What this really highlights is that productivity doesn’t scale at the task level, it scales at the institutional level.
We’re seeing genuine task acceleration, but jobs are embedded in workflows, permissions, review cycles, liability, and coordination structures that haven’t moved. Until those layers change, micro gains get absorbed as slack, quality variance, or reallocation rather than output.
The Solow paradox wasn’t about computers being weak. It was about institutions being slow.
Thanks, Alex. I really enjoyed reading this. With another group of coauthors, we tried to bridge this macro-micro divide by looking at open source software but when a country banned it (https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5332003).
Maybe this is worthy of a longer conversation, but I wonder if simple measures of adoption are not very informative. The best way I've heard it said is by Arvind Narayanan and Sayesh Kapoor:
Imagine an alternate universe in which people don’t have words for different forms of transportation, only the collective noun “vehicle.” They use that word to refer to cars, buses, bikes, spacecraft, and all other ways of getting from place A to place B. Conversations in this world are confusing. There are furious debates about whether or not vehicles are “environmentally friendly,” but (even though no one realizes it) one side of the debate is talking about bikes and the other side about trucks. There is a breakthrough in rocketry, but when the media focuses on how vehicles have gotten faster, people call their car dealer (oops, vehicle dealer) to ask when faster models will be available. Meanwhile, fraudsters have capitalized on the fact that consumers don’t know what to believe when it comes to vehicle technology, so scams are rampant in the vehicles sector.
Now replace the word “vehicles” with “artificial intelligence,” and we have a pretty good description of the world we live in.
If this is the case, then are we measuring adoption in a way that is helpful? Even in Shen and Tamkin (2025), the way the tool is used is so different across participants, is it even the same tool? I am asking these questions because I don't know the answers. And I am simultaneously trying to make progress in my other work on algorithmic hiring, where I think the challenges are similar. A resume screening tool is different from a voice interviewer is different from having candidates play games that are trained on performance of your current employees. But we are calling these 'algorithmic hiring' and trying to draw broad conclusions over inherently different technologies.
One thing worth questioning though: the assumption that the micro gains will eventually show up in the macro.
Not all technologies do. In fact, most don’t. The nail gun made framing 60% faster and never moved TFP. There’s a tradition of declaring GPT only to find out that what we have is just a t.
The current AI evidence—narrow task gains, no rush to deploy despite problems, wage premium going to users not builders—looks more like a nail gun than like electrification.
The micro-macro gap is the key here. Productivity runs through coordination systems, not tasks. A faster task inside an unchanged workflow just creates slack or gets absorbed. The bottleneck is the review cycle, the approval chain, the meeting cadence, the organizational defaults that were designed around the old speed. Until the coordination architecture changes, task-level gains don’t compound. They dissipate. Less a J-curve delay than a design problem.
One terawatt of AI compute using the latest models, Claude Opus and Gemini, about doubles the global cognitive-work workforce (guesstimated at 500M humans, assuming 24/7 work by the AI instances). Cognitive work being administration, management, planning, marketing, sales, government, teaching, diagnosis, interpretation, reporting, and so on. Any job mostly done in an office building, not directly moving physical things.
Jobs are collections of tasks, but over time those collections can be, and often are, reorganised: split up and recombined as new jobs, simpler or more complex depending on individual context. Mostly simpler, though. Some reorganisations greatly reduce the impact of formerly bottlenecking tasks.
Beware any paper that falls for the "static world" fallacy, implicitly assuming that jobs are fixed in amber for all time.
Awesome, thanks for putting this all together. As a former economist who now works as a software engineer in an ML field, this topic is of great interest.
I am overall skeptical of the micro evidence. A couple caveats in addition to those you mentioned:
1. Which tasks are completed? The set of tasks to be completed doesn’t fall from the sky; at least in software engineering for people past entry level, it’s the worker’s responsibility to figure out how best to break a larger goal into tasks. Breaking the goal into tasks isn’t done entirely up front, but rather is done iteratively along with implementation. One can ask an agent to do the wrong thing and get a “successful” PR out of it. E.g. Say I'm tasked with adding feature Y. I determine that adding feature X is a prerequisite. I have an agent write a PR for feature X, and I land it. If I had written the code by hand, I might have noticed that we don’t actually need feature X. Or maybe I do notice while working with the agent, but I decide I might as well have the agent finish that PR anyway.
A study such as Cui et. al. that counts pull requests or lines of code will not capture the issue of task selection. Nor is it fully captured by a study like Paradis et. al. that uses relatively small tasks (0.5 to 2.5 hours). Becker et. al. used tasks predicted to take less than 2 hours. For the Peng et. al. Copilot study, you can read the task specification on page 14. The task took the control group ~2.5 hours, but is specced out on a very detailed level, making it more of a collection of smaller tasks.
2. Quality. I see that some of the micro studies mentioned here show negative quality effects of AI for code, and a Google search for “AI code quality” will reveal a few studies and a lot of complaints; I want to emphasize this point, because I’m not sure it's obvious to people who haven’t worked on large-scale software just how much it slows you down when a component superficially seems to solve the problem but has a subtle bug and/or incorrect documentation. Working on small prototypes and personal projects won’t give you the right intuition. For the economists, think O-ring theory (and it's hard to find the O-ring, and it has a sticker on it that says “this is the engine and it definitely works"). A common maxim is that you need to be twice as smart to debug code as to write it.
3. Loss of human knowledge of the codebase. This is an unmeasured externality in that if one task is completed with AI assistance, future tasks become harder to complete. This is an issue with experiments that randomize at the task level or even the worker level. An experiment that captures this would need to give the same treatment to everyone touching the same code. (Of course, you could posit a similar effect in the opposite direction, where workers learn valuable skills from seeing how an AI solves a problem -- I have picked up some tricks that way.)
I think these issues are all pretty important, and jointly they make me very skeptical of the micro evidence and concerned that AI could have a fairly negative effect on team productivity (especially when measured in real economic terms, not just # of PRs or LOC) on longer time horizons.
Great write-up and interesting to see the disconnect.
One additional thought to the ones you and commenters mentioned on why macro impacts might be lower:
A project might have a set amount of time to be completed. For example, suppose you need to provide analyses in 10 hours. Without AI, you'd prioritize on what is the best bang for your buck - so you complete 2 analyses for the same project. With AI, you can now complete 3 analyses for the same project. The likelihood, however, that the 3rd analysis is a true value add might be low. So although we can do more, the incremental value add of the additional work is low.
When Microsoft Windows was newly introduced in the 1990s, a lot of time was wasted in "faffing around" - changing the desktop background, adjusting colors and the shape of the cursor etc. I guess it used to give people a sense of being productive without actually doing anything useful. I see much the same going on right now with AI at the individual level.
I also have a few thoughts on why productivity gains from AI aren't showing up at the macro level (some overlap a little with what you've already said).
1. It's very difficult to disentangle the impact of AI from other macroeconomic factors. Several studies are using dif-in-dif to attempt this (with some finding evidence that AI is already having negative effects on employment.)
2. People are still figuring out how to make the most of AI - that figuring out phase might drag productivity, rather than enhancing it. Workers may not realise that AI is slowing them too (as the Becker et al. study described above suggests).
3. If AI is increasing workers' productivity - I suspect a lot of people will be taking that time back as more leisure/less effort. That promise surely must be drawing a lot of workers to use AI. If that's the case, I suppose the question is "how long until the managers cotton on and start demanding more from staff?"
The bottleneck framing in the "jobs are a collection of tasks" section is one of the clearest explanations I have encountered of why micro gains fail to compound into macro outcomes. The Faros finding about PR review time increasing 91% despite 98% more pull requests is almost a perfect real-world confirmation of exactly this mechanism. The code arrives faster; the human approval layer does not.
What strikes me about the J-curve argument is that it predicts not just delay but a non-linear jump, once organizations cross some threshold of genuine restructuring. That threshold seems to be the part most firms are underestimating right now. I just published a piece tracing the parallel to the 1980s Solow paradox and what the current macro data actually implies about timing, if any of it is useful to you.
Thanks for putting this all together. Great resource!
Generally I'm watching this trend of: Task productivity increases > Firm level productivity > Job posting slowing + Macro productivity gains. Not sure where job losses fit in there yet, but sense it's very lumpy and coming soon!
Trying to synthesize all of the noisy data and predications and corporate announcements here: https://labor.mattzieger.com/
Thanks for review. I avoided temptation to get chatgpt to summarise it and find most useful things. Mixed feelings about AI. Somethings it is good. But, causing a lot of disruption too.
Great roundup, thanks for putting this together. Productivity from AI is a topic I really find fascinating, since I can very clearly quantify the hours that it saves me every month, but the data are so much less clear.
The productivity impact is most dramatic in knowledge work with high repetition. Accounting is the perfect example: same email types daily, same document reviews weekly, same reports monthly. AI doesn't just save time on each instance — it eliminates the cognitive load of starting from scratch every time. That's the hidden productivity gain nobody measures.
Publicly listed companies tend to be reporting their quarterly financials several days earlier than even a few years ago. AI has been a significant factor in the speed up. This is a quality improvement that isn't necessarily going to show up in productivity statistics.
The J-curve framing creates a measurement problem that the analysis doesn't address. Complementary investments are, by definition, a leading indicator; firm MFP (aggregated to TFP) is a lagging one. Including both in a single 'AI impact' figure means the J-curve isn't showing AI impact rising over time—it's showing the sum of (leading investments) + (lagging productivity) rising. Those two quantities have different dynamics and shouldn't be added without justification.
The shape is then treated as evidence for the mechanism. But the shape is equally consistent with the mechanism being absent or broken. Penn Wharton projects AI's contribution to TFP growth at roughly 0.01 percentage points in 2025: essentially zero. The late-2025 'uptick' looks more capex-driven (data centres etc.) than widespread evidence of firms extracting value from AI intangible. If complementary investments are climbing while TFP contribution is near zero, that's a widening gap between leading and lagging indicators, not a J-curve descent. Calling it the latter is a reading the data underdetermines.
The J-curve requires not just that intangible capital accumulates, but that firms eventually extract it as output. There's currently little empirical signal that either condition (accumulation, extraction) is being met. Babina et al. find zero relationship with sales per worker or TFP—though the study predates genAI and the replication question remains open—while Humlum and Vestergaard find zero effect on earnings or hours in Danish administrative data through end of 2024, even for workers who self-reported large productivity gains. We also need to consider that 'leading indicators' might be consumption rather than investment. If a firm buys AI because 'every other firm is buying AI' (a Red Queen effect), it's a defensive expense (consumption of capital) rather than a complementary investment.
The parsimonious reading of the current data is that the gap between leading and lagging indicators is widening. That's consistent with the J-curve—but it's equally consistent with the leading indicators leading somewhere else.
What this really highlights is that productivity doesn’t scale at the task level, it scales at the institutional level.
We’re seeing genuine task acceleration, but jobs are embedded in workflows, permissions, review cycles, liability, and coordination structures that haven’t moved. Until those layers change, micro gains get absorbed as slack, quality variance, or reallocation rather than output.
The Solow paradox wasn’t about computers being weak. It was about institutions being slow.
Yes agreed completely.
Thanks, Alex. I really enjoyed reading this. With another group of coauthors, we tried to bridge this macro-micro divide by looking at open source software but when a country banned it (https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5332003).
Maybe this is worthy of a longer conversation, but I wonder if simple measures of adoption are not very informative. The best way I've heard it said is by Arvind Narayanan and Sayesh Kapoor:
Imagine an alternate universe in which people don’t have words for different forms of transportation, only the collective noun “vehicle.” They use that word to refer to cars, buses, bikes, spacecraft, and all other ways of getting from place A to place B. Conversations in this world are confusing. There are furious debates about whether or not vehicles are “environmentally friendly,” but (even though no one realizes it) one side of the debate is talking about bikes and the other side about trucks. There is a breakthrough in rocketry, but when the media focuses on how vehicles have gotten faster, people call their car dealer (oops, vehicle dealer) to ask when faster models will be available. Meanwhile, fraudsters have capitalized on the fact that consumers don’t know what to believe when it comes to vehicle technology, so scams are rampant in the vehicles sector.
Now replace the word “vehicles” with “artificial intelligence,” and we have a pretty good description of the world we live in.
If this is the case, then are we measuring adoption in a way that is helpful? Even in Shen and Tamkin (2025), the way the tool is used is so different across participants, is it even the same tool? I am asking these questions because I don't know the answers. And I am simultaneously trying to make progress in my other work on algorithmic hiring, where I think the challenges are similar. A resume screening tool is different from a voice interviewer is different from having candidates play games that are trained on performance of your current employees. But we are calling these 'algorithmic hiring' and trying to draw broad conclusions over inherently different technologies.
Great survey of the evidence.
One thing worth questioning though: the assumption that the micro gains will eventually show up in the macro.
Not all technologies do. In fact, most don’t. The nail gun made framing 60% faster and never moved TFP. There’s a tradition of declaring GPT only to find out that what we have is just a t.
The current AI evidence—narrow task gains, no rush to deploy despite problems, wage premium going to users not builders—looks more like a nail gun than like electrification.
The lag might not be a lag at all.
Yeah macro productivity is super noisy so you might be right.
The micro-macro gap is the key here. Productivity runs through coordination systems, not tasks. A faster task inside an unchanged workflow just creates slack or gets absorbed. The bottleneck is the review cycle, the approval chain, the meeting cadence, the organizational defaults that were designed around the old speed. Until the coordination architecture changes, task-level gains don’t compound. They dissipate. Less a J-curve delay than a design problem.
One terawatt of AI compute using the latest models, Claude Opus and Gemini, about doubles the global cognitive-work workforce (guesstimated at 500M humans, assuming 24/7 work by the AI instances). Cognitive work being administration, management, planning, marketing, sales, government, teaching, diagnosis, interpretation, reporting, and so on. Any job mostly done in an office building, not directly moving physical things.
Jobs are collections of tasks, but over time those collections can be, and often are, reorganised: split up and recombined as new jobs, simpler or more complex depending on individual context. Mostly simpler, though. Some reorganisations greatly reduce the impact of formerly bottlenecking tasks.
Beware any paper that falls for the "static world" fallacy, implicitly assuming that jobs are fixed in amber for all time.
Awesome, thanks for putting this all together. As a former economist who now works as a software engineer in an ML field, this topic is of great interest.
I am overall skeptical of the micro evidence. A couple caveats in addition to those you mentioned:
1. Which tasks are completed? The set of tasks to be completed doesn’t fall from the sky; at least in software engineering for people past entry level, it’s the worker’s responsibility to figure out how best to break a larger goal into tasks. Breaking the goal into tasks isn’t done entirely up front, but rather is done iteratively along with implementation. One can ask an agent to do the wrong thing and get a “successful” PR out of it. E.g. Say I'm tasked with adding feature Y. I determine that adding feature X is a prerequisite. I have an agent write a PR for feature X, and I land it. If I had written the code by hand, I might have noticed that we don’t actually need feature X. Or maybe I do notice while working with the agent, but I decide I might as well have the agent finish that PR anyway.
A study such as Cui et. al. that counts pull requests or lines of code will not capture the issue of task selection. Nor is it fully captured by a study like Paradis et. al. that uses relatively small tasks (0.5 to 2.5 hours). Becker et. al. used tasks predicted to take less than 2 hours. For the Peng et. al. Copilot study, you can read the task specification on page 14. The task took the control group ~2.5 hours, but is specced out on a very detailed level, making it more of a collection of smaller tasks.
2. Quality. I see that some of the micro studies mentioned here show negative quality effects of AI for code, and a Google search for “AI code quality” will reveal a few studies and a lot of complaints; I want to emphasize this point, because I’m not sure it's obvious to people who haven’t worked on large-scale software just how much it slows you down when a component superficially seems to solve the problem but has a subtle bug and/or incorrect documentation. Working on small prototypes and personal projects won’t give you the right intuition. For the economists, think O-ring theory (and it's hard to find the O-ring, and it has a sticker on it that says “this is the engine and it definitely works"). A common maxim is that you need to be twice as smart to debug code as to write it.
3. Loss of human knowledge of the codebase. This is an unmeasured externality in that if one task is completed with AI assistance, future tasks become harder to complete. This is an issue with experiments that randomize at the task level or even the worker level. An experiment that captures this would need to give the same treatment to everyone touching the same code. (Of course, you could posit a similar effect in the opposite direction, where workers learn valuable skills from seeing how an AI solves a problem -- I have picked up some tricks that way.)
I think these issues are all pretty important, and jointly they make me very skeptical of the micro evidence and concerned that AI could have a fairly negative effect on team productivity (especially when measured in real economic terms, not just # of PRs or LOC) on longer time horizons.
Great write-up and interesting to see the disconnect.
One additional thought to the ones you and commenters mentioned on why macro impacts might be lower:
A project might have a set amount of time to be completed. For example, suppose you need to provide analyses in 10 hours. Without AI, you'd prioritize on what is the best bang for your buck - so you complete 2 analyses for the same project. With AI, you can now complete 3 analyses for the same project. The likelihood, however, that the 3rd analysis is a true value add might be low. So although we can do more, the incremental value add of the additional work is low.
Yup, I discuss something similar as one of the factors blunting macro impact: AI being applied to unproductive tasks e.g,. the 7th analysis.
Thanks for the kind words!
When Microsoft Windows was newly introduced in the 1990s, a lot of time was wasted in "faffing around" - changing the desktop background, adjusting colors and the shape of the cursor etc. I guess it used to give people a sense of being productive without actually doing anything useful. I see much the same going on right now with AI at the individual level.
This is a really excellent summary.
I also have a few thoughts on why productivity gains from AI aren't showing up at the macro level (some overlap a little with what you've already said).
1. It's very difficult to disentangle the impact of AI from other macroeconomic factors. Several studies are using dif-in-dif to attempt this (with some finding evidence that AI is already having negative effects on employment.)
2. People are still figuring out how to make the most of AI - that figuring out phase might drag productivity, rather than enhancing it. Workers may not realise that AI is slowing them too (as the Becker et al. study described above suggests).
3. If AI is increasing workers' productivity - I suspect a lot of people will be taking that time back as more leisure/less effort. That promise surely must be drawing a lot of workers to use AI. If that's the case, I suppose the question is "how long until the managers cotton on and start demanding more from staff?"
I totally agree on all of this.
The bottleneck framing in the "jobs are a collection of tasks" section is one of the clearest explanations I have encountered of why micro gains fail to compound into macro outcomes. The Faros finding about PR review time increasing 91% despite 98% more pull requests is almost a perfect real-world confirmation of exactly this mechanism. The code arrives faster; the human approval layer does not.
What strikes me about the J-curve argument is that it predicts not just delay but a non-linear jump, once organizations cross some threshold of genuine restructuring. That threshold seems to be the part most firms are underestimating right now. I just published a piece tracing the parallel to the 1980s Solow paradox and what the current macro data actually implies about timing, if any of it is useful to you.
Thanks for putting this all together. Great resource!
Generally I'm watching this trend of: Task productivity increases > Firm level productivity > Job posting slowing + Macro productivity gains. Not sure where job losses fit in there yet, but sense it's very lumpy and coming soon!
Trying to synthesize all of the noisy data and predications and corporate announcements here: https://labor.mattzieger.com/
Wow this is a pretty incredible resource!!! I’m going to track this too.
Thanks for review. I avoided temptation to get chatgpt to summarise it and find most useful things. Mixed feelings about AI. Somethings it is good. But, causing a lot of disruption too.
https://open.substack.com/pub/glenroberts911399/p/ai-and-my-reflection-the-architecture?r=1pgr4n&utm_medium=ios&shareImageVariant=overlay
This is fantastic, thanks!
Great roundup, thanks for putting this together. Productivity from AI is a topic I really find fascinating, since I can very clearly quantify the hours that it saves me every month, but the data are so much less clear.
The productivity impact is most dramatic in knowledge work with high repetition. Accounting is the perfect example: same email types daily, same document reviews weekly, same reports monthly. AI doesn't just save time on each instance — it eliminates the cognitive load of starting from scratch every time. That's the hidden productivity gain nobody measures.
Publicly listed companies tend to be reporting their quarterly financials several days earlier than even a few years ago. AI has been a significant factor in the speed up. This is a quality improvement that isn't necessarily going to show up in productivity statistics.
The J-curve framing creates a measurement problem that the analysis doesn't address. Complementary investments are, by definition, a leading indicator; firm MFP (aggregated to TFP) is a lagging one. Including both in a single 'AI impact' figure means the J-curve isn't showing AI impact rising over time—it's showing the sum of (leading investments) + (lagging productivity) rising. Those two quantities have different dynamics and shouldn't be added without justification.
The shape is then treated as evidence for the mechanism. But the shape is equally consistent with the mechanism being absent or broken. Penn Wharton projects AI's contribution to TFP growth at roughly 0.01 percentage points in 2025: essentially zero. The late-2025 'uptick' looks more capex-driven (data centres etc.) than widespread evidence of firms extracting value from AI intangible. If complementary investments are climbing while TFP contribution is near zero, that's a widening gap between leading and lagging indicators, not a J-curve descent. Calling it the latter is a reading the data underdetermines.
The J-curve requires not just that intangible capital accumulates, but that firms eventually extract it as output. There's currently little empirical signal that either condition (accumulation, extraction) is being met. Babina et al. find zero relationship with sales per worker or TFP—though the study predates genAI and the replication question remains open—while Humlum and Vestergaard find zero effect on earnings or hours in Danish administrative data through end of 2024, even for workers who self-reported large productivity gains. We also need to consider that 'leading indicators' might be consumption rather than investment. If a firm buys AI because 'every other firm is buying AI' (a Red Queen effect), it's a defensive expense (consumption of capital) rather than a complementary investment.
The parsimonious reading of the current data is that the gap between leading and lagging indicators is widening. That's consistent with the J-curve—but it's equally consistent with the leading indicators leading somewhere else.