Science communication with and under AI

Lecture

Government, the public and other scientists during COVID-19, software as communication, and what the person is for when writer and reader both have a language model.

Combining Infectious Disease Modelling and AI · Monday 14 September 2026, 13:40–14:10 (Europe/Brussels) · Learning Theatre, VUB campus Etterbeek · Sam Abbott

Slides Workshop page How I am LLM epinowcast community EpiNow2

Abstract

During COVID-19 we communicated model outputs to three audiences. Government, through weekly estimates and reports to the UK’s scientific advisory groups and the consensus statements they produced. The public, through a dashboard of reproduction number estimates that ran from 2020 to 2022. Other governments and their scientists used it too. Checking at that scale was very difficult, and we sometimes published estimates that made no sense. Other scientists, through collaborative forecast hubs, where a forecast becomes part of an ensemble and can be hard to learn from. Agreeing how forecasts are scored is communication too.

The software we built was a fourth channel. A package carries the methods, assumptions and defaults of the people who wrote it to every analyst who runs it. Users run the defaults, so the defaults are the advice. The epinowcast community was an attempt to communicate through software and the people around it.

In the second half of the talk we turn to communication in the age of AI. A live outbreak report drafted with agents puts its limitations first, and the feedback is that this is off-putting, so how should we warn? I show how we make agent use visible, with a bot account, a published prompt log and a review bot that reads each pull request and is not a person, and what happens when my bots and other people’s bots file issues on each other’s repositories. Many people now get their information from a language model, so I ask what one says about this work and whether it is right. I end by asking what the person is for when the writer and the reader both have a language model.

The organisers asked for a communication-focused lecture of about 30 minutes, tentatively titled Science Communication with and under AI, as a starting point to make my own. They suggested my experience during COVID-19 with government bodies, the public and fellow scientists, including building widely used R packages for the modelling community, as the material.

Around it

Karolien Poels’s keynote, Risk Communication and Uncertainty in Infectious Disease Modelling and Mitigation, is before this. Andres Algaba’s lecture, under the same title as this one, is after.

Resources

The pandemic

Communities

With and under AI