Benevolent Bias in Multi-Turn Human-Agent Dialogue
It is suggested that fair monitoring of human-agent dialogue must look beyond surface cues to whether the agent's treatment is disparate, as off-the-shelf detectors reliably flag overt bias yet largely miss benevolent bias, LLM judges catch more under more explicit detection criteria but increasingly mis classify neutr...
Qian-Qi Liu, Jin Huang, Fethiye Irmak Doğan et al.
· 0 citations