Q&A: Is AI democratizing global health or reinforcing old inequities?

UNIVERSITY PARK, Pa. — Artificial intelligence (AI) is rapidly transforming the tools that are central to global health decision-making in areas like disease control policies, financing and vaccination strategies, such as infectious disease modeling. This brings new opportunities to the modeling landscape, but could also exacerbate existing disparities, according to Matt Ferrari, professor of biology and director of the Center for Infectious Disease Dynamics at the Huck Institutes of Life Sciences at Penn State.

“There’s an irrational optimism that AI, machine learning and computational tools are going to democratize access to science and policy influence,” Ferrari said. “AI, if treated naively, will likely reinforce existing inequities so we should be proactive about avoiding that.”

Ferrari is the scientific co-chair of the Measles Analytics Hub (MAH), a global network dedicated to reducing the burden of measles and funded by the Gates Foundation. The MAH has developed a framework for equitable research partnerships to develop high-quality contextualized analytical evidence that enables better informed decision making. In a recent commentary in the journal PLoS Global Health, Ferrari and his MAH colleagues shared how what they’ve learned through the MAH could be applied to AI and infectious disease modeling.

In this Q&A, Ferrari spoke about the role of mathematical modeling in public health, promises and pitfalls of AI and large language models (LLMs) in infectious disease modeling, and how intentional collaboration could reshape the future of infectious disease science.

Q: Why do we use mathematical modeling in infectious disease research?

Ferrari: There are lots of questions that can't be answered through conventional observational or experimental techniques — sometimes for ethical reasons, sometimes for operational reasons. There are also some quantities that can’t possibly be measured through observational studies in the real world.

Take vaccination policy. If we want to know whether changing the timing or delivery of a vaccine program would save more lives, we can’t ethically run a trial where we deliberately withhold vaccines from one population to find out. Mathematical models let us explore that problem in a safe, computational space and build a body of evident to justify trying something new.

I think most of us that make mathematical models recognize that we're the last resort. If you can do an experiment or an observational study, do an observational study, do that. When you can’t, that’s when you use mathematical models.

Q: LLMs, machine learning and other AI tools are increasingly used in infectious disease modeling, automating processes and rapidly processing data. Are there concerns about these tools and how they are being integrated?

Ferrari: There’s rapid innovation in computational tools, machine learning and AI, and everyone’s excited. And because they’re excited, they’re charging ahead without picking their heads up and looking at the consequences because they only see the potential. But that potential won’t be realized unless we stop and make these tools ethical and equitable. That means recognizing the innovation, recognizing who’s been left behind and then building a bridge between the two.

Q: You and your colleagues at the MAH have been very intentional about addressing that tension between innovation and who’s left behind. Why?

Ferrari: The MAH was born out of the recognition that no one person or group should be the go-to source of modeling and computational analysis for policy makers and big organizations — like the World Health Organization, the Gates Foundation or Gavi, the Vaccine Alliance. We’re better off if we have multiple groups, multiple perspectives sharing information and working on different approaches to solving big policy problems.

The problem was that policy makers kept going back to the same trusted groups over and over. It meant that we weren’t getting a diversity of ideas or methodologies and, instead, we were getting individual and group-level biases baked into the answers. The groups with the most access to policy makers were disproportionately in North American and Europe. Modeling analyses originated in high-income countries where measles wasn’t endemic, even when they are supposed to inform policy in low- and middle-income countries.

At the same time, brilliant scientists in measles-endemic countries see the epidemiology, the operational and logistical constraints, and the financial realities on a daily basis. They bring valuable context to the problems we’re trying to solve and help direct our models and analyses to practical needs. But they have less access to computational and academic resources and don’t have access to policymakers. They work in isolation.

The MAH’s goal is to intentionally leverage access and opportunities and to bring the community together so we aren’t working in isolation or from afar.

Q: What parallels do you see between what has happened with infectious disease modeling and what is now happening with AI?

Ferrari: There are similar structural challenges and inequities in access to resources and scientific influence through the policy sphere that are getting reinforced through the rapid development of AI.

For example, if you look at the business model for LLMs, for-profit companies develop and release free versions to get people engaged. Once they hit computing capacity, they start charging fees. People who can afford the premium access get cutting edge tools, creating differential access to the best tools. Policymakers who want the best evidence go to the groups using the best tools, resulting in a widening gap between academic haves and have-nots.

There’s also a data problem. Right now, the corpus of data that AI models and tools are trained on barely includes most of the low- and middle-income countries in the world. And therefore, the outcomes that would come from these models are not going to represent those countries. If you want a model that can contextualize responses in a way that can be representative of outcomes in Sub-Saharan Africa, South Asia or the Middle East, you need to train on data from those places.

Q: How has the MAH addressed these equity challenges? What lessons can be applied to AI and LLMs?

Ferrari: With the MAH, we recognize that bringing groups historically left out into the discussion and increasing the scope of expertise, experience and context that we have in the intellectual community will make our science better. For that to work, it requires those of us who have benefited from innovation, academic resources and access to and trust with policymakers to be willing to give away some of that in order to bring others along.

What we build is co-created and co-led with in-country experts and global partners to ensure contextual relevance and equity. We use intentional tools to do this. For example, we cover the costs for researchers from low- and middle-income countries to attend international meetings for networking and mentoring. We regrant money from funders like the Gates Foundation and require that any funded proposal includes researchers from measles-endemic countries.

The work of the MAH gives us a toolkit and roadmap for rebalancing the scales in the AI space. It would require proper and intentional investment — paying for data, paying for the digitization of data — where you’re creating market value for the information from communities that can’t afford licensing fees and for whom that content is their primary valuable resource.

I’m hoping that the development of AI, machine learning and computational tools — and even the business model — is still new enough that there’s an opportunity to get ahead of it and make these tools ethical and equitable now so we don’t have to dismantle a bad business model later.