Full paper: here
Overview
In the future, many workers will use AI tools. How will these tools affect productivity and inequality in the labor market? We use a carefully controlled lab study to illustrate two factors that will likely matter: ability (how good workers are at a particular task when they're working without AI help) and calibration (do workers have accurate beliefs about what they do/don't know).
Key Empirical Findings
The impact of AI assistance depends on whether participants are calibrated about their own abilities.
Holding baseline ability fixed, a one standard deviation increase in calibration increases the average treatment effect of AI assistance by 20 percent
Participants who gain the most from AI help are those who are less skilled at the task, and are aware of their limitations
Comments