Human-AI Complementarity: A Goal for Amplified Oversight

​This talk is presented by Rishub Jain and Sophie Bridgers from Google DeepMind.

How do we ensure that humans can continue to oversee increasingly powerful AI systems?

Research on Amplified or Scalable Oversight aims to use AI to improve humans’ abilities to train, evaluate, and monitor AI systems, even as these systems become more capable and potentially surpass human performance in certain domains.

In this talk, our guests will argue that a critical goal for Amplified Oversight is how to achieve human-AI complementarity, that is, how to leverage the complementary strengths of both AIs and humans to generate an oversight signal that is stronger than using AI raters or human raters alone.

This is fundamentally a human computer interaction problem (HCI), and they’ll present learnings from HCI that can help to inform this research.

They’ll also discuss two promising approaches for achieving complementarity:
- AI rating assistance (giving humans an AI that can help to evaluate model outputs);
- hybridization (combining AI ratings and human ratings).

The researchers find positive evidence by combining these approaches, previously studied in isolation, suggesting a path forward for keeping humans in the loop even when AI outperforms humans.

Join BuzzRobot Slack to connect with the community