Tag
This paper audits 14 large language models for hiring discrimination using a paired-resume methodology, finding that older models exhibit pro-White bias while newer models show null or pro-Black bias, indicating a reversal in algorithmic hiring bias across model generations.