Responsible AI in the Generative Era with Michael Kearns - 662

AI Evolution

Michael Kearns, a professor at the University of Pennsylvania and an Amazon scholar, highlights the transformative impact of generative AI on responsible AI practices. He notes that the open-ended nature of these models, which allows them to generate rather than predict, introduces new challenges in ensuring responsible AI. This shift requires a reevaluation of traditional AI metrics and practices to accommodate the unique capabilities and risks of generative models 1. Kearns emphasizes the importance of adapting to these changes, stating:

The power of these models is in their open-endedness. They're not making numerical predictions about inputs or point predictions or solving classification problems. They are truly generative.

--- Michael Kearns

Service cards, which provide summaries of model properties and recommended use cases, have evolved to become more informative and sophisticated, reflecting these new challenges 2.

   

Societal Influence

Societal elements such as activism and audits play a crucial role in shaping responsible AI. Michael Kearns discusses the AI activism movement, which he views as a positive force in the industry, encouraging transparency and accountability through external audits 3. This movement includes initiatives like bias bounties, inviting external communities to influence AI principles cooperatively rather than adversarially. Kearns explains:

It's anticipating what journalists, nonprofits, researchers might do in a less than friendly audit of your model.

--- Michael Kearns

He also shares insights from his work at AWS, highlighting the importance of understanding modality in data and the challenges of ensuring fairness without explicit demographic data 4.

   

New Challenges

The generative AI era introduces new challenges in responsible AI, particularly in adapting metrics and practices. Michael Kearns points out that traditional models did not face issues like hallucinations, which are now prevalent in generative AI 5. This requires new metrics and qualitative guidance to address these challenges effectively. Kearns notes:

There are also things that we still don't have a quantitative handle on yet. And so we have to think about those things more qualitatively.

--- Michael Kearns

He also discusses the balance between specificity and generality in model performance, highlighting the need for tailored metrics for specific use cases to optimize the benefits of generative AI 6.