Amount Awarded: $25,000
Trust in AI concerns human willingness to accept Artificial Intelligence (AI) as a guide or tool, to have confidence in its outcomes, and to share data and information for its operation and improvement. Our research groups in Oslo and Berkeley work to understand and respond to the demands of trustworthy AI: to develop systems that are worthy of people’s trust. Veridical AI is a central component of our efforts. The novelty of this project is the framing and development of the new methods and practices of veridical AI.
Veridicity is the quality of being true, honest, accurate. Achieving this in the context of AI is challenging. Current systems are rarely verdical as decisions about data (imputations, resolving data conflicts, ownership, data privacy) are often undocumented and poorly justified. Drawing on Bin Yu’s new framework on veridical data science, based on predictability, computability, and stability (PCS) and documentation, we will investigate veridical AI. PCS principles will enable detailed analyses and adaptations to guarantee trustworthy results, addressing the capacity to interpret and apply findings to new data, and stability and consistency across different analytical approaches.
Moreover, our proposal underscores the necessity of uncertainty quantification (UQ) in AI's trustworthy application. Veridical UQ requires, in our view, a PCS workflow and documentation for the data science life cycle. It requires taking into account other sources of uncertainty beyond the sample-to-sample variability under a data generation model. Veridical AI will help ensure that AI systems remain reliable under new data scenarios or operational contexts. UQ is also integral to AI alignment, ensuring AI’s behaviours align with ethical and legal standards, and can detect deviations (AI "hallucinations"). We will study governance mechanisms that would encourage veridical AI practices, with particular reference to UQ. Veridical AI will be tested, optimised and validated in concrete innovation projects.