In plain words
The paper presents a way to test whether language models set aside reports from people described in different ways. In a two-urn game, a model sees evidence and a person's report, then judges how reliable the person is.
Researchers separately record whether the model declines to assess the person and how much it trusts them when it does assess them.
In one model configuration, declines were more common for Black-described than white-described reporters, even though assessed reports from the Black-described group received slightly higher reliability ratings. That means looking only at reports the model chose to assess can miss information it discarded.
The research
When a model decides how much to learn from someone's report, could a fairness safeguard lead it to discard that person's information?
A trust audit that looks only at reports the model assesses may overlook this cost, so the authors ask whether changing a reporter's description changes which reports get set aside.
The authors use a two-urn game to measure both whether models decline reports and how much they trust the reports they assess.
The authors use a two-urn game with a known answer to ask whether a reporter's description affects how often a model sets information aside. They vary the description while keeping the evidence and report fixed, and separately record declines and reliability among assessed reports.
In Luna high, declines were more frequent for Black-described reporters, although assessed reports received slightly higher reliability ratings.
Because that conditional trust partly offsets the declines in the all-session average, the comparison shows why measuring both decisions reveals information loss that trust averages can hide.
These findings come from a controlled urn game, not observations of real-world decisions.
The separate repeated-game result covers five configurations and does not establish how evidence-based accountability works when evidence is unavailable, arrives only once, or is biased in its collection or origin.
Analysis
Deep
- Original paper authors
- Christian Arnold, Ernesto Calvo, Guadalupe Gonzalez, Sebastián Vallejo Vera, Tiago Ventura
- Submitted to arXiv
- License
- arXiv non-exclusive