HCII Seminar Series - Hoda Heidari
Speaker
Hoda Heidari
K&L Gates Career Development Assistant Professor of Ethics and Computational Technologies at Carnegie Mellon University
When
-
Where
Newell-Simon Hall 1305
Video
Panopto
Description
"GenAI Evaluations: The Broken Bus from Transparency to Accountability"
Evaluations have become a central mechanism for governing generative AI systems, underpinning claims about transparency, safety, and accountability. This talk begins with an overview of the contemporary GenAI evaluation landscape, highlighting persistent challenges related to validity, reliability, scalability, and—critically—actionability. Focusing on actionability, I ask whether existing evaluation practices meaningfully support the decisions that key stakeholders must make about AI adoption, deployment, and oversight.
Drawing on evidence from two recent empirical studies, I argue that they often do not. First, I present findings from an interview study of GenAI developers documenting models on open-source platforms, which reveals substantial uncertainty about what evaluations should communicate, for whom, and for what purposes. Second, I examine how local government agencies engage with vendor-provided evaluations and documentation during AI procurement, showing how information asymmetries, disclosure constraints, and institutional context limit the practical utility of evaluation results.
Together, these studies suggest a disconnect between evaluation as a transparency practice and evaluation as a mechanism for accountability. I conclude by discussing implications for future research and practice, including the need to reconceptualize GenAI evaluations as sociotechnical and institutional artifacts rather than purely technical measurements.
Host
Hong Shen