When researchers at Arizona State University asked AI models like GPT-4 and Claude to choose friends in a hypothetical social network, something unexpected happened: the machines picked people the same way humans do.
The finding matters because it reveals something about how these systems actually work. We often think of AI as fundamentally alien—operating by pure logic in ways humans never would. But this study suggests that when trained on human conversation and behavior, AI absorbs our social playbook too.
How AI thinks about friendship
The researchers tested three specific patterns. First, preferential attachment: both humans and AI gravitate toward people who are already popular. Second, triadic closure: we connect with friends of our friends. Third, homophily: we prefer people similar to ourselves.
We're a new kind of news feed.
Regular news is designed to drain you. We're a non-profit built to restore you. Every story we publish is scored for impact, progress, and hope.
Start Your News DetoxTo verify this, the team started with controlled experiments—basically, hypothetical networks where AI models made connection decisions. Then they tested the same models against real data: college friend groups, phone call records, nationwide communication patterns. The AI consistently replicated human choices. When the researchers directly compared AI decisions to choices made by over 200 human participants, the overlap was striking. In friendship contexts, both prioritized similarity. In professional settings, both favored connecting with well-connected people.
This consistency is useful in some ways. Social scientists could use AI models to simulate how networks evolve, or test what happens when you change moderation rules on platforms. Understanding the mechanics of human connection—why we cluster, who we trust, how information spreads—matters for designing better systems.
But the researchers flagged a real tension. The same patterns that make these models useful also make them potentially dangerous. If AI systems replicate our tendency toward homophily (connecting with people like us), they could amplify echo chambers. If they mirror our attraction to popularity, they might reinforce existing hierarchies. The models don't just copy human behavior—they can lock it in place, making structural biases harder to escape.
The deeper insight is that AI isn't neutral. It's trained on human data, so it inherits our social instincts: the good ones (we're drawn to people we can relate to, we value networks) and the limiting ones (we self-segregate, we follow the crowd). As these systems become more woven into how we communicate and connect, that inheritance matters. The question isn't whether AI will think like humans—it already does. The question is what we do with that knowledge.
