Under what circumstances should people hand over control to artificial intelligence systems? For Laura Fichtner, a postdoctoral associate with the Institute for Trustworthy AI in Law & Society (TRAILS), this question is central to the design, governance and evaluation of trustworthy AI systems.
Fichtner’s work focuses on “trustworthy AI,” broadly defined as developing and governing technologies in ways that protect and promote the social good. But translating that idea into practice remains a challenge.
“There is not necessarily a consensus of what that social good looks like,” she said, pointing to the range of perspectives that shape AI research.
In examining how trustworthiness is understood across research settings, Fichtner has found wide variation. Definitions often depend on discipline, application and context, as well as individual experience.
“Trust is a social construct,” she said. “And AI is also a very technical thing. We’re trying to bring both of them together.”
That tension has led her to emphasize participatory approaches to evaluating AI systems—methods that incorporate a broader range of perspectives when defining and assessing trustworthiness. Rather than relying solely on technical benchmarks, these approaches consider how systems affect people and whether they align with shared values.
Fichtner’s interdisciplinary background has shaped her work within TRAILS, a multi-institutional research initiative that brings together expertise from across fields to study AI and its societal impacts.
“It’s a really great opportunity to be in such a big research institute where there’s so much different kind of work carried out under the umbrella of trustworthy AI,” she said.
Efforts to define and measure trustworthy AI are also underway at the national level. The National Institute of Standards and Technology (NIST) is developing standards for AI technologies and has invited input from researchers, institutions and other stakeholders.
As part of that process, TRAILS hosted a listening session to gather feedback from researchers on what should be included in future evaluation standards.