Skip to main content
Back

UNSW Replaces Human Surveys With LLMs to Map Disease Reporting Gaps

Researchers built an LLM-powered epidemic simulator that reveals how income and education -- not geography -- drive who reports getting sick

Paper, Full Access
  • New ABM framework uses LLMs to simulate whether census-grounded agents report flu-like symptoms, replacing costly surveys.
  • Income and education dominate reporting rate variation; LLM model choice and message framing have smaller but consistent effects.
  • Three scenarios tested: independent reasoning, household influence, and public health message framing -- across San Francisco and Atlanta.
  • Reporting rates range 40-65% across models, producing realistic behavioural spread rather than uniform decisions.
  • Spatial grounding is the key innovation: agents are mapped to census tracts, enabling neighbourhood-level bias analysis in disease surveillance.
  • Paper accepted at KDD 2026; no public code repository announced yet.

Modelling individual decision-making during infectious disease outbreaks is crucial for understanding behavioural dynamics and informing effective public health interventions. The problem is that real surveys are expensive, slow, and often fail to capture the spatial and demographic diversity that actually shapes how a disease spreads. A new paper from researchers at UNSW Sydney and George Mason University proposes a different path: replace survey respondents with LLM-powered agents, grounded in real census data, and let them decide whether to report their symptoms.

Prior work has shown that large language models can simulate realistic human behaviour by generating agent decisions based on demographic prompts and situational context. Earlier approaches relied on rule-based mechanisms to simulate decisions like symptom reporting or vaccination. In contrast, this work proposes a generative approach where agent decisions are produced by large language models prompted with detailed personas and contextual information, including city-specific pandemic scenarios.

The gap that needed filling

Traditional agent-based models (ABMs) -- simulations where individual "agents" follow rules and interact to produce population-level behaviour -- have long been used in epidemiology. Traditional ABMs often rely on simplified rules to represent behaviour, limiting their ability to capture complex decision-making processes and cognitive dynamics. The result is that these models tend to treat all people in a demographic bucket as identical, missing the nuanced, context-sensitive variation that drives real-world outcomes like under-reporting in low-income neighbourhoods.

Recent work has begun to incorporate reporting bias into ABMs, enabling the generation of synthetic datasets that reflect how demographic and behavioural traits influence observed disease outcomes compared to true infections. But those methods still rely on pre-fitted statistical models trained on historical survey data -- which means they can't easily adapt to new diseases, new cities, or new public health messaging.

How the simulation works

The team extended a prior simulation environment in three key ways. First, they incorporated an infectious disease transmission model based on SEIR dynamics, allowing agents to progress through susceptible, exposed, infectious, and recovered states according to physical interactions. Second, they initialised the synthetic population using attributes sampled from real-world census data, providing a realistic demographic foundation for each agent. Third, they replaced the prior regression-based decision mechanism with LLM-based decision-making -- instead of manually defining behaviour rules, they generate agent decisions using LLMs prompted with demographic and contextual information.