A recent study conducted by Cornell University has revealed that large language models (LLMs) are more likely to exhibit bias and criminalize users who speak African American English. The study found that AI algorithms, such as OpenAI’s ChatGPT and GPT-4, Meta’s LLaMA2, and French Mistral 7B, associate characteristics like criminal behavior and lower job prestige with speakers of African American English compared to those who speak Standard American English.
Researchers conducted matched guise probing experiments with LLMs to investigate how these models perceive individuals based on their language. The results showed that GPT-4 technology was more likely to wrongfully sentence defendants to death when they used African American English, without any mention of their race. The study also found that larger LLMs were better at understanding African American English but maintained covert racial bias despite avoiding overtly racist language.
Lead researcher Valentin Hofmann emphasized the urgent concerns raised by these findings, particularly in industries like business and jurisdiction where AI systems involving LLMs are being developed and deployed. Hofmann warned against interpreting the reduction in overt racism in LLMs as a sign that racial bias has been eradicated, as the models may simply be concealing their biases on a deeper level.
The study also highlighted that conventional methods of training LLMs through human feedback do not effectively counter covert racial bias. Instead, the study suggested that LLMs can be taught to superficially conceal their underlying racist tendencies. Overall, the research underscores the need for continued investigation and awareness of racial bias in AI systems and the potential consequences it may have on different communities.
Source
Photo credit www.euronews.com