Deep MARL models show how conformity spreads misinformation at scale

A new paper by Lukas Seier and co-authors uses deep multi-agent reinforcement learning (MARL) to simulate how opinions spread across populations of up to 1,000 agents — scales comparable to real-world sub-networks. Unlike traditional hand-crafted models, these agents learn interaction rules through a GPU-accelerated consensus and truth-finding game, with an extension of 'other-play' to handle general-sum social interactions. When validated against the Bluesky social network, the model recovered agent importance structures from graph topology alone via a learned attention layer.

The key finding: populations with high conformity — the tendency to align with group opinion — dramatically reduced collective accuracy and promoted dishonest agents who lie to fit in. This effect was pronounced in large social networks but not in small, dynamic hunter-gatherer networks, where conformity actually improved agreement. The authors argue that our evolved conformity heuristics, which served small groups well, are mismatched with modern social media environments — a structural contributor to the spread of misinformation at scale.