Human-AI Augmentation Gap

New research summarized by Ethan Mollick suggests that models which perform a task well on their own are not necessarily the models that help a person perform it better. This is a concrete challenge to agent-first evaluation: optimizing autonomous completion may reduce the quality of human-AI collaboration. The finding is currently represented by one research report. Independent task studies, workplace experiments or augmentation-specific benchmarks would confirm whether a separate coworker-model category is needed.