Define, detect, report, coordinate: what epidemiology offers the Paris Call

November 18, 2026

I’m an epidemiologist currently exploring ways in which epidemiology could be useful to AI harm monitoring, and useful to the MIT AI Risk Initiative. So I know I’m in danger of having a hammer and seeing only nails, but I dare any epidemiologist to read the third commitment of the Paris Call for AI Safety and Security and not think ‘well, that sounds like a job for me’.

On the 9th of November, 18 heads of state, and representatives of AI companies and civil society signed a declaration committing to lay the necessary foundations for safe AI adoption. The first commitment is to a common approach to evaluating and managing the risks of frontier AI models, the second to responsibility and accountability across the AI ecosystem, and the third, at least to my biased epidemiologist eyes, is to treating AI harms as public health problems to be defined, measured, and mitigated.

Epidemiology is basically the science of counting exposures, behaviours and pathogens that are diffuse throughout society and therefore difficult to count. Given how diffuse and difficult to count AI harms are, I’ve been thinking for a while that epidemiological principles and methods could be useful to the field of AI harm monitoring. The Paris Call has me feeling vindicated. It declares that we need:

• Alignment on incident classification, which screams ‘case definition’ to me – a standardised set of criteria used to identify and classify cases in outbreak investigation

• Early warning systems to detect precursor signals of serious harm, which is epidemiology’s bread and butter. An example would be wastewater monitoring, which on a bad day tells you there’s cholera out there somewhere nearby.

• Reporting arrangements to share timely information. ‘Do you even share timely information along the appropriate channels?’ is the ‘Do you even lift, bro?’ of epidemiology.

• Exploring a coordination mechanism, which sounds a lot like the start of a WHO for AI safety

In public health, all of this defining, detecting, reporting and coordinating is in service of effectively preventing further harm. Let’s take the 2014 Ebola outbreak in West Africa as an (idealised) analogy of the type of effective action that the Paris Call commits to building towards.  A cluster of unexplained deaths was detected, investigated, and once Ebola was confirmed by labs, this was reported to the WHO, which declared an outbreak. This declaration triggered a pre-agreed international response, and hundreds of experts ranging from clinicians to anthropologists were mobilised to contain the epidemic – with contracts pre-arranged through effective coordination mechanisms.

At MIT’s AI Risk Initiative, I’m currently exploring which epidemiological approaches could be used for AI harm monitoring, and how. As a first step I’m interviewing AI-adjacent epidemiologists, and as a second, people working in AI harm monitoring. Please get in touch if you’d like to be one of them – you’ll be contributing to research unknowingly endorsed by the Paris Call.

Contact Branwen Owen at owen_b@mit.edu

‍

Featured blog content