Preprint
Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence (working paper)
Shakked Noy and Whitney Zhang, MIT Economics working paper, 2023
Read the originalRead on October 16, 2026
- What it is
- A randomized experiment with college-educated professionals who wrote short workplace texts, with and without ChatGPT. Writing time and blinded evaluator grades are the two outcomes.
- How much weight it carries
- A working paper, not the final journal version. The authors note they did not check the accuracy of the generated writing. Check the published Science version before quoting its figures.
Facts taken from this source
- In the Noy and Zhang working paper, average evaluator grades for the treatment group rose by 0.45 standard deviations.
“Average evaluator grades in the treatment group increase by 0.45 standard deviations” (Results)
- In the Noy and Zhang working paper, time taken on the second writing task fell by 10 minutes, or 37 percent, for the group with ChatGPT.
“Time taken on the post-treatment task drops by 10 minutes (37%) relative to the control group” (Results)