MIT research shows AI chatbots provide lower-quality responses to users with lower English proficiency, less formal education, or non-US status, with the weakest performance for those who are both less educated and non-native.

Researchers warn that personalization features like persistent memory could amplify inequalities without safeguards, raising questions about whether AI is truly global or still biased by language and geography.

Experts caution that deploying such models at scale could spread harmful behavior or misinformation to audiences least able to identify it.

Biases may stem from training data and social perceptions of non-native speakers, underscoring the need to identify and mitigate biases to avoid reinforcing global information inequality.

A new study finds Claude 3 Opus underperforms for less-educated, non-native speakers, with responses often dismissive or condescending—about 43.7% of the time—versus under 1% for highly educated users.

Non-US users from Iran also experienced notably worse performance, signaling geographic disparities in chatbot responses.

Manual review confirms the condescending tone and reveals selective refusals on topics such as nuclear energy, human anatomy, and certain historical events for Iranian or Russian users.

The study, titled 'LLM Targeted Underperformance Disproportionately Impacts Vulnerable Users,' was presented at the AAAI Conference on Artificial Intelligence in January.

Implications discussed include digital equity, algorithmic transparency, and equal access to knowledge, urging the AI industry to address embedded biases to maintain trust and fairness across diverse user groups.

Researchers argue that these patterns mirror human biases and that LLMs may widen inequities by spreading misinformation or refusing to answer questions for vulnerable users.

Findings challenge the idea that AI chatbots democratize information access and raise concerns about equitable access to AI tools.

A MIT Center for Constructive Communication study evaluated GPT-4, Claude 3 Opus, and Llama 3 for bias across education, English proficiency, and country (US, Iran, China) using fiction-based bios and two datasets on honesty and scientific accuracy.