AI chatbots provide less-accurate information to vulnerable users (news.mit.edu)

Recent research from MIT’s Center for Constructive Communication highlights concerning biases in large language models (LLMs) like OpenAI's GPT-4 and Anthropic's Claude 3, suggesting that these AI chatbots may provide less accurate information to users who are non-native English speakers, have lower education levels, or come from specific countries. The study examined how these models responded to questions from users with varying backgrounds and found significant drops in accuracy for those deemed to have less formal education or lower English proficiency. Notably, responses were not only less accurate but also frequently dismissive or condescending, particularly towards users from regions like Iran. This research underscores critical implications for the AI/ML community, challenging the narrative that LLMs inherently democratize information access. As these models are increasingly deployed, there is a risk of perpetuating and even exacerbating existing social inequities. The findings raise alarms about how LLMs might unintentionally misinform vulnerable groups, emphasizing the need for ongoing scrutiny and adjustments in model training to mitigate biases, particularly as personalization features become more prevalent. As the technology evolves, it is crucial to ensure that it truly serves all users equitably, rather than reinforcing harmful stereotypes and disparities.