Can AI help scientists write code they can trust?

University of Michigan School of Information lecturer Elle O’Brien has earned a grant from Schmidt Sciences. Her project will support research on how scientists and software engineers can use agentic AI to develop trustworthy code.

The project, “When Specs Meet Science: How Researchers and RSEs Build Trustworthy Software with AI Agents,” asks a timely question: Can scientists use agentic AI to create software that is reliable, trustworthy and maintainable?

The problem: undetected coding errors

In nearly every field of science, researchers use code to collect data, find patterns and draw conclusions. But that code is often written quickly and informally, which means errors can go undetected and affect published findings. 

“Just because it runs doesn't mean it's doing the right thing,” O’Brien says. “My research suggests that AI tools for coding are getting widely adopted across the sciences, but keeping an eye on how they’re being used and what is changing as a result is really hard.”

The research: spec-driven development

O’Brien’s project will study an approach called spec-driven development, where researchers write out exactly what their code needs to do before AI builds it. Those instructions can then be used to check whether the final code actually works as intended.

The study will follow collaborations between scientists and research software engineers at the Virtual Institute for Scientific Software, a Schmidt Sciences program with sites at Georgia Tech, the University of Washington, Johns Hopkins University and Cambridge University. Over the next two years, O’Brien will use interviews and observations to understand how these teams work with AI coding tools and what makes scientific software more reliable when AI is doing much of the code writing.

For O’Brien, the project grows out of her own training as a computational scientist specializing in auditory neuroscience at the University of Washington.

“When I was a PhD student, I struggled a lot with a certain kind of existential dread — how do I know if anything in my code is really correct?” O’Brien says. “I think we ask a lot out of the public to trust science. And we have to hold up our end of the bargain and make sure it deserves that trust.”

O’Brien said she is excited to study this moment in scientific computing as it unfolds.

“I feel like a fortunate explorer — I’m getting to observe this absolutely seismic moment in computing up close,” she says. “I fully expect to see things I’ve never seen before.”

RELATED

Learn more about Elle O’Brien’s by visiting her UMSI faculty profile.