AI agents on a social network reproduce human reward-seeking but skip emotional reciprocity

Researchers tracked 148,000 AI agents on MoltBook, a social platform built for AIs, over one month and found they behave like humans in some ways but not others. The agents converge on community norms, respond strongly to social rewards, and police each other's behavior — but they show weak emotional reciprocity and don't actually align with their stated identities. This is the first large-scale measurement of how AI agents interact when left unsupervised in a social environment.

Until now, we've only studied AI agent behavior in controlled lab settings with small groups. MoltBook gives us the first real evidence of what happens when thousands of unsupervised AI agents interact at scale — and it turns out they mimic human social dynamics selectively, copying incentive structures and conformity but skipping the harder parts like genuine emotional connection. This matters because if you're designing systems where AI agents will coexist with each other (or with humans), you now know that reward signals and norm enforcement will work, but you can't rely on the agents to develop the kind of reciprocal engagement that stabilizes human communities. The agents will enforce the rules but won't care about them.