🧵 Untitled Thread

Anonymous at Thu, 7 Mar 2024 19:33:58 UTC No. 16061697

https://twitter.com/vjhofmann/status/1764687418626576445
>We discover a form of covert racism in LLMs that is triggered by dialect features alone, with massive harms for affected groups.
>For example, GPT-4 is more likely to suggest that defendants be sentenced to death when they speak African American English.
Scientifically speaking, why are language models prejudiced against people who speak Ebonics?
>We find that the covert, raciolinguistic stereotypes about speakers of African American English embodied by LLMs are more negative than any human stereotypes about African Americans ever experimentally recorded, although closest to the ones from before the civil rights movement.
Scientifically speaking, why are they more racist than even pre-civil rights humans?
>How does model size affect dialect prejudice? We find that increasing scale does make models better at understanding African American English and at avoiding prejudice against overt mentions of African Americans, but perniciously it makes them *more* covertly prejudiced.
Scientifically speaking, why does making models larger make it claim to be *less* prejudiced while actually displaying *more* prejudice?