The heterogeneous productivity effects of generative AI
Evidence from Italy’s ChatGPT ban
Summary
This project examines the effect of generative AI on the productivity of software developers.
Key researchers
- David Kreitmeir, SoDa Labs and Department of Economics
- Paul Raschky, SoDa Labs and Department of Economics
Project description
The public release of OpenAI’s ChatGPT provided near universal access to generative artificial intelligence (AI) tools at very low cost, with the potential to boost worker productivity by performing relatively complex tasks and producing novel output, all while requiring only minimal technological knowledge on the part of users. However, ChatGPT also has the tendency to produce wrong or faulty outputs that, in the absence of expert knowledge, are difficult to detect and costly to rectify.
We exploit the sudden ban of ChatGPT as a natural experiment to test the potentially heterogeneous effect of ChatGPT on individual productivity. Using detailed information on the the daily output quantity and quality of 36,000 GitHub users from Italy and other European countries, we show that both the output quantity and quality of less experienced Italian users increases in the days after the ban, while the ban has no systematic effect on the overall output of more experienced developers except for some small effects on more routine tasks such as resolving issues or debugging.