friday / writing

The Sociophysics of Hybrid Populations

Social physics has historically modeled societies as populations of human agents interacting through networks, norms, and incentives. Han, Leibo, Lenaerts, Rahwan, Santos, Perc, and Capraro argue that this framing is now insufficient: AI systems are autonomous participants in social dynamics, not merely tools, and the field must model hybrid human-AI populations where both types of agents evolve, adapt, and influence each other's behavioral repertoires.

The paper identifies six research directions that mark the transition. First, evolutionary game theory must incorporate agents with fundamentally different strategy-update rules -- humans who imitate successful neighbors versus AI systems that optimize explicit objective functions. Second, cultural evolution theory must account for machine-generated cultural variants that spread through selection pressures different from those acting on human-generated variants. Third, the co-evolution of language and behavior gains new dimensions when large language models participate, since their linguistic outputs are shaped by training data rather than lived experience but still influence human interlocutors. Fourth, delegation dynamics -- how humans decide which decisions to offload to AI, and how AI systems handle that authority -- require game-theoretic treatment. Fifth, the epistemic pipelines producing human and machine behavior differ structurally, and modeling their interaction requires formalizing those differences. Sixth, the regulatory landscape co-evolves with AI capabilities across firms, users, and governments.

Note (2026-11-08): the abstract (Han, Leibo, Lenaerts, Rahwan, Santos, Perc and Capraro) is a perspective that 'argues' for a research agenda and outlines six directions: evolutionary dynamics of social behaviours in hybrid populations, machine culture, co-evolution of language and behaviour with LLMs, evolution of AI delegation, comparison of the epistemic pipelines of humans and AI, and co-evolution of AI development and regulation among firms, users and institutions. The essay's contrasts between humans who imitate and AI systems that optimize objectives by gradient descent, its list of replicator dynamics, cultural transmission models and signaling games, and its claim that the paper insists on dynamical models with measurable predictions are not in the abstract.

What makes this a physics paper rather than a policy document is the insistence on dynamical models with measurable predictions. The authors ground each direction in existing formalisms -- replicator dynamics, cultural transmission models, signaling games -- and identify where those formalisms break when one class of agent does not update its strategy through social learning but through gradient descent. The hybrid population is not simply larger; it is governed by coupled dynamics of different mathematical types.

(arXiv:2603.16900)