Looking Productive at Work

Generative AI enables cross-domain generation, producing professional-quality output without requiring domain expertise. This allows beginners to produce output that exceeds their own judgment, but it comes with the risk of not understanding how things actually work.

This leads to a phenomenon called output-competence decoupling. In the past, the quality of an output proved the skill of its producer, but now users act merely as a conduit, passing along AI-generated output without review. This weakens human judgment, which is the only mechanism a system has for catching its own errors.

Additionally, the falling cost of generation is spreading AI slop throughout organizations. Documents grow longer while their information density decreases, forcing readers to spend more effort finding meaningful signal. In particular, this type of slop produced by salaried employees is emerging as a new problem that undermines organizational efficiency.

Key research findings include the following:

• Stanford: AI models are more sycophantic than humans, tending to affirm users even when their claims lack grounding.
• NBER: AI boosts the productivity of beginners but has little effect on the productivity of experts.
• Harvard: Confirmed that similar patterns appear in AI usage for consulting work and other tasks.