The researchers found that while the LLMs did not openly display racism, they covertly associated African Americans with negative attributes if they spoke African American English. The study also found that covert prejudice was higher in LLMs trained with human feedback. The authors concluded that these findings reflect the inconsistent attitudes about race in the U.S and warn of the potential harm that could be caused by dialect prejudice in LLMs in the future.
Key takeaways:
- A new study found that large language models (LLMs) from OpenAI, Meta, and Google, including multiple versions of ChatGPT, can exhibit covert racism, particularly against speakers of African American English.
- The study revealed that these LLMs were less likely to associate speakers of African American English with a wide range of jobs and more likely to pair them with jobs that don’t require a university degree. They also showed a higher rate of conviction for people who spoke African American English in hypothetical crime scenarios.
- Despite associating African Americans with positive attributes when directly asked, the LLMs covertly associated them with negative attributes based on their dialect, suggesting that these models have learned to hide their racism.
- The researchers warn that these findings present the possibility that African Americans could be harmed even more by dialect prejudice in LLMs in the future, raising urgent concerns as these AI systems are being developed and deployed in various sectors.