In a controlled experiment published in Science, Noy & Zhang (2023) assigned 453 college-educated professionals to writing tasks with and without ChatGPT access:
- ChatGPT reduced average completion time by 40% (from 27 to 17 minutes)
- Output quality improved by 18% as rated by blinded evaluators
- The tool particularly benefited lower-performing writers, compressing the quality distribution
- However, workers with ChatGPT access showed reduced effort and more homogeneous outputs
The Inequality Compression Effect
Both this study and Brynjolfsson et al. (2023) found the same pattern: AI most benefits lower-performing individuals, compressing the performance distribution. This is "sweet" for equality but raises a "bitter" question: if everyone's output converges to the same AI-enhanced level, what happens to individual expertise and career differentiation?
Source
Noy, S., & Zhang, W. (2023). Science, 381(6654), 187-192.