Back to feed
arXiv cs.CL·

Evaluating Second-Order Bias of LLMs Through Epistemic Entitlement

Signal
72
Hype
18
In three linesStudy of second-order bias in LLMs: how models judge biased content, beyond generation. Grounded in entitlement epistemology, the method evaluates whether LLMs infer demographics without sufficient support. Findings: systematic bias across target groups, evasion of safety guardrails, persistence of demographic triggers.
Read source
Your take?
EvalsAI safetyAlignmentBenchmarks

Summary generated by Claude — human-verified