Epidemics as Dialogue: How Language Models Teach Epidemiologists to Listen to Human Behavior ✛

In March 2020, France emptied out: people stopped going to work, school, and cafés, and the streets filled with silence. Epidemiologists stared at the graphs and understood that the virus was not spreading on its own, but through handshakes, conversations, and metro rides. But how could they predict how many contacts a person would keep if the government announced new restrictions? They sought the answer through an unusual alliance: a language model and a mathematical model of the epidemic.

The language model acts as a translator that reads the news summary and decides how an ordinary person would behave. It is shown what contacts people had before the pandemic and given a description of the current situation: whether schools are closed, whether gatherings are allowed, how many people are in hospitals. The model responds by indicating which contacts it would keep and which it would cut. These decisions are compiled into a table showing how many contacts occur between different age groups. This table is called an age-specific contact matrix, and it is what drives the virus's transmission from person to person in the computer simulation.

The simulation does not just count infections: it sends the results back to the language model. The model sees that hospitals are filling up, and in the next round it cuts contacts even more. This creates a feedback loop where human behavior and virus spread influence each other. Previously, epidemiologists set such reactions manually with rigid rules, but now the language model does it flexibly, based on the text of the policy description. This approach is called GABLE, which stands for Generative Adaptive Behavioral Layer for Epidemics.

The researchers tested three language models: GPT-4o mini, Gemini 2.5 Flash, and Grok 3 mini. All of them reproduced the overall picture: contacts dropped sharply during lockdowns and partially recovered in between. The correlation with mobility data ranged from 0.60 to 0.85. But the age structure differed: for example, during the second lockdown, Gemini 2.5 Flash kept more adult contacts but fewer teenage contacts than mobility data suggested. Still, all models reproduced the overall hospitalization waves with an error of only 2.3–3.6%.

The key result emerged in short-term forecasting over four weeks. GABLE with GPT-4o mini achieved a mean weighted interval score of 1298, compared to 2443 for mobility-based forecasts and 3047–4148 for statistical methods. This means the forecasts became more accurate, and the 95% intervals covered the actual values in 96.2% of cases. Moreover, the method allowed assessing the impact of measures not yet implemented. Delaying a lockdown by two weeks increased contacts by 46% and hospitalizations by 45.8%, while maximum measures reduced them by 17.2% and 14.0%, respectively.

Previously, it was unclear how to model adaptive human behavior during an epidemic without rigid assumptions and without real-time contact data. Now it is clear that language models can generate plausible age-specific contact matrices that improve hospitalization forecasts and allow comparing the consequences of not-yet-implemented policies. The method was embedded in a feedback loop with a mechanistic epidemic model: the model decides which contacts to keep, and the epidemic responds by showing how that affects spread. This could help epidemiologists and public health authorities quickly assess intervention scenarios when fresh contact data are unavailable. When France empties again, such a tool could help hear how people respond to new rules before hospital beds reveal it.