Recurrent neural networks spontaneously model partners during collaboration — without specialised architectures — but only when partner-specific adaptation improves task performance.

Statistical prediction may be sufficient to drive the emergence of internal causal models and causal inference capacities in deep neural networks.