How predictions drive consensus: AI's Role in Opinion Formation
The evolution of opinions and beliefs in society has long been a central topic in research, ranging from physics to social sciences. Century-old models of opinion dynamics describe how people update their beliefs through peer interactions, explaining when groups converge or disagree. With the rise of AI platforms, however, the rules of information flow in society have fundamentally changed. We interact with our peers not only explicitly through social interactions but also implicitly through the data we share and the predictions we consume from AI models. AI platforms superimpose a new layer of connectivity onto society, enabling human influence far beyond the topology of underlying social graphs.
Suppose you go on Amazon today, browse a sequence of five products, and buy a laptop. A stranger on the other side of the world viewing those same five products tomorrow will likely receive a recommendation for that exact laptop. In effect, the platform implicitly forges a link between you and this stranger even though you have never spoken to each other.
To understand how these dynamics play out at scale, we study the long-term co-evolution of platform predictions and individual opinions. We introduce a tractable model, illustrated above, to capture this recursive loop: a platform’s predictions influence individuals' opinions, which in turn evolve through peer interactions and form the training data for future model updates. This framework bridges machine learning with social dynamics, by embedding the concept of performative prediction directly into classical models of opinion formation.
Through this framework, we show that a predictive platform qualitatively changes how opinions evolve. In particular, even when AI predictions are perfectly accurate, they exert a powerful homogenizing effect on public opinion. This allows consensus to emerge even within networks where classical models of peer influence predict persistent disagreement. This homogenization occurs because the predictive system dynamically adapts to shifting opinions, thereby reinforcing the effects of peer interactions. Simultaneously, the platform’s learning objectives create algorithmic spillover effects that bypass traditional network boundaries, connecting individuals who would otherwise never exchange information. We validate these insights through both theoretical results and empirical investigations.
More broadly, this work highlights a blind spot in how we study modern belief evolution. Models of opinion dynamics typically ignore the algorithm, while models of platform influence ignore the social structure. By modeling both simultaneously, we expose a primary driver of unexplained homogenization observed in networks. Our framework opens several distinct avenues for future research, including how to effectively audit platform power, engineer platform objectives that actively counter unintended homogenization, and examine interlocking predictive loops that emerge when individuals engage across multiple competing platforms.
We hope this framework can serve as a bridge between the machine learning, network science, and tech policy communities, offering a shared vocabulary to measure the true scale of platform influence.
Join us for the upcoming presentation as a contributed talk at NetSci 2026 in Boston. Thursday, June 4, 11:30am ET in session PS 2.2.
The full paper is available here. This article was written by ELLIS Institute authors Rediet and Celestine.
Find out more about Rediet's and Celestine's work.