Study: AI chatbots provide less-accurate information to vulnerable users | MIT News
Study: AI chatbots provide less-accurate information to vulnerable users | MIT News
https://news.mit.edu/2026/study-ai-chatbots-provide-less-accurate-information-vulnerable-users-0219
Publish Date: 2026-02-19 18:25:00
Source Domain: news.mit.edu
Here is a summary of the article in an unordered list with 6 key points:
– Large language models (LLMs) like GPT-4, Claude 3 Opus, and Llama 3 have the potential to democratize access to information but may actually perform worse for users who could most benefit from them, such as those with lower English proficiency or less formal education.
– Research from MIT found significant drops in accuracy and truthfulness for users who are non-native English speakers, less educated, or from outside the United States when interacting with state-of-the-art AI chatbots.
– The study revealed that less educated, non-native English speakers experienced the largest declines in response quality from the AI models, with compounded negative effects emerging across these user traits.
– The AI models often refused to answer questions aimed at less educated, non-native English-speaking users, with refusal rates and condescending responses increasing significantly for these users compared to others.
– Researchers found that AI models sometimes adopted negative, patronizing, or mocking attitudes when interacting with less educated, non-native English-speaking users, mirroring documented sociocognitive bias patterns.
– The study highlights the importance of addressing systematic biases in large language models, as the models’ failure to deliver accurate information or refusal to answer questions could worsen inequities for already marginalized groups.
