup, measured as EAD^b, E [|ord (Y_LLM) − ord (Y_Human)|]^c.

Model	Education		Sex		Doctor visits		ERd visits
	Low	High	Male	Female	Low	High	Low
Perception-based questions (Q1, Q10; n=8)
Sonnet 4.5	1.29	0.85	0.97	1.02	0.97	1.07	0.97
Opus 4.6	1.11	1.24	1.04	1.06	1.05	1.03	1.04
GPT-5.2	1.03	1.09	0.93	0.99	0.96	0.93	0.96
GPT-4.1	1.11	1.19	1	1.03	1.02	1	1.02
Human (split-half)	0.93	0.93	0.85	0.98	0.94	0.90	0.96
Information-based question (Q2-Q9; n=32)
Sonnet 4.5	0.91	0.93	0.92	0.94	0.93	0.92	0.93
Opus 4.6	0.76	0.82	0.81	0.78	0.80	0.78	0.79
GPT-5.2	0.87	0.82	0.85	0.84	0.85	0.84	0.84
GPT-4.1	0.93	0.89	0.92	0.91	0.91	0.93	0.93
Human (split-half)	0.82	0.75	0.85	0.66	0.80	0.76	0.79

^aLLM: large language model.

^bEAD: expected absolute ordinal distance.

^cThe upper block reports perception-based questions (Q1 and Q10 per discharge summary; n=8) and the lower block reports information-based questions (Q2-Q9 per discharge summary; n=32). Q2-Q5 are the confidence-axis items and Q6-Q9 are the categorical items. Lower values indicate closer alignment; the per-item maximum attainable distance varies from 2 (3-option items) to 5 (6-option items) across the 32 information items, with an item-weighted maximum of approximately 2.9; the 8 perception items have a uniform per-item maximum of 4. An exact-match alignment for information items is provided in Table S10 in Multimedia Appendix 1 and a per-item normalized version is provided in Table S11 in Multimedia Appendix 1.

^dER: emergency room.