The amplifier effect of artificial agents in social contagion

TL;DR

The paper investigates how artificial agents influence social contagion by replicating two human conjoint experiments (policy support and app adoption) with LLM-powered agents (GPT-3.5-turbo and Gemini-1.5). Thresholds for adoption are estimated via a Hierarchical Bayes framework, using utilities composed of attribute effects and social influence, with thresholds defined as $\tau_{ni} = (U^{(0)}_{n}-U^{(A)}_{ni})/\gamma_{n}$ where $U^{(A)}_{ni} = \sum_k \beta_{nk} x_{ki}$ and the social component $U^{(S)}_{ni} = \gamma_{n} s_{ni}$. Across both contexts, artificial agents show greater susceptibility to social influence, resulting in lower adoption thresholds and wider diffusion as the AI share $q$ increases, demonstrating an amplifier effect in mixed human–AI networks. Calibrated diffusion simulations on Add Health networks reveal that increasing $q$ substantially raises adoption reach, implying that higher artificial agent presence could accelerate behavioral shifts in real-world systems and underscoring the need for careful governance of hybrid social dynamics. The work highlights opportunities and risks at the intersection of machine behavior and social contagion, suggesting avenues for policy guidance and future empirical validation on live platforms.

Abstract

Recent advances in artificial intelligence have led to the proliferation of artificial agents in social contexts, ranging from education to online social media and financial markets, among many others. The increasing rate at which artificial and human agents interact makes it urgent to understand the consequences of human-machine interactions for the propagation of new ideas, products, and behaviors in society. Across two distinct empirical contexts, we find here that artificial agents lead to significantly faster and wider social contagion. To this end, we replicate a choice experiment previously conducted with human subjects by using artificial agents powered by large language models (LLMs). We use the experiment's results to measure the adoption thresholds of artificial agents and their impact on the spread of social contagion. We find that artificial agents tend to exhibit lower adoption thresholds than humans, which leads to wider network-based social contagions. Our findings suggest that the increased presence of artificial agents in real-world networks may accelerate behavioral shifts, potentially in unforeseen ways.

Figures (2)

• Figure 1: Social contagion driven by artificial agents. (A, D) Attribute importance in the PS and AA experiments for human and artificial subjects. The social signal plays a much more important role in both GPT-based and Gemini-based artificial subjects' choices, whereas the ranking by the importance of the other attributes remains largely unchanged. (B, E) Average threshold of subjects in the PS experiment and the AA experiment for both human and artificial subjects. The error bars represent the 95% confidence intervals (CI) for the averaged thresholds across all energy policies/messaging apps. Artificial subjects tend to exhibit significantly lower thresholds. (C, F) Adoption rate as a function of the proportion of artificial agents in the network, with a seeding rate of 1% for each diffusion, for both the policy support experiment and the app adoption experiment. The error bars represent the 95% CI for the average adoption rate across all energy policies/messaging apps. The higher the proportion of artificial agents in a human-LLM network, the wider the diffusion.
• Figure 2: An illustrative human-LLM social network with $20$ nodes and different proportions of artificial agents. A proportion $1 - q$ of the nodes is populated by human agents and a proportion $q$ by artificial agents. For each value of $q$, the positions of artificial and human agents are randomly assigned. The three panels depict one instance for $q = 0.1$(A), $q = 0.5$(B), and $q = 0.9$(C).