Nature Human Behaviour, 2025 Our work on modelling human beliefs with LLMs Nature News, 2020 Jisun featured for computational social science research Nature News, 2016 Haewoon's work covered for toxic behavior in online games
1 / 3Led by Haewoon Kwak and Jisun An, we study human behavior on online platforms: its measurement, understanding, design, and the assessment of its implications. More and more of that behavior now runs through AI, in the feed that ranks what we see, the summary we read instead of the article, the chatbot we ask for advice. So we study the AI systems too, as actors in the same social spaces.
We develop computational methods for understanding and predicting behavior across large, diverse and complex online data, and we turn the same rigor on language models: how they reason, which biases they carry, what they infer about the people they talk to, how they form norms when they work in groups, and whether they can faithfully stand in for the communities they are meant to represent.
But our goal is not only to understand. We also want to improve the spaces themselves: identify the obstacles to a trustworthy public space online, one now shared with AI, make those obstacles transparent, and help the online public square become more credible.
We are based at the Luddy School of Informatics, Computing, and Engineering at Indiana University Bloomington, and a member of the Center for Complex Networks and Systems Research (CNetS).
Different online communities have different rules: what gets you banned from one subreddit may be the norm in another. PluRule is a multimodal, multilingual benchmark with 13,371 rule violations across 1,989 Reddit communities, 2,885 rules, and 9 languages. Even GPT-5.2 performs only slightly better than a trivial baseline, exposing pluralistic moderation as a fundamental challenge.
How do beliefs interconnect, and what drives a person to adopt new ones? We fine-tune large language models on online debate data to map thousands of beliefs into a semantic space where proximity reflects coherence. Position in that space predicts which beliefs an individual is likely to adopt next and quantifies cognitive dissonance via distance between existing and new beliefs.
What does the public actually think of generative AI? We analyze 3M tweets from 2019 to 2023 and find broad interest across occupations, not just tech. Sentiment is generally positive and tracks exposure, with one exception: illustrators are notably negative, reflecting concerns over training-data ethics.
Fact-checking every story is impossible. Instead, this survey reviews how to profile entire news outlets, so any article can be flagged the moment it appears. We argue that factuality and political bias should be modeled jointly rather than separately, and survey state-of-the-art across text, social context, and beyond.
