What's this about?
Patients increasingly ask LLM assistants which doctor to see. This paper audits how reputation and demographic signals shape those recommendations, framing these systems as "AI infomediaries" — algorithms that intermediate one person's choice among other people, and thereby affect those people's livelihoods.
What exactly is an algorithm audit?
It's the empirical practice of treating a deployed system as a black box and testing it the way you'd test for discrimination in hiring: send in systematically varied inputs and measure whether outputs differ on characteristics that shouldn't matter. Change only a doctor's name to signal a different ethnicity or gender, hold credentials constant, and see whether the recommendation changes. It's how bias gets demonstrated rather than asserted, and it requires no access to the model's internals.
Why is this worth noticing?
- The "infomediary" framing is the contribution: unlike most AI bias work about how a system treats *you*, this is about how it treats third parties who never interacted with it and can't appeal.
- Physician choice is high-stakes on both sides — patient outcomes, and doctors' practices being materially affected by a recommendation engine nobody regulates.
- It's an early instance of a category that will grow as assistants mediate more choices among people: which contractor, which lawyer, which school.