Where does the time saved with generative AI go?

Research question

When workers save time with generative AI, how much of it goes to more of the same output, to new tasks, to checking AI output, to on-the-job leisure or to shorter hours, measured with time-use diaries or telemetry linked to the same workers' measured output?

Why it matters

Self-reported time savings come to 1.4% of US work hours, executives mostly report no productivity impact yet, and technical workers report speed gains larger than the value gains they report. A METR note shows why time saved on a shifted mix of tasks can overstate value. Whether saved time becomes output is the hinge between task gains and firm outcomes, and it is currently inferred from self-reports rather than measured.

The debate lists this among what we still do not know: Where does saved time go: more output, new tasks, checking, or leisure? No study follows saved time directly; Danish register data and one Korean survey come closest.

Evidence that motivates it

1. In nationally representative US surveys to late 2024, workers' self-reported time savings from generative AI came to 1.4% of all work hours, which the authors translate into a potential aggregate productivity gain of about 1.2%; they note that worker adoption is running ahead of firm adoption. (LD03-SL-C02 v1.0)
   Workers' self-reported time savings equal 1.4% of all US work hours; the potential productivity gain assumes the saved time becomes output.
   Cites: Bick et al. (2025), “The Rapid Adoption of Generative AI”
2. In a six-month randomized trial of an integrated AI assistant across 66 firms and 7,137 knowledge workers, workers who used it spent about two fewer hours a week on email, but the authors detect no shift in the quantity or composition of workers' tasks from individual-level provision, and no change in time spent in meetings. (LD03-SL-C06 v1.0)
   Telemetry shows users spending about two fewer hours a week on email and less work outside regular hours, without tracing where the time went or linking it to output.
   Cites: Dillon et al. (2025), “Shifting Work Patterns with…”
3. A METR note shows that when workers shift towards the tasks AI speeds up, time saved measured on the new mix of tasks overstates the value gained, which lies between the speedup on the old task mix and on the new one under the note's assumptions; the authors suggest this may help explain why reported gains exceed aggregate data. (LD03-SL-C13 v1.0)
   When workers shift towards tasks AI speeds up, time saved measured on the new task mix overstates the value gained.
   Cites: Cunningham & Whitfill (2026), “Task Substitution and Uplift”
4. In a METR survey of 349 technical workers in early 2026, the median respondent reported that AI tools made them about 3 times faster but changed the value of their work by only 1.4 to 2 times; METR warns that such self-reports have overstated measured effects in its own trials. (LD03-SL-C14 v1.0)
   The median surveyed technical worker reports being about three times faster but producing only 1.4 to 2 times the value.
   Cites: Becker (2026), “Measuring the Self-Reported Impact of…”

Claims are shown as they read on the day this pack was built; claim pages keep every earlier version.

A possible design

Panel of workers in several organizations with staggered AI access (randomized timing where the employer agrees, otherwise the rollout schedule as a natural experiment). Before and after access, run two-week experience-sampling waves (several short prompts a day: what are you doing, is it AI-related, are you checking AI output, are you on a break) alongside consented application telemetry and recorded hours. Link to individual or team output records already kept by the employer. Decompose the change in time into more of the same output, new tasks, checking and oversight, on-the-job leisure and shorter hours.

Data needed

One or more employer partners with a staged rollout; an experience-sampling app; consented telemetry; output measures at the individual or team level; ethics approval for workplace monitoring.

Feasibility

A PhD project. Low to medium; experience-sampling software and participant incentives are modest, and the hard part is employer access and consent to link output data. Recording on-the-job leisure is sensitive. Opt-in consent, aggregate reporting and a written guarantee that the data are not used for performance management are needed, or participation and candour will drop.

Has this been done?

Partly done already: related work covers some of this question (first-pass screen of 15 October 2026; not yet checked in depth).

No study was found that follows the same workers' saved time with a time-use diary, experience sampling or telemetry and links it to measured output, so the proposed decomposition is untested. The nearest quantitative evidence is a cross-sectional, self-reported Korean survey that finds time savings almost uncorrelated with output changes and attributes the gap to on-the-job leisure. Danish register data show new tasks and no change in recorded hours or earnings but do not measure within-day time. Vendor-run telemetry measures time on specific tasks without tracing the freed time. An ethnographic study points the other way, to work that intensifies.

Where the search could not reach. Screened by web search (5 queries) and arXiv and NBER abstract pages on the date shown; national time-use-survey work and the CHI and CSCW digital libraries, where diary studies of AI use are likely, were not searched directly. Vendor surveys that appeared in results were left out as grey literature. Absence of a hit is not proof that the work does not exist.

Before building a project on it, run your own search in a scholarly database such as Scopus, Web of Science, Google Scholar or SSRN.

Nearest work found

• Suh & Oh, "Generative AI and the Reallocation of Time: Productivity, Leisure, and Fulfilling Work," arXiv:2602.12695 (2026) Judged from an abstract, summary or search listing only
  Abstract read. A representative Korean worker survey: generative AI cuts working time by 3.8%, but time savings and output changes are almost uncorrelated, which the authors attribute to on-the-job leisure. The closest to the question, but self-reported and cross-sectional, with no diary or telemetry.
• Humlum & Vestergaard, "Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI," NBER Working Paper 33777 (revised March 2026)
  Full text read for the LD-02 recheck (LD02-HQ-C02). Danish adoption surveys linked to administrative records: precise nulls on earnings and recorded hours, and work reorganized into new tasks including AI oversight. Does not measure where within-day time goes.
• Dillon, Jaffe, Peng & Cambon, "Early Impacts of M365 Copilot," arXiv:2504.11443 (2025) Judged from an abstract, summary or search listing only
  Abstract read. Randomized experiment with over 6,000 workers in 56 firms: users spent less time reading email and finished documents faster. Time on specific tasks, not where saved time goes. Run by Microsoft researchers on Microsoft's product.
• Dillon, Jaffe, Immorlica & Stanton, "Shifting Work Patterns with Generative AI," arXiv:2504.11436 (v4) / NBER Working Paper 33795 (2025)
  Full text read for the LD-03 debate (LD03-SL-C06). About two fewer hours a week on email among users and less work outside regular hours: partial evidence that some time goes to more regular hours. No output linkage.
• Ranganathan & Ye, "AI Doesn't Reduce Work—It Intensifies It," Harvard Business Review (February 2026) Judged from an abstract, summary or search listing only
  Judged from search listings and a university news page; not read directly. Reportedly an eight-month ethnography at a US technology company of about 200 people in which workers took on more and wider tasks and longer hours. A practitioner article about research described as in progress.

References

Sources behind the motivating claims:

• Bick, Blandin & Deming, “The Rapid Adoption of Generative AI,” Federal Reserve Bank of St. Louis Working Paper 2024-027F, October 2025 (SRC-SL-BICK-BLANDIN-DEMING-STLFED-WP-2024-027F, original)
• Dillon, Jaffe, Immorlica & Stanton, “Shifting Work Patterns with Generative AI,” NBER Working Paper 33795, revised November 2025 (SRC-SL-DILLON-JAFFE-IMMORLICA-STANTON-ARXIV-2504-11436-V4, original)
• Cunningham & Whitfill, “Task Substitution and Uplift,” METR, 8 May 2026 (SRC-SL-CUNNINGHAM-WHITFILL-METR-2026-05-08, original)
• Becker, “Measuring the Self-Reported Impact of Early-2026 AI on Technical Worker Productivity,” METR, 11 May 2026 (SRC-SL-BECKER-METR-SELF-REPORT-SURVEY-2026-05-11, original)

BibTeX

The first entry cites this opportunity; then its nearest work; then the sources above. Nearest work and sources are recorded as reference strings, so each goes in whole as the title; add authors and year from the original before you cite it.

@misc{sociolens_ro_ld03_02_v1,
  author = {{SocioLens project}},
  title = {Where does the time saved with generative AI go?},
  year = {2026},
  date = {2026-10-15},
  howpublished = {\url{https://sociolens-web.pages.dev/opportunities/ro-ld03-02/v1/}},
  note = {Research opportunity RO-LD03-02, version v1}
}

@misc{ro_ld03_02_nearest_1,
  title = {Suh \& Oh, "Generative AI and the Reallocation of Time: Productivity, Leisure, and Fulfilling Work," arXiv:2602.12695 (2026)},
  url = {https://arxiv.org/abs/2602.12695},
  note = {Reference as recorded by SocioLens; authors, title and year are not separated; Judged from an abstract, summary or search listing only}
}

@misc{ro_ld03_02_nearest_2,
  title = {Humlum \& Vestergaard, "Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI," NBER Working Paper 33777 (revised March 2026)},
  url = {https://www.nber.org/papers/w33777},
  note = {Reference as recorded by SocioLens; authors, title and year are not separated}
}

@misc{ro_ld03_02_nearest_3,
  title = {Dillon, Jaffe, Peng \& Cambon, "Early Impacts of M365 Copilot," arXiv:2504.11443 (2025)},
  url = {https://arxiv.org/abs/2504.11443},
  note = {Reference as recorded by SocioLens; authors, title and year are not separated; Judged from an abstract, summary or search listing only}
}

@misc{ro_ld03_02_nearest_4,
  title = {Dillon, Jaffe, Immorlica \& Stanton, "Shifting Work Patterns with Generative AI," arXiv:2504.11436 (v4) / NBER Working Paper 33795 (2025)},
  url = {https://arxiv.org/abs/2504.11436},
  note = {Reference as recorded by SocioLens; authors, title and year are not separated}
}

@misc{ro_ld03_02_nearest_5,
  title = {Ranganathan \& Ye, "AI Doesn't Reduce Work—It Intensifies It," Harvard Business Review (February 2026)},
  url = {https://newsroom.haas.berkeley.edu/ai-promised-to-free-up-workers-time-uc-berkeley-haas-researchers-found-the-opposite/},
  note = {Reference as recorded by SocioLens; authors, title and year are not separated; Judged from an abstract, summary or search listing only}
}

@misc{src_sl_bick_blandin_deming_stlfed_wp_2024_027f,
  title = {Bick, Blandin \& Deming, “The Rapid Adoption of Generative AI,” Federal Reserve Bank of St. Louis Working Paper 2024-027F, October 2025},
  url = {https://doi.org/10.20955/wp.2024.027},
  note = {Reference as recorded by SocioLens; authors, title and year are not separated}
}

@misc{src_sl_dillon_jaffe_immorlica_stanton_arxiv_2504_11436_v4,
  title = {Dillon, Jaffe, Immorlica \& Stanton, “Shifting Work Patterns with Generative AI,” NBER Working Paper 33795, revised November 2025},
  url = {https://doi.org/10.3386/w33795},
  note = {Reference as recorded by SocioLens; authors, title and year are not separated}
}

@misc{src_sl_cunningham_whitfill_metr_2026_05_08,
  title = {Cunningham \& Whitfill, “Task Substitution and Uplift,” METR, 8 May 2026},
  url = {https://metr.org/blog/2026-05-08-task-substitution-and-uplift/},
  note = {Reference as recorded by SocioLens; authors, title and year are not separated}
}

@misc{src_sl_becker_metr_self_report_survey_2026_05_11,
  title = {Becker, “Measuring the Self-Reported Impact of Early-2026 AI on Technical Worker Productivity,” METR, 11 May 2026},
  url = {https://metr.org/blog/2026-05-11-ai-usage-survey/},
  note = {Reference as recorded by SocioLens; authors, title and year are not separated}
}