Study probes whether language models judge morality or just spot moral cue words
A study of 14 language models examined whether interpretability methods identify attention heads and sparse autoencoder features that genuinely encode normative judgment or merely detect moral cue words.
Read the full story at arXiv cs.CL ↗