Not just what LLMs answer, but how they reason about us.
LLMs as social & cognitive actors
LLMs now sit in counseling, moderation, and policy work, where how a model reasons
matters as much as what it answers. We trace how its moral reasoning switches
framework from one step to the next, where its cognitive biases live inside the
network and how far they can be steered out, how easily persuasion talks it out of
its stated beliefs, and how much it can infer about the people it talks to, from
their politics to the secrets in their inbox.
From raw text to a semantic space of human belief.
Beliefs, opinions & narratives at scale
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 Nature Human Behaviour
paper introduced a "map" of human beliefs that can predict which view a person is
likely to adopt next. We also ask what helps a model predict someone's beliefs,
how faithfully LLMs simulate public opinion in different societies, and how much
interpretive room a country's news leaves open.
Toxicity is not an act; it's an outcome of design.
Online safety, harms & platform governance
Toxicity, hate, and misinformation are rarely the work of lone bad actors. They are
shaped by competition, group norms, and the rules a platform chooses. We have
measured what happened to hate on Twitter once moderation was relaxed, built a
benchmark showing that even frontier models struggle when communities disagree on
what should be allowed, and asked when an altered face starts to count as a
specific person, the point where deepfake harm begins.
Public conversation as a public-health signal.
Digital health & wellbeing
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, test
whether the LLMs entering mental-health care recognise that depression presents
differently across cultures, and build synthetic training data that helps symptom
models generalise to real posts.
Where we started: who gets seen, and who gets heard.
Attention, news & networks
Much of our earlier work asked how attention is distributed: which countries the
world's news covers, how audiences engage with news across platforms, and how
information travels through networks. It runs back to asking whether Twitter is a
social network or a news medium, and the habit it built, measuring the algorithm
and the people together, still runs through everything above.
Measuring how an AI's beliefs drift across a conversation.
Socio-cognitive stability
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.
When data, media, or models speak louder for some than others.
Bias & fairness across systems
Bias isn't only a model problem. It shows up in how people rate an AI partner's
work even when it behaves exactly like a human, and in how a culture's language
genders work and politics across a century. We measure these asymmetries across
models, people, and media, and build tools that make them visible to the people
they affect.
LLM agents that represent specific populations.
Socially faithful digital twins
We have started testing how well LLMs simulate public opinion across societies, and
found that they flatten each demographic group into one "representative" voice.
The goal is agents that faithfully represent the beliefs and behaviors of specific
communities, so researchers and policymakers can test risky interventions, such as
counter-messaging against online radicalization, before trying them on people.
Keeping games and creator platforms accountable.
Safety in participatory systems
Games and creator platforms are where AI turns ordinary fastest. We study who checks
the rules when the company that writes them also runs the championship, what is
lost when an AI metric fills the screen and nobody names its source, and the design
choices behind toxicity at scale and risks in children's media.
