Humans aren’t the only ones with a selective listening problem. A new study from MIT’s Center for Constructive Communication finds that AI chatbots often provide less accurate, less truthful, and sometimes condescending responses to certain users. Specifically, it gives lower quality answers to users who are non-native English speakers, less formally educated, or from outside the U.S., highlighting a troubling gap in accessibility for those who might benefit most from these tools. Testing across science and truthfulness datasets, the researchers found the largest performance drop occurs at the intersection of lower education and non-native English. Models even refuse to answer questions at higher rates or deliver patronizing language, sometimes mimicking broken English. The study suggests that AI may unintentionally withhold knowledge from certain groups, and personalization features like ChatGPT’s Memory could amplify these disparities. The findings echo human biases documented in social science research. How do you talk down a chatbot from talking down?