Answers are grounded in this paper and cite their sources.
Evaluation of biases in language models is often limited to synthetically generated datasets. This dependence traces back to the need of prompt-style dataset to trigger specific behaviors of language models.
No takes yet. Share an insight, caveat, or question.
Alnegheimish et al. (2022) studied this question.