AI-driven Productivity Growth Is Here. And It's Just Getting Started
Yes, I should have said "probably here." But I didn't.
My colleague, co-author, co-founder, and great friend Erik Brynjolfsson just published an editorial in the FT. It’s typical for him, which means it’s great: clearly written, well argued, and data-driven. It’s also typically circumspect; Erik is not yet claiming victory yet in his wager about US productivity growth.
Erik bet fellow economist Bob Gordon1, that, as the Long Bets page formalizing their wager puts it:
Private [US] Nonfarm business productivity growth will average over 1.8 percent per year from the first quarter (Q1) of 2020 to the last quarter of 2029 (Q4).
If this sounds like a dispute between two huge productivity nerds, it is. But this bet deserves a much wider audience because it’s about how quickly we achieve a higher standard of living.
I sometimes wonder if economists come up with the blandest possible phrases to describe the stuff that they study. “Higher standard of living” sounds like something only an accountant could get behind. We should all have it as a goal, though, because a lot of the things that we want fall underneath that label. Better health and health care, less pollution, increased opportunities to learn and acquire new skills, seeing more of the world, dwelling where we want — these are all aspects of a higher standard of living, as measured and obsessed over by economists.
The higher our productivity growth numbers, the faster we get all of that stuff.2 As Paul Krugman put it more than thirty years ago, “Productivity isn’t everything, but in the long run it is almost everything.”
Bob was pessimistic that America’s (almost-) all-important productivity growth would hit Erik’s 1.8% per year benchmark throughout the 2020s. After all, it was a lofty benchmark. As Gordon pointed out, productivity growth had only averaged 1.1% throughout the 2010s. For Erik to be right, productivity growth would have to increase by more than 60% in the following decade.
Erik anticipated that a big ol’ productivity speedup of that magnitude would in fact happen. In large because, as he put it,
AI is a general-purpose technology that is affecting almost every industry while accelerating the pace of discovery.
He cautioned, though, that it would take a while for AI’s productivity benefits to appear in the productivity statistics. Those stats might even take a dip for a while because reconfiguring organizations to take advantage of general-purpose technologies is hard work requiring a lot of investment.3 But to paraphrase Erik paraphrasing Sam and Dave, “Hold on, they’re coming.” (“they” being 2020s productivity growth numbers averaging 1.8%). As he wrote in the FT:
The updated 2025 US data suggests we are now transitioning out of this investment phase into a harvest phase where those earlier efforts begin to manifest as measurable output.
If that’s true, Erik’s going to win the bet. Another colleague, co-author, co-founder, and great friend of mine, Daniel Rock, also thinks he will. As Rock told me when we were talking about the bet and the editorial, the AI that’s likely showing up in the today’s productivity statistics is machine learning and deep learning — the great artificial intelligence Eureka!s of the 2010s. Generative AI, the great Holy S***! of the 2020s, is still too young to be showing up in economy-wide productivity statistics.
But it will.
My X feed is heavily skewed toward the alpha geeks of artificial intelligence and economics. And these days, it’s full of those folk saying some version of “Holy S***! This stuff is amazing. It gives me superpowers.” Very high on the list of superpowers conferred by modern AI is the ability not just to help people write code, but to write large amounts of high-quality code for them. I’ve lost track of the number of folk who have posted or said to me some version of “I wrote a thousand lines of Python yesterday. I don’t know Python.”
Rock does write code, so his wonderment at the power of AI is at a different order of magnitude. At our last Workhelix on-site, he turned to me at the end of a day and said, “I wrote 10,000 lines of production-ready code today.” Anthropic CEO Dario Amodei will be hard to beat in the Holy S***! sweepstakes. I saw him at Davos last month and asked him if it were true that Claude Cowork was essentially entirely written by AI. He got That Look on his face and said “Yes! You have to try it!”
I started using Claude Cowork as soon as I got home, and am continually astonished by it. And Google’s Gemini, OpenAI’s ChatGPT, and who knows what next.
As I wrote when I was visiting the Technology and Society Group at Google in 2024, a big difference between genAI and previous general-purpose technologies is the speed with which it will diffuse throughout the economy. For reasons that I’ve written about a fair bit, and I’m sure will write about more, I don’t I believe that this fast diffusion will be accompanied by massive, sustained technological unemployment.
But I do believe that not paddling hard right now to get in position and catch this genAI wave is a serious mistake for anyone running an organization.4 Because it’s going to be a big wave of productivity growth and innovation.
Anyone want to bet otherwise?
And fellow mensch
Of course, distributional issues matter a lot, too. We don’t just want an economy that generates lots of stuff; we also want an economy that allows everyone a fair shot to share in that bounty and enjoy a rising standard of living.
Erik, Daniel Rock, and Chad Syverson coined the phrase “J-curve” to describe this pattern of falling-then-rising productivity growth as a general purpose technology diffuses.
Erik, Daniel, and I (along with James Milin) cofounded Workhelix to help enterprises with that paddling and positioning. Get in touch with us if you’d like to learn more.
Political appointees and policy 'planners' seriously need to stop making hyperbolic pronouncements though about current-gen AI doing things like "inventing new physics", whatever their reason.. I'm not sure if they realise it or not, but those are falsifiable claims and not things that should be thrown around lightly, EVEN if intentional bluster.
like I'm sorry but just shut the fuck up. That is not the same thing as saying productivity growth (which still needs to be demonstrated..).
I have serious doubts. For one, the only thing the US does right now is variants of software or knowledge work like finance and consulting. To raise the standard of living of the population you need better infrastructure, and more integrated manufacturing capability. That is China's game right now. Productivity growth might meet your target numbers if you count things that actually lower the standard of living, like the social media and doomscrolling sectors, or the wealth extraction activities of private equity that cause immiseration in the wider public. I just don't see the US getting better and faster trains, more and better housing, more nuclear power plants, or more widely available and cheaper health care anytime between now and 2029, regardless of the amount of AI you throw in. The other inputs don't seem to be there.