The AI Productivity Paradox: Why Are Micro Gains Not Reflected in GDP?

In 1987, economist Robert Solow coined one of the most famous phrases in modern economics: "You can see the computer age everywhere but in the productivity statistics." The observation became known as Solow's Paradox.

Decades later, when productivity finally accelerated in the 1990s, it became clear that the computer revolution had simply taken longer than expected to show up in the numbers.

Today, faced with the explosion of artificial intelligence, the question is repeated with renewed urgency: if AI is transforming the way we work, where is the growth? If technology is increasing the productivity of certain tasks by 20%, 30% or 50%, why aren't GDP and aggregate productivity accelerating at the same rate?

The question is not rhetorical. Reports of productivity gains with AI are abundant and impressive. Customer service experiments show increases of up to 15%. Analysis and writing tasks report gains of 25% to 50%. Software developers describe even greater productivity increases in certain activities.

The provocation, then, is direct: if technology companies (which already represent more than a third of the American market value) are growing 5 or 10 times in productivity, economic models should be predicting GDP growth well above what is observed.

And yet, the American economy's productivity grew by just 0.8% in 2025, according to the BLS. The forecast for 2026 is 1.1%, according to the CBO. The macro number seems disconnected from the micro reports. Why?

The answer involves understanding how the productivity of a task is transformed —or not— into aggregate economic growth. Imagine a programmer who dedicates 40% of his time to writing code. AI increases your productivity on this task by 40%. Impressive, but the gain for the entire worker already drops to 16%.

If only half of the programmers intensively adopt the tool, it drops to 8%. If half of this saved time is converted into additional production — not pauses, rework or simply doing the same thing slower —, we reach 4%.

Then we still need to consider the weight of these workers in the total product of the economy. The “40% AI gain” could easily turn into less than 2% effective economic productivity. This funnel (from task to worker, from company to industry, from industry to economy) is where much of the promise of AI gets lost along the way.

Recent NBER research with almost 750 American executives documents exactly this paradox: perceived gains are systematically greater than measured ones, possibly because revenue realization takes longer than operational improvement.

Other adoption surveys reinforce the diagnosis: only 23% of American workers use generative AI at least once a week, and only 9% daily. A 2026 St. Louis Fed study describes the situation as spread across many occupations but shallow within each. High productivity conditional on intensive use is not high average productivity in the economy.

There are two major competing explanations for this disconnect. The first is the J-curve hypothesis, developed by Erik Brynjolfsson and his co-authors: general-purpose technologies — such as electricity in the last century and computers in the 1980s — require complementary investments before fully producing their effects.

Reorganization of companies, training, new processes, new organizational capital. The factory did not become dramatically more productive when it replaced the steam engine with the electric motor. It was necessary to redesign the entire factory around electricity. Perhaps we are making the same mistake by imagining that an old company plus AI equals an old company that is 30% more productive.

The second, more skeptical explanation comes from Daron Acemoglu: the macro effect depends on the fraction of tasks actually affected multiplied by the cost savings on those tasks. His estimate is sober: no more than 0.7% cumulative increase in productivity over ten years.

In other words, either the revolution is coming, but it has not yet reached the statistics, or we are unduly extrapolating extraordinary microeconomic results.

There is an additional distinction that rarely appears in public debate and that seems fundamental to me. Even if productivity gains are confirmed, it is necessary to observe whether we have a level effect (a single jump that generates a few years of additional growth during the transition) or a permanent growth effect.

For the GDP growth rate to accelerate in a lasting way, it would be necessary for the frontier of knowledge itself to continue advancing, because AI accelerates research, scientific innovation and the creation of new technologies. This distinction matters. A world in which AI makes us more productive all at once is very different from a world in which AI permanently accelerates the pace of innovation.

What does all this mean for Brazil and emerging economies? The relevant question is not just whether or not AI will transform global productivity. It's who will capture these gains and at what speed.

Countries that invest in adoption, qualification and digital infrastructure will be able to take advantage of the leap in productivity when it arrives — and position themselves better in the value chain of the new economy.

The Solow Paradox was resolved: computer productivity appeared, but it took more than a decade and required profound transformations in organizations and institutions. Will the AI ​​revolution follow the same path?

The question is not whether growth will come. It's whether we'll be prepared when he arrives. Another question is whether the current prices of AI-linked assets are in sync with this growth curve. But that's a discussion for a future column.