Subscribe

Note on the partial AI Solow Paradox

Note on the partial AI Solow Paradox
Street art on brick wall, Wellington. Photo by author

AI is currently in a partial Solow phase: there is good micro‑level evidence of productivity gains and user benefits, but these are only beginning to show up in aggregate productivity statistics and will likely do so with a multi‑year lag.

Why the paradox reappears with AI

Generative AI and assistants are now ubiquitous in white‑collar work and consumer apps. Official productivity data at the economy level still show only modest shifts so far, even as some firms report local efficiency gains and launch AI projects. There are also a growing number of reports which show the current suite of generative AI projects either failing or delivering little persistent value.

Leading economic statisticians note that AI adoption is still in early stages and that existing business and labour surveys aren’t yet well‑designed to capture AI’s contribution to output, quality and prices. Measures like GDP‑B suggest large consumer surplus from AI that is invisible in GDP and labour‑productivity metrics.

AI is productive in specific contexts

Experimental and field studies find sizable productivity improvements in tasks like customer support, coding, writing and some professional services, often with the biggest gains for lower‑skilled or less experienced workers. For example, one study finds AI adoption raises productivity, but effects are strongest in large firms and appear only after several years—early adopters saw little or no productivity increase at first.

PwC’s analysis of nearly a billion job ads and financial reports suggests that industries most exposed to AI have seen roughly a four‑fold increase in productivity growth and much faster growth in revenue per employee than less‑exposed sectors. Macro projections from structural models foresee AI adding around 1.5% to GDP by 2035, rising further over later decades, but with only a small permanent boost to annual productivity growth once the transition is complete.

Why the macro statistics lag the micro story

Work on the original Solow paradox emphasised that computers only raised measured productivity after firms restructured processes and business models around them. AI similarly requires complementary investment in skills, workflows, and organisational change, so early spending often looks like higher cost without immediate output gains.

Some of AI’s impact comes from shifting activity across firms and sectors, raising productivity in adopters while displacing output or jobs elsewhere, so aggregate statistics can understate early net gains or show them only once reallocation stabilises.

Work using GDP‑B shows that AI is already generating tens of billions of dollars in consumer surplus in the US alone, far exceeding the revenue of AI firms, but this value is missing from standard GDP and labour‑productivity measures. That gap makes the AI era look more paradoxical than it is from a welfare perspective.

How much does the Solow paradox apply?

It's a useful analogy, but not a perfect repeat. As with the computer era, we see widespread AI visibility, modest early macro productivity effects and evidence that benefits are highly uneven across tasks, firms and sectors.

We also already have credible firm‑ and sector‑level data showing substantial productivity gains where AI is deeply integrated, plus forward‑looking estimates that AI will lift productivity growth in the 2030s as diffusion and reorganisation advance. We can also quantify large user surplus that didn't appeared in earlier debates.

This suggests less a puzzle about whether AI is productive, and more a transitional phase where measurement, diffusion and re‑design haven’t yet caught up with the technology.

Sources

  1. The AI Utility Trap, https://jamesthomason.com/the-ai-utility-trap/
  2. Taking AI Commoditisation Seriously, https://www.techpolicy.press/taking-ai-commoditization-seriously/
  3. The Great AI Power Shift - From AI Models to AI Applications, https://aiproem.substack.com/p/the-great-ai-power-shift-from-ai
  4. The Future of Data Context in Enterprise AI, https://www.decube.io/post/data-context-enterprise-ai
  5. Information as an Asset and AI – An Introspective, https://www.mssbta.com/post/information-as-an-asset-and-ai-an-introspective
  6. Turning Business Data into a Valuable Asset for AI Models, https://www.linkedin.com/pulse/turning-business-data-valuable-asset-ai-models-mohammed-shareefuddin-n8xdf