Labor Market Impacts Are Not the Same As Welfare Impacts

AI may not be measured for a while

Wages are not the only thing which a company pays its employees. Companies pay a package of amenities, which includes some mix of culture, working conditions, work responsibilities, benefits, and pay. This is not a trivial concern. When the government changes the minimum wage, for example, employers adjust by spending less on working conditions to partly balance out the increased cost of employment. Jeffrey Clemens (2021) covers how employers will adjust to minimum wage hikes, such as by moving toward employer-driven rather than worker-driven schedules. More recently, Anders Humlum, Mette Rasmissen, and Evan Rose use a survey to estimate how much of the differences in wages between firms are due to differences in productivity, or due to differences in amenities. There are wage premiums due to productivity, but the two trade off against each other. Half of the benefit of moving to a higher paying firm is offset by the worse amenities.

We have also kept working a lot, which is another mystery. Back in the day, economists like Keynes predicted that our descendants would work far less than we do. If goods continued getting cheaper, then eventually the income effect would overcome the substitution effect, and we would consume less and less. Epstein and Kimball (2022) rule out some alternative explanations, like income and substitution effects equaling zero, and are left with two stories – new goods keeping the marginal utility of consumption high, or work improvements keeping the marginal disutility of labor low.

I have written about the introduction of new goods before here, but I’d like to focus on the latter. There are lots of ways in which work has gotten more pleasant. We lift less, and engage in less physically arduous labor; we have air conditioning; we have phones and music; we have the ability to skip commuting and call into the office; we might have more accommodating sick leave policies; we can drive to work. AI is another improvement in pleasantness – it is so much easier to have something else do the rote thinking of writing emails and blog posts. (ope)

We would like to know how much AI is increasing our productivity, but if the productivity increase is coming through people having a more pleasant time on the job and working less, we will not be able to pick it up. We would have to have extremely detailed data on time use, which is simply not available in standard administrative data. This is every bit as much a productivity improvement as an increase in output, but we would totally miss it.

There has been work which tries to include time savings from AI by asking about. For example, Humlum and Vestergaard (2025) include survey data on the amount of time worked. They found that time spent on tasks was reduced (on average) by 2.8%. They don’t refer to this improvement as being all that large, but I disagree with them – it’s a single technology effectively increasing GDP by 2.8%! I also believe this to be an underestimate of the true welfare gains from AI – we can’t include how it would reduce the unpleasantness of thinking.

Please do not take studies which show small labor market effects of AI, and infer that this means that AI is doing nothing to increase productivity. This is incorrect. Labor market impacts are not able to capture the full benefit of AI, even in the long run.