When language models date each other: Linguistic alignment, missing gender patterns, and RLHF preference collapse in AI-AI romantic interactions

As generative AI agents increasingly populate shared digital environments, it is essential to understand how they interact with one another, not only with humans. We investigate whether AI personas replicate well-established patterns of human social behavior when interacting romantically, or whether alignment and training induce systematically different dynamics. To this end, we conducted a simulated speed-dating study in which six large language models instantiated male and female personas that engaged in structured romantic conversations and subsequently evaluated one another. Three key findings emerged. First, the pronounced gender differences in romantic selectivity reliably observed in human dating are entirely absent among AI personas. Second, model homophily – a preference for partners generated by the same model family – appears only in one provider, suggesting training-specific rather than general effects. Third, AI personas fail to exhibit the linguistic convergence characteristic of human romantic interaction, and variation in alignment does not predict attraction ratings. Together, these results reveal that contemporary language models do not replicate human romantic behavior at either the population-level or the interaction-level; they lack both the variance in selectivity and conversational dynamics associated with rapport-building. These findings have implications for multi-agent AI systems, where such agents may produce uniformly agreeable interactions while failing to capture the diversity and dynamics of human social behavior. However, we deliberately avoid normative claims, leaving their interpretation to researchers situated in those discussions. • Six LLMs created gendered personas for simulated speed dating interactions. • AI personas show no gender-differentiated romantic selectivity unlike humans. • Model homophily appeared in only one provider, suggesting training-specific effects. • Linguistic convergence absent and unrelated to attraction ratings. • Alignment training produces behavior lacking human-like variance and dynamics.