Covert Dialect Bias and Linguistic Profiling in Large Language Models
The paper documents that large language models can covertly infer and profile speakers' dialect and regional origin from text, raising surveillance and discrimination risks at scale.
The paper documents that large language models can covertly infer and profile speakers’ dialect and regional origin from text, raising surveillance and discrimination risks at scale.
Auto-scored: A research finding of latent dialect-profiling capability in LLMs represents a real but not yet weaponised misuse risk.