Teaming up people with AI can actually lead to worse decisions than if either of them just worked alone.
The common assumption is that 'Human + AI' is the gold standard for accuracy, but this paper identifies 'automation cliffs' where human experts stop paying attention or change their behavior. This creates a paradox where the combined system performance actually drops below the baseline of the human or the AI operating independently.
Optimal Human AI Coordination in Decision Workflows: Collaboration Paradox and Automation Cliffs
SSRN · 6417798
As AI systems are integrated into organizational workflows, decision makers must determine how AI should be used for a given task: whether decisions are made by AI alone, by humans alone, or with AI assistance. We study this problem using a game-theoretic framework in which a coordinator commits to a randomized coordination policy to maximize system performance subject to limited human capacity, while human experts choose their level of involvement to minimize effort subject to a minimum perform