The AI Productivity Paradox

Why Our AI Bet Has Not Hit GDP

The AI Productivity Paradox (2025): Real gains, slow totals

AI often helps at the task level. Yet the big numbers (firm productivity, GDP per hour) move slowly. That mismatch is the AI Productivity Paradox.

This is not mysterious. It is what general-purpose tech usually does. It boosts some tasks early, then takes years to reshape real production.

1. Micro gains are real, but uneven

In some settings, AI helps a lot:

• Call centers: better outcomes and faster work in measured deployments, with bigger gains for less-experienced workers (Brynjolfsson et al., 2023).

• Writing and knowledge work: controlled studies often find faster output and higher-rated drafts for many users (Noy & Zhang, 2023).

But results are not always positive:

• Software work can go either way. A careful 2025 field study of experienced open-source developers found that AI tools made them slower on average (about 19% more time), even while they believed they were faster (METR, 2025).

So the first rule is simple: AI is not a universal speed button. It can help. It can also add drag.

2. Macro gains depend on diffusion, not demos

The macro question is not "can AI do a task?" It is:

• How many firms adopt it?

• How deep does it get into workflows?

• How much does it change how work is done?

OECD work that maps micro estimates into macro scenarios finds AI could add meaningful points to productivity growth in some cases, but the size and timing depend mostly on adoption speed and breadth (Filippucci et al., 2024).

In plain terms: diffusion is the timeline.

3. Why the paradox persists

3.1 Adoption takes time

Even obvious tech diffuses slowly. It often takes decades to reorganize production and skills (Comin & Hobijn, 2010).

3.2 Tools need complements

You do not get productivity by "adding AI." You need:

• clean data

• system integration

• training

• process redesign

• quality control

• incentives that reward outcomes

Without these, the tool becomes one more thing to manage (Brynjolfsson et al., 2023).

3.3 Measurement misses early value

Early value often shows up as:

• better quality

• faster iteration

• better decisions

• new capabilities

GDP and standard productivity measures can miss those gains, or capture them late (David, 1990).

3.4 Errors can eat the gains

If AI increases output but also increases mistakes, you pay in review and rework. Net productivity can fall. METR is a clean example: more time spent, more confidence, worse reality (METR, 2025).

3.5 Governance adds friction

Rules can be needed, but they add cost and delay. Stanford AI Index reports rising AI-related legal and regulatory activity in the US in 2024, including more state laws and more federal AI-related regulations introduced (Stanford HAI, 2025). The policy context can also shift quickly, which increases planning risk (NIST, n.d.).

4. The right mental model: a J-curve

Expect a J-curve:

1. Early spend and disruption

2. Weak measured gains while workflows get rebuilt

3. Later gains when adoption and complements mature

This is how many big tech shifts look.

5. How to make AI show up in your numbers

If you want measurable productivity (not just "cool"), do the boring work:

1. Pick one workflow with one metric.

   • cycle time, defect rate, cost per ticket, conversion rate

2. Measure the baseline first.

3. Redesign the process around AI. Do not bolt it on.

4. Add cheap quality gates.

5. Train evaluation skills. Prompting is the easy part.

6. Fix the data and integration bottlenecks.

7. Scale only after you can repeat net gains.

Bottom line: AI can raise productivity. But it shows up late, after you rebuild the system that uses it.

---

References

Brynjolfsson, E., Li, D., & Raymond, L. R. (2023). Generative AI at work. National Bureau of Economic Research. https://www.nber.org/papers/w31161

Comin, D., & Hobijn, B. (2010). An exploration of technology diffusion. American Economic Review, 100(5), 2031-2059. https://www.aeaweb.org/articles?id=10.1257/aer.100.5.2031

David, P. A. (1990). The dynamo and the computer: An historical perspective on the modern productivity paradox. American Economic Review, 80(2), 355-361. https://www.aeaweb.org/articles?id=10.1257/aer.80.2.355

Filippucci, F., Gal, P., & Schief, M. (2024). Miracle or myth? Assessing the macroeconomic productivity gains from artificial intelligence. ECOSCOPE (OECD). https://oecdecoscope.blog/2024/11/26/miracle-or-myth-assessing-the-macroeconomic-productivity-gains-from-artificial-intelligence/

METR. (2025, July 10). Measuring the impact of early-2025 AI on experienced open-source developer productivity. https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/

NIST. (n.d.). Executive Order on safe, secure, and trustworthy artificial intelligence. https://www.nist.gov/artificial-intelligence/executive-order-safe-secure-and-trustworthy-artificial-intelligence

Noy, S., & Zhang, W. (2023). Experimental evidence on the productivity effects of generative artificial intelligence. Social Science Research Network. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4375283

Stanford Institute for Human-Centered Artificial Intelligence. (2025). AI Index Report 2025: Chapter 6, Policy and Governance. https://hai.stanford.edu/assets/files/hai_ai-index-report-2025_chapter6_final.pdf