Updating Acemoglu's AI Math: Economics in Real Time

February 14, 2025

This is an exciting time for economists interested in AI. Not to take away from Anton Korineks, Chad Joneses, and Daron Acemoglus who have been writing about the topic for years, but it feels like the field is newly sitting up and paying attention: now driven by industry.

Anthropic this week launched their Economic Index, a research agenda focused on the economic impacts of AI.OpenAI has also recently started an Economic Impacts Team, led by economist Dr Ronnie Chatterji from Duke. The first product is a paper and accompanying dataset describing how their AI tools are used. The form of this dataset is peculiarly well-fitted to existing work on the economic effects of AI. To illustrate how the field is changing, and will have to change, I’m going to use the Anthropic data to update a 2024 paper by Daron Acemoglu: “The Simple Macroeconomics of AI”. I hope this provides a sense of how economics can keep pace with AI’s rapid evolution - even as last year’s results becomes outdated.

Acemoglu’s “Simple Macro” Paper

Daron Acemoglu won the Nobel Prize last year for his work on institutions and growth.Looking at the slate of winners in other fields, one might think that he won it for his extensive work on automation and AI: the physics prize went to Hinton and Hopfield for founding the field of deep learning; the chemistry prize went to Demis Hassabis, the founder of Google DeepMind, alongside two other computer scientists. He has long been a giant in the field of economics, startlingly productive, and the Nobel is another star in a bedecked career.

The “Simple Macro” paper models how AI will affect the economy: which labor tasks will be automated, which tasks will be made more efficient, what new tasks will be invented, and how much does that matter for each task. Acemoglu defines four ways in which AI can increase productivity:This framework is from a 2019 paper of his with Pascual Restrepo.

1. Labor augmentation, or complementarities: This is how economists have traditionally viewed technological advances — increasing the productivity of tasks humans already do. A better shovel, search engine, or saxophone fit in this category.
2. Automation: Taking a task currently done by humans, and having an AI model do it instead. The Roomba automated a janitor. Personally, where I would have hired an RA to classify some data in 2019, I can now get GPT-4 to do it.
3. Deepening of automation: Some tasks, e.g., vacuuming, are already automated — AI might improve the productivity of the Roomba, or GPT-5 might be faster and cheaper at classifying my data than GPT-4.
4. New tasks: The invention of the computer invented the “task” of computer programming; the LLM has already invented the task of prompt engineering.

Acemoglu’s paper seeks to quantify the effects of improvements in AI within this framework. “Simple Macro” ignores the third category,Maxwell Tabarrok enthusiastically critiques this decision here. By ignoring deepening automation, Acemoglu treats the chatbot/language model as the key recent innovation of AI. This underrates the broad applicability of the transformer, which can also improve the productivity of robots, enable self-driving cars, and improve other already-automated tasks. I agree with Tabarrok that we’ll see large increases in capital productivity from recent AI innovations. and is mostly speculative on the fourth, focusing on negative welfare (not productivity) effects of new tasks. The most rigorous section of the paper is focused on the first two categories: how the recent wave of AI tools increase productivity by automating and augmenting tasks currently performed by people. He defines a “task” within the O*NET classification, a standardized system from the Department of Labor that comprehensively describes 19,000+ work activities across occupations.

The paper aims to calculate the change in productivity from new AI tools based on estimates in other papers in the literature. The calculation comes down to this: the change in productivity is equal to the percent of worker tasks affected by AI, times the wage-weighted importance of those tasks, times the cost-savings from using AI over the status quo in those tasks. This falls out of Hulten’s theorem, a general result on how microeconomic changes affect macroeconomic variables.Hulten 1978, modernized in Baqaee and Farhi 2019 Acemoglu finds other papers which provide each variable in his calculation:My notation, illustration only. This roughly corresponds to equation 14 in Acemoglu’s paper. TFP is Total Factor Productivity, a measure of how much output you get per input.

First, he uses Eloundou et al. (2023) to identify \(s_L\), the percent of labor exposed to automation. Eloundou et al. feed the 19,000+ worker tasks categorized by O*NET to GPT-4 and human raters, asking them to rate how “exposed” to automation each task is. Acemoglu aggregates these into occupations, and weights each by its wage — finding that 20% of labor income is exposed to automation.

Second, he uses Svanberg et al. (2024) to evaluate whether these exposed tasks can be profitably automated in the next ten years (before 2034): \(\alpha\). Svanberg et al. model the cost of humans and AI systems performing the same set of computer vision O*NET tasks, and find that 23% of vision tasks can be profitably automated — for the rest, the AI system would be too costly, or the current human behavior is too efficient.This figure assumes AI costs remain stable. If they decrease 10% per year, perhaps 30% may be efficiently automated. Acemoglu addresses this in a robustness section.

Third, Acemoglu uses the average of two studies on productivity improvements due to AI to estimate \(\pi\), the cost savings from automation. Noy and Zhang (2023) find that copy-writers work 40% faster with GPT-3.5 compared to those without; Brynjolfsson et al. (2023) find that customer support agents work 14% faster with an older version of GPT. Acemoglu uses the average of 27% cost-savings. Other studies have roughly agreed with this figure,Dell’Acqua et al. (2023) found BCG consultants completed their tasks 25% more quickly when using AI; Cui et al. (2024) found that developers using AI completed 26% more tasks. some estimate higher values.Toner-Rodgers (2024) found that material scientists using AI assistance discovered 44% more materials than those without the AI tool. There are other reasons to believe that the 27% number is too low when estimating over the next decade. Most notably, Merali (2024) looks at what scaling laws — that model performance improves with model size — imply about Acemoglu’s calculations. This is a good paper; accounting for conservative model growth, he more than doubles the productivity effect used by Acemoglu, from 27% (the average of Noy and Zhange and Brynjolfsson) to 61%.

Finally, Acemoglu uses 54% as the share of total income \(\phi\) that goes to paying wages, rather than other costs or profits.

Putting it all together,

This value is the percent increase in productivity from the new AI tools. Acemoglu’s headline result, based on these values, is that GDP in 2034 will be 1.16% higher due to LLMs than it would have been without.This comes from the 0.66% TFP level increase over 10 years and the estimated 0.93%–1.16% GDP increase, assuming a capital-output response proportional to TFP growth. If investment effects are stronger, the upper bound rises to 1.4%–1.56%.

New Data, More Growth

This is where Anthropic’s new data come in. Acemoglu’s estimates come from forward-looking estimates of tasks which might be automated or augmented; the Anthropic Economic Index provides data on which tasks are already being done by AI. Their first product is a datasetThe citation is Handa et al. (2025) and the authors are here. I’ll refer to the dataset authors as Anthropic. based on four million conversations with their premier LLM, Claude.

Using a system which hides the conversations from human reviewers, Anthropic maps each conversation to one of the tasks in the same O*NET categorization used by Eloundou et al. and Acemoglu. The conversations in Anthropic’s data are from users all over the world, and selects for more tech-savvy people who are comfortable using AI. Even among AI-heads, Claude is not as popular of an LLM as ChatGPT or Google’s Gemini. Claude has a reputation for being a niche tool more suited to developers than every-day use, although this is changing with the advent of reasoning models from other labs. In looking at actual-usage rather than potential-usage, I would expect to see technical tasks overrepresented in the Anthropic data compared to the Eloundou et al. exposure classification. The Anthropic team excludes tasks which have fewer than 15 conversations mapped to them, and report the percent of conversations which performed each task.

I use these data to repeat Acemoglu’s calculations.Thanks to his RA Can Yeşildere for clarity on methods. There are 3,514 tasks in Anthropic’s data which were automated or augmented, compared to Acemoglu’s 4,089. I map these tasks to occupations listed by the Bureau of Labor Statistics, and following Acemoglu, label an occupation x% exposed if x% of its tasks are automated or augmented in the Anthropic data. Finally, I aggregate total labor exposure to automation, weighted by the median wage for each occupation.I use 2019 wage data; Acemoglu uses wage data pooled across 2019-2022. My guess is that using pre-pandemic data probably suppresses my estimates relative to Acemoglu’s due to lower 2019 wage shares for AI-exposed jobs, which grew post-pandemic. This produces a wage-weighted share of tasks being automatable or augmentable by Anthropic’s AI models of 23.7%.Without access to the Eloundou et al. data on AI exposure, it’s unknown how much overlap there is between the 20.0% exposure Acemoglu found in their data, and the 23.7% from the Anthropic data. Because there are fewer tasks in the Anthropic data than exposed tasks from Eloundou et al., the tasks in the Anthropic data must be relatively higher paid to equal a larger share of labor income. I’ve reached out to the authors. I’ll update here if I can make that comparison.

There are two ways to use this number. If I interpret it as the share of labor which is exposed to automation in the future, it replaces \(s_L=0.200\), or 20.0% of labor tasks exposed in Acemoglu’s data. If I interpret it as the share of labor which is already profitably automatable or augmentable, then it replaces the \(s_L \times \alpha = 0.200*0.23 = 4.6\%\) calculation.

In favor of the first interpretation, any task in the Anthropic data is exposed to automation, but we have limited information about how attractive that task is for automation. Anthropic provides the percentage of conversations which perform a task, but not the total number of times that that task is performed in the economy. For example, we know that ~1% of conversations have Claude “prepare, rewrite and edit copy to improve readability, or supervise others who do this work,” but we don’t know from this data how many times humans do that task without AI automating or augmenting them. Therefore, the Anthropic data informs us about exposure, but not about attractiveness.

For the second interpretation, we could assume that any task in the Anthropic data is inherently attractive for automation. The data only include tasks which show up in at least 15 of the ~one million conversations analyzed, and even those tasks at the bottom of the distribution are attractive to automate or augment with AI. For example, the “design and fabricate dental prostheses, or supervise dental technicians and laboratory bench workers who construct the devices” makes up less than 0.002% of all conversations in the Anthropic sample. Therefore, at least 15 times, users turned to Claude to help with their dental prostheses.To be cost-effective means that it’s cost-effective for someone, not for everyone. That it hasn’t scaled — that a million dental protheses were designed and fabricated that week without Claude’s help — doesn’t mean that automating/augmenting the task is unattractive.

I am partial to the second interpretation. All tasks in the Anthropic data are both exposed and attractive to automation/augmentation. Keeping the cost savings and labor share stable, this changes the calculation:

an increase of 5x. This result is so much larger than Acemoglu’s because Anthropic’s data indicates that many more tasks are profitably automatable or augmentable than the predictions of Eloundou et al. and Svanberg et al. This corresponds to a roughly 10x increase (relative to Acemoglu) in GDP by 2034, perhaps around 11% larger GDP in 2034 than without the new AI technology. The larger increase in GDP than TFP is due to capital deepening — larger investment in capital when productivity rises.

Models Racing Models

Last summer, there was a cottage industry of people hating on the “Simple Macro” paper. Repeating or expanding on those critiques would be repetitive, as well as months late. And actually — I kinda like Acemoglu’s paper. For example, the Hulten’s theorem exercise is useful, and gives us a good basis for evaluating how automating and augmenting tasks using the current AI models will change work. And I’m especially glad for the “bad new tasks” section, which I haven’t covered here. Acemoglu calculates the negative welfare effects from AI-super-powered social media, based on estimates from other economists. Economics arrived late to this question, and I’m interested in work on AI following Hunt Allcott’s work on the welfare effects of social media. It will also be important to expand past social media, towards misleading advertising, romantic and social chatbots, deepfakes, and more.

The problem is that AI does not move at the speed of economics working papers. Anthropic has not released any “reasoning”Such as DeepSeek R1, or OpenAI’s o-series. models, which are much closer to automation than the chatbot LLMs assessed here. Both Acemoglu’s original paper and the Anthropic data focus only on chatbots (ignoring any API use), which is at least one step removed from true automation. The underlying technology continues to get better, and firms are learning how to use it more effectively. Acemoglu’s results rely on Eloundou et al. and Svanberg et al., and taken with his model, these numbers undervalue the potential of current AI models. Perhaps they were accurate at the time — but the Anthropic data suggests that an AI can already cost-effectively perform a much higher fraction of tasks.

Moving Faster: Economics in Real Time

Increasing rates of technological progress requires increasing the rate of research on that progress. Anthropic admirably published these data (and their paper) just 23 days after data collection ended.

It’s unreasonable to expect an economist, even the legendarily productive Acemoglu, to produce a novel structural model and write it up in the span of days and weeks. But having frameworks which will work with data known to be forthcoming increases the cycle time, and allows us as researchers to keep up with fast-paced developments. It may be a coincidence that the Anthropic data line up so nicely with Acemoglu’s 2024 paper.The Anthropic paper does not cite Acemoglu’s Simple Macro paper. That I can present this replication immediately is a testament to how well Acemoglu’s model is tailored to the data. Or perhaps the other way around: automation is Acemoglu’s wheelhouse, and it’s not surprising that his task-based framework emerges as the primary framework for academics to evaluate these technologies.

I follow Acemoglu’s approach very closely above; if he had had these data, it’s likely he would have taken a different approach. For example, I use the Anthropic data to indicate whether a task is automatable at all, and follow the original paper in measuring intensity at the occupation level. But the new data allow researchers to measure the relative popularity of automation between tasks. A more tailored approach might weight occupations by how deeply the technology is being adopted, or assess the cost savings of automating specific tasks rather than occupations.

Clearly, AI models have become more capable between when Eloundou et al. measured AI exposure and when Anthropic measured their users’ tasks. Economists will continue to write papers about the AI exposure of different tasks, and the number of exposed tasks will increase over time — the models are never going to be worse than they are today. Anthropic has been good at reaching out to researchers; we should reciprocate! If we want to stay up to date on the economic effects of these models, we should do the legwork that we can before the data are released. I was only able to come up with the figure of “5% higher GDP in 2034” with Anthropic’s data because Acemoglu had done the hard work of setting up the structural model — ready to plug in data — last year.

Future Directions

There are other ways to use these data; Anthropic only released the dataset this week, and I expect more to come. Now that there are teams at Anthropic and OpenAI dedicated to studying the economic effects of AI, there is so much more I’m interested in seeing. I’ll close with some closely related examples.

The Anthropic paper provides high level data on which tasks are automotive vs. augmentative. While I reference this above, they do not include these crosstabs on the main tasks in the public dataset. I could measure exposure to automation or augmentation with more detailed data.

Similarly, providing similar data from API use, not just online chatbots, will provide further detail into automotive behavior. I use claude.ai and chatgpt.com for augmenting my work — but when I want an AI to do batch work for me, I use the APIs.

The Anthropic data also covers conversations with two of their models, Opus and Sonnet, although the task data used above is based on a third model, Haiku. Model choice is super interesting and largely vibes-based today. Similarly, OpenAI does A/B testing with their models — I so want those data.

Acemoglu has a digression in his paper on easy vs. hard tasks: his team used a simpler ML system to assign a probability that each of the O*NET tasks is hard. This classification finds that exposed tasks according to Eloundou et al. tend to be easier.

The O*NET system also provides complexity and importance scores for each task. Neither Acemoglu or I use these. Complexity could be used in place of his easy/hard distinction. Importance could be used to weight tasks within occupation. Currently, all tasks are weighted the same.

You too can have input on what data Anthropic release next: they’re requested feedback from researchers here.