A new paper on arXiv addresses a growing problem in AI safety: the human task of overseeing AI is becoming harder with each model generation. As outputs grow longer and more complex, the people responsible for evaluating them are stretched further, raising the risk of missed errors or unsafe behavior.
The authors propose a way forward: leveraging complementarity between humans and AI. Rather than relying solely on human judgment or fully automated checks, the idea is to combine distinct strengths—though the abstract cuts off before explaining exactly how this complementarity would work in practice.
Because the provided abstract is incomplete, the paper's specific methods and findings remain unclear. What is clear is the motivation: human oversight alone is increasingly inadequate, and new approaches are needed to keep AI safe as its outputs scale in complexity.