MIT research finds GPT-4, Claude 3 Opus, and Llama 3 provide less accurate responses to non-native English speakers and less educated users.