AI tools feel transformative day to day, but economy-wide productivity gains so far add up to roughly 1%, not a revolution.

This explores how much AI has actually raised productivity across the economy so far, as opposed to what forecasts and personal anecdotes suggest. Measured carefully, the answer is about 1%, not a revolution. Survey data on how much time people save with generative AI points to roughly a 1.1% gain in total productivity What is generative AI's actual aggregate productivity gain today?. That is close to Acemoglu's expert-based estimate of about 0.7%. Looking ahead, Acemoglu's task-by-task model caps AI's ten-year boost to overall productivity at no more than 0.66%, and possibly below 0.53% once you account for tasks AI learns poorly How much will AI actually boost productivity over ten years?. Those numbers sit far below the popular forecasts of 1.5% per year.

The opposing camp doesn't claim the gains are already large. It argues the gains are hidden. Brynjolfsson, Rock, and Syverson describe a 'productivity J-curve.' When companies adopt a technology like this, they make large investments that national statistics don't count: retraining staff, redesigning processes, building internal software. Measured productivity first dips below the true figure, then overshoots it once those investments start producing output Why does AI productivity appear weak when investment is high?. Brynjolfsson has argued that 2025's mix of slow job growth and continued GDP growth, about 2.7% productivity growth, is the upswing of that curve finally arriving. Other economists answer that AI still leaves no clear trace in employment, productivity, or earnings data outside the big tech firms Is AI productivity finally showing up in economic data?. So the honest answer to 'over the last decade' is that the measured contribution is small, and the live debate is whether that reflects reality or a measurement lag.

The most interesting material sits between the individual worker and the economy as a whole. Workers say AI saves them time almost everywhere: 86% of general workers and 97% of creatives in one interview study Why do workers hide productivity gains from AI use?. Much of that saving leaks away before it reaches the economy. Nearly 40% of saved time goes into fixing and checking AI output, and only 14% of employees consistently come out ahead Where does AI's time savings actually go in practice?. Self-reported productivity rises with up to three AI tools, then falls with four or more as the burden of supervising them grows Does using more AI tools always boost worker productivity?. Executives also see bigger gains than their financial results show, partly because revenue trails behind day-to-day improvements Do AI productivity gains feel larger than they actually measure?.

There is also a structural reason individual speedups don't add up neatly. AI mostly speeds up the 'doing' part of knowledge work. Deciding what to do and delivering the results remain, and sometimes grow, which is why fields like translation and law haven't shrunk the way you might expect Does AI really compress all layers of knowledge work equally?. When each person makes faster but more varied decisions, organizations face new coordination problems. They only capture the full benefit if they restructure how those decisions fit together Does AI adoption require rethinking how organizations coordinate?. And the gains depend on skill: AI helps people do work they already know how to do, but the boost disappears, and learning suffers, when they use it to pick up new skills When does AI actually boost worker productivity?.

The twist you might not expect is that what's productive for individuals may not be productive for the whole. In science, researchers who use AI publish three times as many papers and receive 4.8 times as many citations. Yet science as a whole covers fewer topics and involves less collaboration, as work clusters around problems with plenty of data Does AI help individual scientists while narrowing scientific focus?. Productivity statistics capture the first effect and miss the second. One caveat: this corpus mostly covers the generative AI era (2023 onward), so it says little about the earlier machine-learning decade.