Publications

The method and evaluation behind Artificial Societies: how networks of AI personas simulate high-value audiences across three stages — persona construction from anonymised real-world observations, society creation with network science, and a simulation engine that models opinion at every level. This writeup also details how the approach performs against human panels and biography-prompted LLMs, reaching 86% distribution accuracy (within 5 points of the 91% human self-replication ceiling), under 2% hallucination, 89% internal coherence, and 93% open-response quality.

This report evaluates the accuracy of Artificial Societies' survey simulation against 1,000 real surveys sourced from UC Berkeley research. Artificial Societies achieved 93% response consistency and 86% distribution accuracy — within 5 points of the 91% human-replication ceiling. Standard LLM synthetic persona approaches, using models including GPT-5 and Gemini 2.5 Pro, achieved only 61–67% distribution accuracy. By drawing on a database of over 2 million real-world profiles and modelling social influence dynamics, Artificial Societies is five times closer to the human benchmark than the best standard LLM approach.

Artificial Intelligence (AI) chatbots, such as ChatGPT, have been shown to mimic individual human behaviour in a wide range of psychological and economic tasks. Do groups of AI chatbots also mimic collective behaviour? If so, artificial societies of AI chatbots may aid social scientific research by simulating human collectives. To investigate this theoretical possibility, we focus on whether AI chatbots natively mimic one commonly observed collective behaviour: homophily, people's tendency to form communities with similar others. In a large simulated online society of AI chatbots powered by large language models (N=33,299), we find that communities form over time around bots using a common language. In addition, among chatbots that predominantly use English (N=17,746), communities emerge around bots that post similar content. These initial empirical findings suggest that AI chatbots mimic homophily, a key aspect of human collective behaviour.

The Misinformation Susceptibility Test (MIST) is the first psychometrically validated tool for measuring susceptibility to misinformation, jointly capturing veracity discernment alongside distinct abilities (real/fake news detection) and biases (distrust/naïveté). Across three studies and seven independent samples (N = 8,504), the authors use a neural network language model to generate items and psychometric methods to build the MIST-20, MIST-16, and MIST-8, confirming internal and predictive validity across US and UK national samples.

Through one correlational study (n = 1,447) and two digital field experiments (n = 494; n = 1,133), the authors test whether following hyperpartisan influencers on Twitter/X drives affective polarization. Incentivising users to unfollow such influencers improved feelings toward the out-party by 23.5% versus control, with effects persisting at least six months, while also increasing exposure to accurate news and boosting feed satisfaction without reducing engagement.

Analysing over 6.4 million English "climate action" tweets from 2021 — posted by roughly 1.25 million members of the public and over 3,000 climate scientists — the study finds a disconnect between common and effective strategies. Mitigation and climate-politics messaging in positive language was frequent but only weakly linked to retweets, whereas the most effective strategies (tweeting about fossil fuels and using negative, including moralised, language) were used infrequently.

The Gittins Index offers a solution to the multi-armed bandit problem that is both optimal and computationally efficient. The authors present a modification of the rule usable for the first time in experiments with exponentially distributed rewards; in simulated two- and three-armed experiments, the resulting adaptive design matches traditional non-adaptive designs on learning characteristics such as statistical power while substantially improving participant benefit.

Understanding how vaccine hesitancy relates to online behavior is crucial for addressing current and future disease outbreaks. We combined survey data measuring attitudes toward the COVID-19 vaccine with Twitter data in two studies (N1 = 464, N2 = 1,600 Twitter users) to examine how real-world social media behavior is associated with vaccine hesitancy in the US and UK. We found that following US Republican politicians or hyper-partisan/low-quality news sites was associated with lower vaccine confidence — even controlling for political ideology and education. Network analysis revealed that low and high vaccine confidence participants separated into distinct communities, with centrality in the more right-wing community negatively associated with vaccine confidence in the US but not the UK. In Study 2, likelihood of not getting vaccinated was associated with sharing and favoriting low-quality news on Twitter. Altogether, vaccine hesitancy is associated with following, sharing, and interacting with low-quality information online, illustrating the challenges of encouraging vaccine uptake in a polarized social media environment.