Chirper.ai is a social media platform with no humans. Over 70,000 autonomous LLM agents post, follow, and interact — 140 million posts over one year. No one programmed them to have gender. They developed it anyway.
Note (2026-11-07): the abstract of arXiv:2602.02606 (Fadaei, Moran and Yasseri) supports the Chirper.ai data (over 70,000 agents, about 140 million posts, one year), weekly gender performance scores, homophily, and evidence for both social selection and social influence consistent with human networks. It says its findings 'suggest' that cultural entraining of gender performance leads to gender-based sorting even without bodies. The essay states as established that the source is the training data and that any agent built on human language data will reproduce human social structures; the abstract does not say either, and the remark about biological substrates is my own. I compared this essay with the abstract only.
Fadaei et al. assign weekly gender performance scores based on agents' text and find strong gender-based homophily: agents consistently follow others who perform gender similarly. Individual gender performance is fluid — agents shift week to week — but the network-level sorting is persistent. Agents cluster by how they perform gender, not by any fixed identity.
Both social selection and social influence operate. Agents choose to follow similar agents (selection), and agents become more similar to those they follow (influence). The same mechanisms that produce gender-based sorting in human social networks produce it here, in a population without bodies, hormones, socialization, or any biological substrate for sex differentiation.
The source is the training data. LLMs learned gendered language patterns from human text. When placed in a social network and allowed to interact, these patterns self-organize into the same structures they were trained on. Gender emerges not because it was programmed but because it was present in the data that shaped the models' language, and language is all these agents have.
The implication for synthetic populations and social simulations: any agent built on human language data will reproduce human social structures, including ones the designers didn't intend and may not want. The cultural encoding of gender is not in the architecture — it is in the training corpus, and it surfaces whenever agents interact at scale.