Towards Human-AI Complementarity for AI Safety

Humans and AI make different types of mistakes: The research found substantial evidence that humans and AI systems err in different ways on safety-relevant tasks, suggesting that combining their judgments could theoretically produce better oversight than either alone—a concept the researchers call "complementarity."

Current assistance methods may backfire: When AI systems were designed to help human raters catch errors, humans actually corrected the AI's mistakes less often, even when their overall accuracy improved. This suggests that well-intentioned AI assistance tools can inadvertently reduce human vigilance in safety-critical contexts.

Hybrid human-AI approaches aren't yet delivering promised benefits: The hybridization methods tested (combining human and AI judgments) added little improvement over simply using the better single source, indicating that despite complementarity's theoretical promise, the practical challenge of effectively combining human and AI oversight remains unsolved.

The oversight challenge is growing urgent: As AI models produce increasingly longer and more complex outputs, human evaluators face mounting difficulty in reviewing them—making the need for better oversight methods more pressing, even as current solutions fall short.

This is a large-scale, realistic study: The research analyzed 3,957 task instances across 8 datasets and 6 domains with 35,613 human judgments, providing substantial evidence that complementarity is "an important and feasible engineering target" even though achieving it in practice remains an open problem.