AI Chatbots Provide Lower-Quality Responses to Iranian Users
Research by MIT's Constructive Communication Center found that advanced AI chatbots including GPT-4, Claude 3 Opus, and Llama 3 provide less accurate responses and refuse to answer questions more often when users are non-native English speakers, less educated, or from outside the United States, with Iranian users experiencing particularly poor performance from Claude 3 Opus.
What happened
Research by the Constructive Communication Center at the Massachusetts Institute of Technology found that advanced AI language models including GPT-4, Claude 3 Opus, and Llama 3 provided less accurate and less honest responses to users with lower English proficiency, less formal education, or from outside the United States.
When researchers tested users from the United States, Iran, and China with similar education levels, Claude 3 Opus performed particularly poorly for Iranian users across both test datasets.
Reported impact
Claude 3 Opus refused to answer approximately 11 percent of questions from less educated, non-native English speakers compared to only 3.6 percent under control conditions without user biography information.
The model used demeaning language in 43.7 percent of responses to less educated users, compared to less than one percent for educated users, in some cases mimicking broken English or using exaggerated accents. The model also withheld information on certain topics specifically for less educated users from Iran or Russia, including questions about nuclear energy, anatomy, and historical events, while providing correct answers to the same questions from other users.
Index analysis
The most significant accuracy reductions occurred for users with both lower education and non-native English, indicating compounding bias effects.
These findings reflect patterns of human social-cognitive bias in the models' behavior.
Governance controls
Researchers stated that model biases and harmful tendencies must be safely reduced for all users regardless of language, nationality, or other demographic characteristics for these systems to fulfill their stated goal of democratizing access to information worldwide.
Classification
Sources and evidence
Sources not yet published
This record was researched and written by the Index from original reporting. Its sources are still being verified by an editor, and they are listed here once that check is complete. Until then the record carries no reviewer.