I study the risks and impacts from AI models influencing humans’ beliefs. My research combines large scale human experiments with simulation-based AI evaluations.
Professionally: I’m currently at Transluce, where I lead persuasion and manipulation research projects for UK AISI and the EU AI Office. Previously I’ve consulted for OpenAI on evaluating AI persuasion risks, and for public information campaigns in areas such as vaccination and climate change.
Academically: My PhD was in computational cognitive science at MIT, then at Stanford I researched AI for simulating human experiments. My work has been published in leading journals including Nature, Science, PNAS, and APSR, and I sometimes organize academic workshops such as AIMII.
I’m glad to advise policymakers and labs — reach me by DM on Twitter.
→ Predicting results of social science experiments using large language models
→ AI systems out-persuade expert humans
→ DeliberationBench: A normative benchmark for the influence of LLMs on users’ views
→ Artificial intelligence can persuade people to take political actions
→ The levers of political persuasion with conversational AI
→ GPT-4o System Card: Persuasion
→ How experiments help campaigns persuade voters: evidence from a large archive of campaigns’ own experiments
→ AI systems out-persuade expert humans
→ Predicting results of social science experiments using large language models
→ The AI Epistemic Deference Index: A Continuous Measure of Sycophancy
→ AI Epistemic Risks: Emerging Mechanisms & Evidence
→ Artificial intelligence can persuade people to take political actions
→ DeliberationBench: A normative benchmark for the influence of LLMs on users’ views
→ The levers of political persuasion with conversational AI
→ How will advanced AI systems impact democracy?
→ Outcome-based Reinforcement Learning to Predict the Future
→ Encouraging vaccination using the creativity and wisdom of crowds
→ Large language models are more persuasive than incentivized human persuaders
→ The impact of AI message-testing on public discourse
→ Quantifying the returns to persuasive message-targeting using a large archive of campaigns’ own experiments*
→ GPT-4o System Card: Persuasion
→ How experiments help campaigns persuade voters: evidence from a large archive of campaigns’ own experiments
→ Using survey experiment pre-testing to support future pandemic response
→ Listening with generative models
→ Quantifying the persuasive returns to political microtargeting
→ Emotion prediction as computation over a generative Theory of Mind
→ DreamCoder: growing generalizable, interpretable knowledge with wake-sleep bayesian program learning
→ Rank-heterogeneous effects of political messages: Evidence from randomized survey experiments testing 59 video treatments
→ Hybrid memoised wake-sleep: Approximate inference at the discrete-continuous interface
→ DreamCoder: Bootstrapping Inductive Program Synthesis with Wake-Sleep Library Learning
→ Estimating the Persistence of Party Cue Influence in a Panel Survey Experiment
→ Learning to learn generative programs with memoised wake-sleep
→ Inferring structured visual concepts from minimal data
→ Learning to infer program sketches
→ The Variational Homoencoder: Learning to learn high capacity generative models from few examples
→ Auditory scene analysis as Bayesian inference in sound source models