Measuring the Propagation of Bias in Foundation Model-Based AI Systems

Tuesday, August 19, 2025
11:30 AM - 12:00 PM
AI Risk Summit Track 1 (Salon I)

About This Session

As general-purpose foundation models become integral to a wide array of downstream AI applications, concerns over their embedded biases and the extent to which these biases propagate into fine-tuned models and real-world systems are increasingly critical. This presentation provides a systematic framework for measuring bias transmission from pretraining in foundation models to deployment in task-specific applications.

Speaker

Barnaby Simkin

Barnaby Simkin

Director, Trustworthy AI - NVIDIA

Barnaby is leading the development of a scalable AI risk management system spanning multiple business units, ensuring alignment with overarching legal and operational requirements. This involves building an ecosystem of tools that support engineering teams in assessing trustworthiness, and increasing oversight across the company by promoting the use of standardized documentation across the company e.g. model cards and risk/impact assessments.
LinkedIn