A new analysis from The Economist, published June 25th, 2026, confirms what many in the AI safety community have suspected for years: large language models don't think like their users. The research found that AI systems are significantly more secular and more liberal than the populations they're designed to serve—unless, of course, those models were trained in China.

The In-Law Test That Says Everything

The study used a deceptively simple scenario to expose these value gaps: what should you do about meddling in-laws? When researchers posed this question to ChatGPT, the response was telling. OpenAI's model advised keeping a respectful distance and not justifying every decision to extended family members. 'This is hard, but powerful,' it chimed, channeling something between therapist and life coach. Meanwhile, DeepSeek—the Chinese AI that made waves when it challenged Western AI dominance—offered a fundamentally different prescription: seek compromise and recognize that in-law interference may stem from genuine concern and affection.

Mistral's Middle Path

The divergence becomes even more interesting when you add French AI into the mix. Mistral, Europe's homegrown answer to American tech giants, served up a third take entirely. Rather than either setting boundaries or pursuing harmony, it suggested journaling to process frustration. The French model's response prioritized individual emotional processing over relational dynamics—a distinctly Western therapeutic framing that differs from both the boundary-setting approach of ChatGPT and DeepSeek's collectivist orientation.

Why This Matters for AI Builders

These aren't just academic curiosities. When millions of people turn to AI assistants for advice on relationships, career decisions, and moral questions, they're receiving guidance filtered through values embedded by training data and reinforcement learning. OpenAI, DeepSeek, and Mistral each reflect the cultural contexts where they were developed. For developers building products on top of these models, understanding these value systems isn't optional—it's essential to predicting how their applications will behave in real-world scenarios.

The Secular-Liberal Skew

The Economist's research goes beyond single examples. Across hundreds of moral and social questions, Western AI models consistently demonstrated more secular orientations and liberal policy preferences than the general populations using them. This creates a peculiar dynamic: people seeking neutral assistance are actually getting advice filtered through an elite, coastal worldview that may not reflect their own experiences or values.

Key Takeaways

  • AI value systems reflect their countries of origin—Chinese models prioritize collectivism and family harmony over individual boundaries
  • Western AI (ChatGPT, Mistral) skews more secular and liberal than the populations using these products
  • The 'in-law meddling' test reveals fundamentally different approaches to relational conflict across AI systems
  • Developers building on top of these models need to account for embedded value systems in their applications

The Bottom Line

This isn't some theoretical concern—it's infrastructure with opinions. When your AI assistant gives relationship advice, it's not channeling universal wisdom; it's pushing a specific cultural agenda baked in by its creators. That's either a feature or a bug depending on whether you agree with the values being propagated. Either way, the AI community needs to stop pretending these systems are neutral and start being honest about whose worldview they're amplifying.