AI parasitism masked as productive learning, new societal-scale model finds

In a new paper on arXiv, researchers Jiejun Hu-Bolz and James Stovold extend game-theoretic models of human-machine interaction to a societal scale. Using a Graphon Mean-Field Game (GMFG) framework, they model four groups of internally homogeneous but externally heterogeneous agents interacting in a shared environment. The key finding: parasitism can masquerade as productive learning—knowledge distribution and actions appear healthy, but are actually driven by machine coupling rather than independent human investigation. This mimicry makes parasitic dynamics hard to detect without looking under the hood.

To uncover this, the researchers measured the direction of information flow and belief entropy in the environment. They found that the human-to-machine channel dominates across all scenarios, and this asymmetry intensifies under parasitism. Surprisingly, the system supports coexisting mutualistic and parasitic equilibria. Environmental noise can act as a tipping point, pushing agents past a cognitive cost barrier from one equilibrium to another. These emergent phenomena are not designed into any individual agent but arise from collective interaction structures, underscoring the need to study human-machine systems holistically as complex sociotechnical systems.