Scaling Laws for Economic Impacts

How to cite

merali, Ali. 2024. "Scaling Laws for Economic Impacts." AEA RCT Registry. AEARCTR-0013743. https://doi.org/10.1257/rct.13743-2.0

Download: BibTeX · RIS — for reference managers without a browser translator. Zotero and most others pick up this page's citation metadata directly.

Ali merali*

* Primary investigator

Registry ID
AEARCTR-0013743
DOI
https://doi.org/10.1257/rct.13743-2.0
First registered
Last updated
Status
In development
Trial dates
2024-07-22 – 2024-08-09
Countries
Private
Registry record
http://www.socialscienceregistry.org/trials/13743

Abstract

This study investigates the productivity impacts of thirteen generative artificial intelligence (AI) technologies with varying training compute sizes. By analyzing scaling laws, we measure how economic impacts evolve as the size of generative AI models increases. Our research is conducted in a real-world context, employing professional online translators to perform tasks akin to their routine professional work. Additionally, we explore the relationship between the skill-bias of technological change and the degree of automation. Our findings offer critical insights into the scaling effects of AI on productivity and the dynamics of skill-biased technological advancements in the labor market.

Experimental design

Recruiting 336 professional translators on online freelancer websites (eg. Freelancer.com, fiverr.com) across three languages (Spanish, Arabic, and Hindi). Each translator will first complete a baseline task of language ability. Then they will complete five short follow-up tasks. In each follow-up task individuals will equally be split into one of fourteen groups. One with no AI assistance (control), and the others will be randomly assigned one of thirteen different generative AI models. They will also all complete some short questions (eg. on their experience as a professional translator, familiarity with AI tools, etc.)

Design details

Randomization method
Randomization done in office by a computer
Randomization unit
Individual
Clusters
336 (112 individuals for each of the three languages)
Observations
336 individuals
Arms
120 tasks completed * 14 arms = 1680 tasks Total= 1680 tasks/ 5 tasks per individual= 336 individuals
Primary outcomes
See Analysis Plan for full regression specifications/ definitions: 1. Productivity by log model compute (Main) 2. Interaction between respondent baseline ability and log model compute (Main 2) 3. Heterogeneity of productivity impacts by task complexity (interaction between log model compute and language in a logistic regression)

Cited by

No citation data available for this registration yet.

Keywords