We study these dynamics to help design healthier digital spaces.
Online platforms are shaped both by algorithms (what gets recommended, who is followed) and by people (what we share, talk about, ignore). We build tools that measure both sides together, so we can see how attention concentrates, how information spreads, and which voices end up loudest, without reducing it all to engagement counts.
What do people believe, and how do those beliefs change? Surveys can ask, but they're slow and small. We build computational tools that read meaning (framing, stance, belief) directly from online text. Our recent Nature Human Behaviour paper introduced a "map" of human beliefs that can predict which view a person is likely to adopt next, given the ones they already hold.
Toxicity, hate, and misinformation are usually not the work of lone bad actors. They are shaped by competition, group norms, and the design of the platform itself. We study toxic behavior in online games and communities, how false stories spread, and how moderation rules change what gets said, including the harder question of what to do when different communities disagree on what should be allowed.
LLMs are increasingly used for moderation, counseling, and policy work, tasks that demand social and cultural judgment, not just accuracy. We test how reliably they detect hate speech, how easily their stated beliefs can be talked out from under them, and whether they can simulate public opinion across different cultures. We also study how trust in AI varies by profession and country, because the same model does not land the same way for everyone.
When something goes wrong with collective wellbeing (a pandemic, a vaccination controversy, a mental-health crisis) it shows up in online discourse before it shows up in clinics. We work with health researchers to read those signals, study how language around illness differs across cultures, and build datasets and models that help campaigns reach the people they're meant to serve.
Most AI evaluations test a single answer to a single question. But real use is back-and-forth. We're developing new ways to measure how an AI's beliefs, norms, and emotional tone shift as a conversation unfolds, so "alignment" becomes something we can check over time, not just at one moment.
Bias isn't only a model problem. It shows up in which voices news outlets amplify, which faces get recognised by APIs, which jobs an ad reaches. We measure these asymmetries across media, data, and AI systems, and build pipelines and literacy tools that make them visible to the people they affect.
Some social experiments are too risky to run on real people. Testing counter-messaging against online radicalization, for example. We're working toward AI agents that realistically represent the beliefs and behaviors of specific communities, so researchers and policymakers can simulate before they intervene.
Online games and user-generated platforms are evolving fast, with AI agents and player-made content blurring the line between user and system. We study the design choices that shape harmful behavior at scale, from toxicity in team-based games to risks in children's media.
To see what we have been working on, browse our publications.
