New AI Model Reveals When More Help Actually Hurts Output

Researchers from the computer science and game theory community have published a theoretical model on arXiv that exposes a critical flaw in the assumption that more AI assistance always boosts productivity. The paper, led by Ali Aouad, models how human agents with varying skill levels choose how much effort to exert when aided by AI. The key insight: when humans adjust their effort based on AI reliability, or when skill acquisition becomes endogenous, increasing AI help can actually degrade output. This productivity paradox emerges because workers may reduce their own effort too aggressively if they trust the AI too much, or they may fail to develop crucial skills if the AI handles too much of the task.

The study also examines long-term effects on skill distribution, finding that heterogeneity in "AI literacy" — the ability to identify and adapt to inaccurate AI outputs — leads to skill polarization over time. Those with higher AI literacy benefit disproportionately, widening the gap with less literate workers. The model provides simple conditions under which these paradoxes arise, helping explain the mixed empirical evidence on AI's real-world productivity impact. For businesses, this implies that simply adding more AI assistance without considering worker adaptation and literacy levels could backfire, potentially reducing overall productivity and exacerbating inequality.