Yves here. While this post flags an important central finding, that the more a job is exposed to AI use, the longer its hours become, its authors seem awfully credulous about the supposed benefits. Specifically, they consider that the additional time required is to clean up AI messes due to fad-enamored managers requiring its use when it is not ready for prime time.
For instance, earlier this year, IM Doc reported on the train wreck of AI writing up patient visit notes which seem entirely plausible (as in accurate) Wwhen rife with serious errors. And that is routine. From his e-mail:
The visits are now being recorded and about 10 minutes later – the AI generated visit notes appear in the chart. On almost 2/3 of the charts that are being processed, there are major errors, making stuff up, incorrect statements, etc. Unfortunately – as you can see it is wickedly able to render all this in correct “doctorese” – the code and syntax we all use and can instantly tell it was written by a truly trained MD.
I have noted to my dismay that most of my colleagues do not even look at these – they simply sign off on them. That is a fatal tragedy.
He then recounted a particular case, with enough anonymized backup to support his claims.
The patient had come for a routine visit. No new issues, a recitation of routine aches and pains, and a review of recent bloodwork and other regular tests.
The AI invented that the patient was on dialysis when the patient had never had kidney issues. There there was no discussion of renal issues or tests in the visit, much the less the AI’s claim that a congenital defect caused the (supposed) kidney issues. It also claimed the patient had had a particular operation (falese) and had recently been septic. The AI also depicted the patient as unable to drive due to cataracts resulting from the administration of steroids, another fabrication.
As IM Doc pointed out:
1) Had I signed this and it went in his chart, if he ever applied to anything like life insurance – it would have been denied instantly. And they do not do seconds and excuses. When you are done, you are done. If you are on dialysis and have cataracts and cannot drive – you are getting no life insurance. THE END.
2) This is yet another “time saver” that is actually taking way more time for those of us who are conscientious. I spend all kinds of time digging through these looking for mistakes so as not to goon my patient and their future. However, I can guarantee you that as hard as I try – mistakes have gotten through. Furthermore, AI will very soon be used for insurance medical chart evaluation for actuarial purpose. Just think what will be generated.
3) These systems record the exact amount of time with the patients. I am hearing from various colleagues all over the place that this timing is being used to pressure docs to get them in and get them out even faster. That has not happened to me yet – but I am sure the bell will toll very soon.
4) When I started 35 years ago – my notes were done with me stepping out of the room and recording the visit in a hand held device run by duracells. It was then transcribed by secretary on paper with a Selectric. The actual hard copy tapes were completely magnetically scrubbed at the end of every day by the transcriptionist. Compare that energy usage to what this technology is obviously requiring. Furthermore, I have occasion to revisit old notes from that era all the time – I know instantly what happened on that patient visit in 1987. There is a paragraph or two and that is that. Compare to today – the note generated from the above will be 5-6 back and front pages literally full of gobbledy gook with important data scattered all over the place. Most of the time, I completely give up trying to use these newer documents for anything useful. And again just think about the actual energy used for this.
5) This recording is going somewhere and this has never been really explained to me. Who has access? How long do they have it? Is it being erased? Have the patients truly signed off on this?
6) This is the most concerning. I have no idea at all where the system got this entire story in her chart. Because of the fake “Frank Capra movie” style names in the document [inclusion of towns and pharmacies that do not exist] I have a very unsettled feeling this is from a movie, TV show, or novel. Is it possible that this AI is pulling things “it has heard” from these kinds of sources? I have asked. This is not the first time. The IT people cannot tell me this is not happening.
Admittedly, to the authors’ credit, they do discuss one downside that IM Doc highlighted: that of even more extensive/intensive management surveillance.
By Wei Jiang, Junyoung Park, Rachel Xiao and Shen Zhang. Originally published at VoxEU
Technological progress is typically expected to lighten the burden of work. But as artificial intelligence has been integrated into workplaces, early evidence suggests a paradox: instead of reducing workloads, many AI-equipped employees are busier than ever. This column examines the relationship between AI exposure, the length of the workday, time allocation, and worker satisfaction. Though AI-driven automation and delegation allow workers to complete the same tasks more efficiently, the authors find that employees in AI-exposed occupations are working longer hours and spending less time on socialisation and leisure.
For much of modern history, technological progress has been expected to lighten the burden of work. Keynes (1930) predicted that by 2030, rising productivity would allow people to work 15 hours a week. As artificial intelligence is integrated into workplaces, early evidence suggests a paradox: instead of reducing workloads, many AI-equipped employees are busier than ever. While AI-driven automation and delegation allow workers to complete the same tasks more efficiently, employees in AI-exposed occupations might well be working longer hours and spending less time on socialisation and leisure.
The rapid diffusion of AI, exemplified by the introduction of ChatGPT in late 2022, has reignited concerns about its effects on employment. A large body of work examines how AI is replacing some job functions while augmenting others – the extensive margin of employment (e.g. Acemoglu et al. 2022, Albanesi et al. 2023, Bonfiglioli et al. 2024, Felten et al. 2019, Gazzani and Natoli 2024). Less attention has been given to how AI influences time allocation – the intensive margin of employment. If workers retain their jobs, does AI make them work more or less? Our study (Jiang et al. 2025) examines the relationship between AI exposure, the length of the workday, time allocation, and worker satisfaction.
Using nearly two decades of time-use data from the American Time Use Survey (ATUS), we link AI-related patents with occupational descriptions to construct a measure of AI exposure across jobs. We then distinguish AI that complements human labour – enhancing worker productivity – from AI that substitutes for it, potentially displacing workers.
AI complementary/substitutive exposure varies significantly across occupations, as illustrated in Figure 1. At the forefront are computer and information system managers, bioinformatics technicians, and management analysts, fields where AI enhances productivity rather than substituting labour. In contrast, jobs like data entry keyers, tellers, and office machine operators face high AI substitutive exposure but little complementarity, facing the risk of displacement rather than augmentation. Meanwhile, occupations such as dancers and barbers sit at the bottom of the AI spectrum, largely untouched by AI advancements.
Figure 1 Complementarity with AI
A third category of AI exposure – AI-driven monitoring – captures how surveillance technologies track employee effort. This framework allows an examination of whether AI lengthens or shortens the workday and whether these effects vary across labour markets.
The findings, summarised in Figure 2, reveal a pattern: higher AI exposure is associated with longer work hours and reduced leisure time. Over the 2004–2023 period, workers in AI-intensive occupations increased their weekly work hours relative to those in less exposed jobs. An increase from the 25th to the 75th percentile in AI exposure corresponds to an additional 2.2 hours of work per week. This relationship is strengthened over time, suggesting that as AI becomes embedded in workplaces, its effect on working hours intensifies. This challenges the expectation that automation enables workers to complete tasks more quickly and reclaim leisure time.
Figure 2 Weekly working hours and AI exposure
The introduction of ChatGPT provides an unexpected shock to generative AI adoption and serves as a natural experiment (Hui et al. 2023). Occupations more exposed to generative AI saw a rise in work hours immediately following the release of ChatGPT. Compared to workers less exposed to generative AI (such as tire builders, wellhead pumpers, and surgical assistants) those in high-exposure occupations (including computer systems analysts, credit counsellors, and logisticians) worked roughly 3.15 hours more per week in the post-ChatGPT period. This shift was accompanied by a decline in leisure time, reinforcing the idea that AI complements human work in a way that increases labour supply rather than reducing it. When leisure hours are cut, non-screen-based activities – especially entertainment and socialisation – bear the brunt. Screen-based leisure activities, such as watching television and playing video games, remain relatively stable, suggesting that workers are more likely to sacrifice activities that require active participation rather than passive consumption, signalling a shift towards more isolated and sedentary downtime.
Two key mechanisms help explain this result. First, AI raises worker productivity, creating incentives for longer hours. When AI complements human labour rather than replacing it, the process makes each hour of work more valuable. This effect is strongest in jobs where AI helps employees perform tasks more efficiently, such as finance, research, and technical fields. Employers may expect more output; workers, incentivised by productivity-linked pay, may extend their hours. AI-exposed occupations have indeed seen wage increases, suggesting that firms are sharing some productivity gains. However, higher wages have not translated into more leisure time. Instead, workers appear to be substituting additional earnings for longer hours, a pattern consistent with the economic principle that when work becomes more rewarding, people may choose to do more of it.
The second mechanism is AI-driven performance monitoring. Digital surveillance tools have expanded, particularly in remote and hybrid work environments. AI enables real-time tracking of employee effort, leading to longer working hours. Our study examines the COVID-19 period as a natural experiment, when AI-driven monitoring surged due to remote work. Jobs that were more ‘remote-feasible’ at the onset of COVID-19 experienced dramatic improvement in remote work monitoring during the next two years. Occupations with high exposure to AI surveillance technologies – such as customer service representatives, stockers and order fillers, dispatchers, and truck drivers – experienced longer work hours post COVID even after workers returned to the office. This effect was absent among the self-employed, confirming that it is not simply the nature of AI-exposed jobs but the principal-agent dynamics of employment that drive longer work hours. Monitoring increases employer oversight and tightens performance expectations, often at the cost of work-life balance. Some AI-intensive roles saw the introduction of automated performance scores, leading employees to work harder to avoid falling behind peers in algorithm-driven assessments.
The broader question is: who benefits from AI-driven productivity gains? While AI-exposed workers may see wage increases, these gains do not translate into improved wellbeing. Employee satisfaction data from Glassdoor show that higher AI exposure is associated with lower job satisfaction and work-life balance ratings. While AI may boost output and compensation, it does not necessarily enhance workers’ quality of life. Many of AI’s productivity gains accrue to firms and consumers rather than to workers.
Labor and product market competition shapes these dynamics. First, AI’s impact on work hours is amplified in competitive labour markets, where workers have less bargaining power because there are only a few employers who dominate local hiring. In such environments, employees are less able to demand shorter hours or better compensation for increased productivity. Second, in highly competitive product markets (where products in the industry are similar), firms have an incentive to pass productivity gains on to consumers in the form of lower prices or better services, rather than sharing them with workers through reduced workloads. The result is that while AI makes workers more productive, they do not necessarily see corresponding improvements in work-life balance. Instead, they work more hours to maintain their employment.
AI’s role in the future of work is not predetermined. The extent to which it leads to longer or shorter hours depends on how firms deploy the technology and how policymakers respond. Our research provides insight into this debate, showing that AI is not inherently liberating or oppressive. The impact of AI on work hours is shaped by the incentives and constraints of labour, product, and capital markets. If AI is to improve lives, more deliberate and well thought-out approaches will be necessary to distribute its benefits fairly.
See original post for references
