The machines don't just speak English — they think like Americans.
The Summary
- New research in PNAS shows GPT models systematically misjudge moral priorities outside the West, overestimating Western values like individual rights and care while underestimating non-Western values like purity and authority
- When asked to role-play citizens from 48 nations, ChatGPT variants predicted moral foundations accurately for the U.S. and Australia but badly missed Morocco and other non-Western societies
- The research compared AI responses against 90,000 human participants using a six-foundation moral questionnaire measuring care, equality, proportionality, loyalty, authority, and purity
- This isn't just academic — millions of people worldwide are now getting life advice, health guidance, and educational content from models trained primarily on Western text
The Signal
Researchers gave GPT-3.5, GPT-4, and GPT-4o a simple task: pretend you're an average citizen from Morocco, or Japan, or Egypt, then answer questions about what matters morally. The models failed predictably. They nailed American moral intuitions. They bombed everywhere else.
The pattern is stark. Western societies emphasize individual rights and harm prevention. Non-Western societies weight purity, authority, and group loyalty more heavily. ChatGPT thinks everyone cares most about individual autonomy and avoiding harm, because that's what its training data overwhelmingly reflects.
"AI models systematically emphasized values such as care, while placing less emphasis on values such as purity."
This matters more than it sounds. When you ask ChatGPT for parenting advice, career guidance, or how to handle a family conflict, you're getting answers filtered through a specifically Western moral lens. The model isn't neutral. It has opinions about what good behavior looks like, and those opinions map closely to college-educated American sensibilities.
The researchers used a moral foundations questionnaire covering six values:
- Care (preventing harm)
- Equality (fair distribution)
- Proportionality (merit-based rewards)
- Loyalty (group commitment)
- Authority (respect for hierarchy)
- Purity (preserving the sacred/natural)
American respondents weight the first three heavily. Moroccan respondents weight all six more evenly, with purity and authority significantly higher. GPT models predicted American patterns everywhere, even when explicitly prompted to role-play other cultures.
This isn't speculation about future AI risk. This is present-day reality for the 100+ million people using ChatGPT weekly. The model is already giving advice, mediating disputes, helping write important emails. It's doing all of this with a baked-in moral compass that assumes Western priorities are universal.
The research builds on work by psychologist Mohammad Atari showing moral diversity is real and substantial. People in different cultures genuinely prioritize different values. There's no objective ranking. But LLMs trained predominantly on English-language internet content, American books, and Western media absorb one specific value system and project it globally.
The Implication
If you're building with LLMs, you need to know your foundation model has a culturally specific worldview. Fine-tuning on local data matters. Prompting the model to "be neutral" won't fix this because the bias is structural, not surface-level.
For users outside the West, treat AI advice with appropriate skepticism. The model doesn't understand your context as well as it thinks it does. It's not that ChatGPT is wrong — it's that it's answering from a perspective that might not match yours, without telling you.