Common language carries more gender bias than objective metrics show

Evidence

As the authors explain, “Common language is more biased than one would expect based on external, objective metrics.” Baseline aside, however, there is a consistent trend across time that shows the gender bias in word embeddings for professions moving in lockstep with the change in the workforce itself. — Brian Christian, The Alignment Problem, 221

Categorization

• XREF: Directly addresses the reader’s interest in unstated assumptions—even fifty-fifty professions skew male in language, revealing a deeper baseline bias. — connects to: Hundreds of papers missed blatant sexism in word embeddings — Brian Christian, The Alignment Problem, 186; Large-scale English text corpora reflect real-world societal biases; Even massive, comprehensive text datasets reflect biases present in the real world (+5 more)
• QUESTION: What does this ‘male baseline’ in even neutral language say about how we should treat widely-shared cultural datasets?
• corroborated: 2 independent passages support this claim — Brian Christian, The Alignment Problem, 221

Broader Connections

• Large-scale English text corpora reflect real-world societal biases #corroborates (0.80)
• Even massive, comprehensive text datasets reflect biases present in the real world #corroborates (0.80)
• Language can distort as well as transmit human culture #expands (0.78)
• Awareness of sexist discrimination does not equal eliminating sexist oppression #corroborates (0.70) — bell hooks, Ain’t I a Woman_ Black Women and…, 554
• Machine-learning bias often stems from unrepresentative training data #expands (0.65)
• Partial debiasing can make bias problem worse by hiding it #complicates (0.60) — Brian Christian, The Alignment Problem, 210