Does AI Save Time at Work?

Does AI save time at work? A National Bureau of Economic Research working paper built on two decades of US government time diary data finds close to the opposite: an interquartile rise in a worker's occupational exposure to AI is linked to about 2.2 additional hours of work, not less. The tool sold as a time saver is, in the data collected so far, adding to the workday more often than shortening it.

Does AI save time at work, or just move it around?

The paper draws on the American Time Use Survey, a US government survey of roughly 26,400 people a year who log their day in 15 minute diary blocks, spanning 2004 to 2023. The researchers, Wei Jiang, Junyoung Park, Rachel Xiao and Shen Zhang, scored each occupation's AI exposure by matching its task descriptions against the language of AI patents. After controlling for personal characteristics, other innovation measures, and a wide set of fixed effects, an interquartile shift in occupational AI exposure corresponds to about 2.2 additional hours of work in the full sample, and roughly 40% of that magnitude when comparing workers within the same occupation over time.

What changed after ChatGPT launched

The end of 2022 gave the researchers a natural test. Occupations with higher exposure to generative AI saw a 3.15 hour increase in daily work time per interquartile increase in exposure after ChatGPT's launch, an effect strongest in occupations where AI complements the work rather than replacing it, and in regions where Google search interest in AI ran higher. This is a National Bureau of Economic Research working paper, meaning it has not yet cleared peer review, and a single study's estimate of this size is a finding to note carefully rather than treat as settled.

Why more AI would mean more hours, not fewer

The paper's explanation borrows from classic principal agent theory, not a claim about laziness or ambition. AI can raise how much a given hour of effort produces, which gives a worker reason to put in more hours for more reward. It can also make effort easier to observe. The researchers use the sudden shift to AI monitored remote work during the 2020 pandemic as a natural experiment and find that remote workers with greater exposure to AI surveillance technologies worked longer hours afterward, an effect not present among the self employed, who answer to no employer's dashboard. When effort becomes easier to see, it becomes easier to expect more of it.

Higher pay, the same satisfaction

The extra hours came with extra pay, but not extra contentment. Workers with greater AI exposure reported lower satisfaction on Glassdoor despite earning more, a gap the paper ties to reservation utility that fails to keep pace with productivity gains, particularly in competitive labor and product markets where workers have less power to keep the value AI helps create. The productivity did not disappear. It mostly went somewhere else.

What keeps the gains on your side of the ledger

If AI makes effort easier to measure, the open question is what gets measured. Frontier Professionals, the group of workers Microsoft found gets the most out of AI, do not just log more hours with it, they pause to decide what should go to AI and what should not, which is a judgment call a monitoring tool cannot make for them. The six domains this site measures, judgment among them, are exactly the kind of work that stays hard to automate and hard to reduce to a logged hour. A specific, checkable result is what turns a longer workday into a stronger case for your own value, not just more of it. The check gives you that kind of evidence in about seven minutes.

The honest limit

This is one working paper, not yet peer reviewed, describing an association across a large government dataset, not a law that applies to every AI user or every job. The size of the effect varies by how complementary a job is to AI, by whether a market is competitive, and by whether the work is monitored at all, and the self employed in this same data showed no rise in hours tied to AI surveillance. What the paper does settle is the more common assumption: AI has not, on the evidence collected so far, been handing workers their time back.