Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-08-17 · high
Models cited gender or ethnicity in their own stated reasoning in at most 0.03% of cases, so self-report cannot catch the bias.
Why now: The audit's design was prespecified before a single response was scored, which is what makes the reversal hard to write off as prompt luck.
A randomized audit ran seven LLMs through 40,068 doctor-recommendation choices and found race and gender swayed picks by up to $14 a visit, per arXiv.
That money moved for reasons the models almost never admit to. Researchers found gender or ethnicity showed up in the models' own stated reasoning in at most 0.03% of responses.
Reputation still drove the biggest swings. A doctor rated 3.9 to 4.7 stars gained 31.4 points of choice probability over lower-rated peers. Being listed first alone was worth about $11 in perceived value.
The demographic effects ran opposite the direction found in prior human-audit studies of real patients. Female-signaled names gained 2.5 points of choice probability. Hispanic-, South Asian-, and Black-signaled names gained 1.3 to 2.9 points over White-signaled names.
The study built in three personas and nine paraphrases across 3,024 choice sets.
A transparency rule built on asking a model to explain itself catches none of this, because the bias lives in outcomes, not in stated reasoning. Auditing healthcare AI for fairness means testing what a model actually recommends across demographic swaps, not what it says about why.
The audit's design was prespecified before a single response was scored, which is what makes the reversal hard to write off as prompt luck.
Each link below shares sources, entities, or timing with this story.
Shared entity: LLMs / Same source domain / Shared topic / Earlier coverage / Tension
Both cover LLMs; reported by the same outlet (arxiv.org); overlapping topics (choice, model).
Both cover LLMs; reported by the same outlet (arxiv.org); overlapping topics (explanation, model).
Shared entities / Same source domain / Earlier coverage
Both cover LLMs, Models; reported by the same outlet (arxiv.org); earlier LLMs coverage from 2026-08-16.
Shared entities / Earlier coverage / Tension
Both cover LLMs, Models; earlier LLMs coverage from 2026-07-31; pushes against this story (against).
Shared entity: LLMs / Same source domain / Shared topic / Earlier coverage
Both cover LLMs; reported by the same outlet (arxiv.org); overlapping topics (gain, model).
Shared entity: Models / Same source domain / Earlier coverage / Tension
Both cover Models; reported by the same outlet (arxiv.org); earlier Models coverage from 2026-08-03.
Shared entity: LLMs / Same source domain / Earlier coverage / Tension
Both cover LLMs; reported by the same outlet (arxiv.org); earlier LLMs coverage from 2026-07-31.
Both cover LLMs; reported by the same outlet (arxiv.org); earlier LLMs coverage from 2026-07-26.