Semantics derived automatically from language corpora contain human-like biases

On the reading lists

• #51 in Safety, alignment & interpretability (2026 edition)

Why it matters

Introduced WEAT and demonstrated that language embeddings reproduce measurable human social biases.

Condensed from the expert lists' notes.

The record

First released

2016 (from arXiv 1608.07187)

Published

2017

Citations

3,338 citations, as counted by Papers With Code (CC BY-SA 4.0)

Named by

the chatgpt list only

Score

0.339 · computed rank 505 · consensus 0.31, must-read 0, citations percentile 0.702 (2010s · safety), awards 0, existing lists 0. Method