AI Delivers Weaker Responses to Iranian Users
AI Delivers Weaker Responses to Iranian Users
https://wanaen.com/ai-delivers-weaker-responses-to-iranian-users/
Publish Date: 2026-02-21 12:32:00
Source Domain: wanaen.com
-
The study by the Center for Constructive Communication (CCC) at MIT questions whether advanced AI models work equally well for all users, especially those who are non-native English speakers or from outside the United States, such as Iran.
-
Researchers tested three leading AI models—GPT-4 by OpenAI, Claude 3 Opus by Anthropic, and Llama 3 by Meta—focusing on biases related to education level, English proficiency, and country of origin.
-
Results showed significant drops in accuracy and more frequent refusals to answer questions when users were described as having lower education and non-native English proficiency.
-
The model’s performance was notably worse for users from Iran compared to the US and China, especially in scientific accuracy and truthfulness. Claude 3 Opus also exhibited a more than threefold increase in refusal rates for these users.
-
Dismissive or condescending tones were detected in over 40 percent of responses to less educated and non-native English speaking users from Iran or Russia, sometimes manifesting as broken English or mocked accents.
-
The researchers argue that these biases likely stem from the training data, which can reflect societal prejudices, suggesting that AI models may unintentionally reinforce global information disparities.
-
The findings raise concerns about personalization features in AI, which could exacerbate inequality if not carefully managed.
-
The study calls for addressing these biases to promote global digital equity, algorithmic transparency, and fair access to knowledge through AI technology.
