Empirical Economics of AI
This is a reading group designed to offer an accessible introduction on how economists think
about the potential economic impacts of AI. We are going to engage with early evidence on
how people actually use frontier LLM’s (both for personal and for professional use). The
series will close with some readings on early evidence of AI raising productivity for specific
types of workers.
We will convene every 2 weeks, for 60-90 minutes each time. Possible questions to be
discussed will be distributed a couple of days before each session. The required reading
amounts to about ~ 3 hours for each session.
Your facilitator
Demosthens Kollias
Fellow, Economic of AI, Institute for Labour
Board Member, DIKTIO - Network for Reform in Greece and Europe
Topic 1 | The foundations | Exposure Metrics
The fundamental causal mechanism through which AI will affect the economy is via the
labour market. In order to understand how this may possibly play out, we need to first study
models that aim to capture how work might be affected. We start by the seminal work by
Osborne and Frey, which in 2013 raised the alarm about the potential of advanced AI being
able to automate routine and non-routine tasks. We then proceeded to discuss two more
recent papers, one from OpenAI and another one from the International Labour Organisation.
OpenAI (Eloundou et al, 2023) uses GPT 4 to classify occupations in terms of their potential
to get substituted by Large Language models. ILO (Gmyrek et al, 2025) uses a balanced and
multilayered approach that asks frontier LLM’s to give numerical estimates on the extent to
which specific work tasks could be performed by AI at an adequate level, before asking
similar questions to people working on relevant tasks and labour market experts, therefore
providing a multi-layered and balanced approach that is worth discussing in more depth.
-The Future of Employment: How susceptible are jobs to computerisation? (https://oms-
www.files.svdcdn.com/production/downloads/academic/future-of-employment.pdf )
-GPTs are GPTs: Labor market impact potential of LLMs
(https://www.science.org/doi/10.1126/science.adj0998 )
-Generative AI and Jobs: A Refined Global Index of Occupational Exposure
(https://www.ilo.org/publications/generative-ai-and-jobs-refined-global-index-occupational-
exposure )
Topic 2 | Evidence and Metrics from AI labs
The first 2 papers of Week 2 explore research work prepared by OpenAI and Anthropic, that
tried to categorise user queries into broad categories. Notably, the Anthropic Economic Index
attempted to map Claude conversations to US occupations and tasks, and also classify them
as augmentative (meaning users actively collaborate, iterate, or learn with Claude), or
automative (users delegate a task to the AI to fulfill with minimal human involvement).
Lastly, the GDPval paper by OpenAI asked professionals to complete common tasks they
face in their work (including multimodal ones, where one handles pdfs, images, etc). Frontier
LLM’s were also asked to perform the exact same work. Experienced professionals were then
asked to rank the outputs in a blinded evaluation and explain their reasoning. For the initial
analysis, frontier LLM’s were almost on par with seasoned industry professionals (Summer
2025), but more recent models (like GPT 5.4) tend to overperform them.
-Which Economic Tasks are Performed with AI? Evidence from Millions of Claude
Conversations (https://www.anthropic.com/news/the-anthropic-economic-index )
-How People Use ChatGPT
(https://www.nber.org/system/files/working_papers/w34255/w34255.pdf )
-GDPval: Evaluating AI Model Performance on Real-World Economically Valuable Tasks
(https://arxiv.org/abs/2510.04374 )
Topic 3 | Views on Productivity
This week we are going to examine 2 micro studies on whether AI boosts productivity in 2
specific domains; customer support and book production. In Generative AI at Work,
Brynjolfsson, Li, and Raymond study the rollout of an AI assistant to more than 5,000
customer-support agents and find a roughly 15% productivity gain on average, with the
largest benefits accruing to less experienced and lower-skilled workers, By contrast, Reimers
and Waldfogel show that the diffusion of LLMs has been associated with a huge expansion in
output of new books, but average quality has fallen, implying a world with much more
“slop”.
-Generative AI at Work (https://academic.oup.com/qje/article/140/2/889/7990658 )
-AI and the Quantity and Quality of Creative Products: Have LLMs Boosted Creation of
Valuable Books?
(https://www.nber.org/system/files/working_papers/w34777/w34777.pdf )
Very cool podcast discussion of the paper, here: https://empiricrafting.substack.com/p/is-ai-
making-books-on-amazon-worse
Additionally, you can find here an excellent summary of empirical studies on micro- and
macro- studies (regularly update), curated by Alex Imas, here:
https://aleximas.substack.com/p/what-is-the-impact-of-ai-on-productivity
Topic 4 | Early data on Labour Market Implications | A terrible time for recent university graduates?
Given that there have been 3+ years since the commercial release of ChatGPT, if those AI
exposure metrics bear some degree of truth, shouldn’t we be able to observe some kind of
labour market effects? This is the hypothesis this week’s papers try to tackle. Both studies
conclude that employment in high-exposure occupations has slowed down for young workers
in both the US and Sweden, and this is especially true for occupations where AI usage is
classified as automotive, rather than augmentative. Overall trends for more senior workers do
not point in the same direction. Put together, the papers suggest that the early labour-market
effect of generative AI may not yet be a dramatic collapse in total demand, but a quieter
reordering of who gets hired, trained, and given a foothold at the bottom of the ladder.
-Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial
Intelligence
(https://digitaleconomy.stanford.edu/app/uploads/2025/12/CanariesintheCoalMine_Nov25.pd
f )
-Same Storm, Different Boats: Generative AI and the Age Gradient in Hiring (Sweden)
(https://cms.ratio.se/app/uploads/2026/03/lodefalk_march16_paper_v2-kombinerades.pdf)
——————————————-
Additional Resources that could be of interest
-Situational Awareness, by Leopold Aschenbrenner (https://situational-awareness.ai/ )
-Machines of Loving Grace, by Dario Amodei (https://darioamodei.com/essay/machines-of-
loving-grace )
-The Intelligence Age, by Sam Altman (https://ia.samaltman.com/ )
-Economics and Transformative AI, by Tom Cunningham
(https://tecunningham.github.io/posts/2025-09-19-transformative-AI-notes.html )
-Justified Posteriors (blog and podcast), by Andrey Fradkin and Seth Benzell
(https://empiricrafting.substack.com/podcast )
-AK or just okay? AI and economic growth, by J. Zachary Mazlish
(https://jzmazlish.substack.com/p/ak-or-just-okay-ai-and-economic-growth)
-Can AI companies become profitable? Lessons from GPT-5’s economics, by aime Sevilla,
Hannah Petrovic, and Anson Ho of Epoch AI (https://epoch.ai/gradient-updates/can-ai-
companies-become-profitable)
-The Economics of Transformative AI | Summer Course that ran at Stanford in the past
summer, consisting of lectures by Phil Trammell and J. Zachary Mazlish. Lecture slides,
recordings, and supplementary material can be found here,
https://digitaleconomy.stanford.edu/about/education/the-economics-of-transformative-ai/ .
-Joshua Gans’ Newsletter, https://joshuagans.substack.com/