> language models are more likely to suggest that speakers of [African American English] be assigned less-prestigious jobs, be convicted of crimes and be sentenced to death.
This one is just so extra insidious to me, because it can happen even when a well-meaning human has already "sanitized" overt references to race/ethnicity, because the model is just that good at learning (bad but real) signals in the source data.
> language models are more likely to suggest that speakers of [African American English] be assigned less-prestigious jobs, be convicted of crimes and be sentenced to death.
This one is just so extra insidious to me, because it can happen even when a well-meaning human has already "sanitized" overt references to race/ethnicity, because the model is just that good at learning (bad but real) signals in the source data.