
Chatbots Say Nice Things About Black People, Then Sentence Them to Death Anyway
Ourtake
Anti-racism training taught the chatbots to hide the bias, not to lose it.
What happened
Researchers led by Valentin Hofmann at the Allen Institute for AI found covert racial prejudice in a dozen large language models, including OpenAI's GPT-4 and GPT-3.5. The models linked African American English speakers to negative stereotypes, steered them away from degree-level jobs, and were more likely to convict them and assign the death penalty, even as they showed positive sentiment toward African Americans in general.
Sources
- arxiv.org · Dialect prejudice predicts AI decisions about people's character, employability, and criminality (opens in a new tab)
- newscientist.com · AI chatbots use racist stereotypes even after anti-racism training | New Scientist (opens in a new tab)
- MIT Technology Review · LLMs become more covertly racist with human intervention (opens in a new tab)
- digitalinformationworld.com · AI Chatbots Have a Racism Problem Despite Anti-Racism Training (opens in a new tab)
- knowridge.com · Chatbots persist in using racial stereotypes despite anti-bias training (opens in a new tab)