Zak FentonWorking notes

· Evidence review

Doomworking: what the evidence supports

The note this review supportsNobody to hand it to

The note asks what an employer should consider when an agent keeps producing output after the person responsible for checking it has finished for the day. The sources below address parts of that question. None of them studied that arrangement directly, and this review does not claim a search that establishes no such study exists.

Where the pattern sits

Blayney’s history records the September 1915 committee and a research programme that defined fatigue as diminished capacity for work C1 C2. That is a history of a scientific category, not an account of why earlier working-time law was passed. The Directive names shift work as a way of organising continuous processes C4 and gives a physiological rationale for its night-work rules C5. Shift work itself carries a measured association with myocardial infarction, a risk ratio of 1.23 from ten studies in a 2012 review C6, which says nothing directly about reviewing agent output.

The on-call literature is dated in the note as it is dated here. A 2004 review judged it under-researched C7; a 2017 review of eight studies found they did not separate nights with calls from nights without C8; a 2025 review of fatigue management included 17 original studies C35.

What the AI-at-work data can and cannot say

In Humlum and Vestergaard’s Danish data, 35 per cent of the new AI-related tasks workers described were classified as quality review, ethics and compliance C18. Among users who reported saving time, 7.2 per cent selected more breaks and 5.8 per cent more leisure as expected uses C19. The registers show precise null effects on earnings and recorded hours in that setting C20. None of these observes unpaid evening work or recovery. Lee and colleagues’ survey of 319 knowledge workers reports effort shifting from gathering information to verifying it C33, as experience rather than measured time. METR’s trial found developers slower with AI tools while believing themselves faster C23, and the technostress literature carries a caution about overlapping measures C24. Brynjolfsson, Li and Raymond used turnover as a broad proxy because they could not observe stress or satisfaction C22. The meta-analytic figure of -0.23 for human-AI combinations is measured against the better of the two alone C21.

What the automation literature says

Bainbridge wrote that an operator will not monitor automatics effectively once they have run acceptably for a long period C16; Parasuraman and Manzey describe complacency arising when manual tasks compete for attention C17. Her other observations, about automating the easy parts C13, meta-level monitoring in real-time control C14 and the difficulty of time pressure C15, come from process control. None establishes the effect of reviewing an agent’s overnight output.

What the recovery literature says

In one cross-sectional sample, the experience of being on call, especially unpredictability, was associated with fatigue and work-home interference more than exposure was C9. Detachment correlates -0.36 with burnout exhaustion C10, and shows no significant average association with physiological indicators and negative associations with contextual performance and creativity C11. These are associations. The Recovery Experience Questionnaire has a 16-item, four-dimension structure C12; that establishes the instrument, not its permissions or its fit to a new use.

The law

Article 2(1) provides the EU definition C3; Great Britain applies the Working Time Regulations. The classification of an entire standby period is separate from work actually done during it C25, and a worker’s own choices are excluded only from the whole-period classification C26. Employers cannot set standby so long or frequent as to risk health, whatever the classification C27. The UK guidance judgment concerns effective rest rights and acknowledges limits to the employer’s responsibility C28. Ontario’s provisions define disconnecting but create no right C29. The platform work directive addresses automated systems that pressure platform workers C30, within its own scope C31. HSE’s Demands standard includes workload and work patterns C32; the absence of particular words from a 2019 workbook says nothing about whether its concepts apply.

What cannot be evidenced here

The opening question, whether recovery differs when work continues in the worker’s absence, is not tested by any study in this ledger. That is a statement about this ledger, not about the literature. One practitioner account corroborates the direction of concern, with its commercial interest declared C34.

Claims ledger

Each claim is worded exactly as it was checked. Assessed claims show the source, passage, limits and date checked. If a source could not be checked, the entry records what was attempted and why. Claims still being checked are marked in review.

  1. C1Supported

    The Health of Munition Workers Committee was appointed in September 1915 to consider questions of industrial fatigue and hours of labour.

    Source
    Blayney, S. (2017). Industrial Fatigue and the Productive Body: the Science of Work in Britain, c. 1900-1918. Social History of Medicine, 32(2), 310-328. Open source (full text read)
    Passage
    In September 1915, Lloyd George appointed the Health of Munition Workers Committee (HMWC), to ‘consider and advise on questions of industrial fatigue, hours of labour
    Where
    section 3
    Limits
    A history of a scientific category. It does not establish when fatigue was first considered, nor why earlier working-time protections were introduced. Passage and location confirmed against the White Rose author manuscript in review on 16 September 2026.
    Checked

    Back to text↩ 1

  2. C2Supported

    In that research literature industrial fatigue was defined as diminished capacity for work and measured through declining work performance.

    Source
    Blayney, S. (2017). Industrial Fatigue and the Productive Body: the Science of Work in Britain, c. 1900-1918. Social History of Medicine, 32(2), 310-328. Open source (full text read)
    Passage
    It was defined as the diminished capacity for work, and, through analysis of empirical or statistical data relating to factory work, it could be measured in terms of declining work performance.
    Where
    section 1
    Limits
    Describes how researchers defined fatigue, not how employers, courts or earlier legislation used it.
    Checked

    Back to text↩ 1

  3. C3Supported

    Article 2(1) defines working time through the worker doing work, being at the employer's disposal and carrying out their activity or duties, in accordance with national law or practice.

    Source
    Directive 2003/88/EC concerning certain aspects of the organisation of working time, Article 2(1). Open source (full text read)
    Passage
    any period during which the worker is working, at the employer's disposal and carrying out his activity or duties
    Where
    Article 2(1)
    Limits
    The definition does not classify an AI notification or establish equivalence with a colleague. For Great Britain, apply the Working Time Regulations 1998; see the note's UK sources.
    Checked

    Back to text↩ 1

  4. C4Supported

    Article 2(5) defines shift work as workers succeeding each other at the same work stations according to a pattern, which may be continuous or discontinuous.

    Source
    Directive 2003/88/EC, Article 2(5). Open source (full text read)
    Passage
    "shift work" means any method of organising work in shifts whereby workers succeed each other at the same work stations according to a certain pattern, including a rotating pattern, and which may be continuous or discontinuous
    Where
    Article 2(5)
    Limits
    A definition. It neither requires around-the-clock operation nor classifies an individual who reviews an agent's output.
    Checked

    Back to text↩ 1

  5. C5Supported

    Recital 7 of the Directive states that research has shown the human body to be more sensitive at night to environmental disturbances and to certain burdensome forms of work organisation.

    Source
    Directive 2003/88/EC, recital 7. Open source (full text read)
    Passage
    Research has shown that the human body is more sensitive at night to environmental disturbances and also to certain burdensome forms of work organisation
    Where
    recital 7
    Limits
    A recital giving the rationale for night-work obligations. It says nothing about the history of working-time law or about reviewing agent output.
    Checked

    Back to text↩ 1

  6. C6Supported

    In a 2012 systematic review of 34 studies, the pooled estimate from the ten studies recording myocardial infarction associated shift work with a risk ratio of 1.23, 95 per cent confidence interval 1.15 to 1.31.

    Source
    Vyas, M. V., et al. (2012). Shift work and vascular events: systematic review and meta-analysis. BMJ, 345, e4800. Open source (full text read)
    Passage
    Ten studies recorded myocardial infarction, 28 recorded any coronary event, and two recorded ischaemic stroke
    Where
    Results, primary outcomes; risk ratio in the abstract and pooled analyses
    Limits
    Observational studies with residual confounding possible; the review found no association with mortality. Included shift patterns were classified in several ways, not only rotating shifts. It concerns human shift work, not agent supervision. Full text read via Europe PMC on 16 September 2026.
    Checked

    Back to text↩ 1

  7. C7Supported

    A 2004 review stated that on-call work had received significantly less research attention than shift work and overtime.

    Source
    Nicol, A.-M., & Botterill, J. S. (2004). On-call work and health: a review. Environmental Health, 3(1), 15. Open source (full text read)
    Passage
    on-call work has received significantly less research attention than other work patterns such as shift work and overtime hours.
    Where
    abstract
    Limits
    The reviewers' assessment in 2004. It says nothing about the literature since, which includes at least one 2025 systematic review of on-call fatigue management.
    Checked

    Back to text↩ 1

  8. C8Supported

    The eight studies in a 2017 systematic review of on-call work, stress physiology and sleep did not differentiate nights on call from home with calls from nights without.

    Source
    Hall, S. J., et al. (2017). The effect of working on-call on stress physiology and sleep: A systematic review. Sleep Medicine Reviews, 33, 79-87. Open source (full text read)
    Passage
    However, studies did not differentiate between night’s on-call from home with and without calls.
    Where
    abstract
    Limits
    Concerns the eight studies that met inclusion in that review, on sleep and stress physiology. Not a statement about all on-call research or about detachment.
    Checked

    Back to text↩ 1

  9. C9Supported

    In a cross-sectional survey of 157 on-call workers, the experience of being on call, especially stress from unpredictability, was associated with fatigue and work-home interference more than the amount of on-call exposure was.

    Source
    Ziebertz, C. M., et al. (2015). The Relationship of On-Call Work with Fatigue, Work-Home Interference, and Perceived Performance Difficulties. BioMed Research International, 2015, 643413. Open source (full text read)
    Passage
    it is employees’ experience of being on-call, especially the experience of stress due to the unpredictability, rather than the amount of exposure
    Where
    abstract, conclusions
    Limits
    Cross-sectional and self-report. It cannot establish an exclusive causal pathway through unpredictability, nor that exposure is harmless, nor anything about agent output.
    Checked

    Back to text↩ 1

  10. C10Supported

    Across 23 samples and 7,007 workers, psychological detachment correlated -0.36 with burnout exhaustion.

    Source
    Wendsche, J., & Lohmann-Haislah, A. (2017). A Meta-Analysis on Antecedents and Outcomes of Detachment from Work. Frontiers in Psychology, 7, 2072. Open source (full text read)
    Passage
    Burnout (exhaustion) 23 7007 −0.36 −0.42 −0.30
    Where
    Table 1, row for burnout (exhaustion)
    Limits
    Correlational, with heterogeneity, and the primary studies are mostly self-report.
    Checked

    Back to text↩ 1

  11. C11Supported

    Detachment showed no significant average association with physiological stress indicators and significant negative associations with contextual performance and creativity.

    Source
    Wendsche, J., & Lohmann-Haislah, A. (2017). A Meta-Analysis on Antecedents and Outcomes of Detachment from Work. Frontiers in Psychology, 7, 2072. Open source (full text read)
    Passage
    average relationships between detachment and physiological stress indicators and work motivation were not significant while associations with contextual performance and creativity were significant, but negative.
    Where
    abstract
    Limits
    The physiological estimate rests on three samples and 219 people. Associations do not establish that increasing detachment causes a performance loss or identify an optimal amount.
    Checked

    Back to text↩ 1

  12. C12Supported

    The Recovery Experience Questionnaire's final model comprised 16 items across four recovery dimensions.

    Source
    Sonnentag, S., & Fritz, C. (2007). The Recovery Experience Questionnaire: Development and validation of a measure for assessing recuperation and unwinding from work. Journal of Occupational Health Psychology, 12(3), 204-221. Open source (full text read)
    Passage
    The final model comprised 16 items with an acceptable fit
    Where
    Study 2, printed page 212; instructions at page 211
    Limits
    Four dimensions of four items each. Establishes the instrument's structure, not its permissions, a diary version, or validation for any new recall period.
    Checked

    Back to text↩ 1

  13. C13Supported

    Bainbridge argued that automating the easy parts of a task can make the difficult parts of the operator's task more difficult.

    Source
    Bainbridge, L. (1983). Ironies of automation. Automatica, 19(6), 775-779. Open source (full text read)
    Passage
    By taking away the easy parts of his task, automation can make the difficult parts of the human operator's task more difficult.
    Where
    section 3, printed page 777
    Limits
    A discussion paper about real-time process control, with no data of its own. An argument, not an effect estimate for agent work.
    Checked

    Back to text↩ 1

  14. C14Supported

    In the context of real-time automated decisions that an operator cannot check as they run, Bainbridge argued that the operator can only monitor the computer's decisions at a meta-level, judging whether they are acceptable.

    Source
    Bainbridge, L. (1983). Ironies of automation. Automatica, 19(6), 775-779. Open source (full text read)
    Passage
    One can therefore only expect the operator to monitor the computer's decisions at some meta-level, to decide whether the computer's decisions are 'acceptable'
    Where
    section 1.1.3, printed page 776
    Limits
    Concerns control rooms where the operator cannot follow the computer's reasoning in real time. That context is material to the claim.
    Checked

    Back to text↩ 1

  15. C15Supported

    Bainbridge concluded that humans working without time pressure can be impressive problem solvers, and that the difficulty is that they are less effective under time pressure.

    Source
    Bainbridge, L. (1983). Ironies of automation. Automatica, 19(6), 775-779. Open source (full text read)
    Passage
    humans working without time-pressure can be impressive problem solvers. The difficulty remains that they are less effective when under time pressure.
    Where
    section 4, printed page 779
    Limits
    Her own conclusion, without supporting data, about human problem solving generally. It does not follow that morning delivery of agent output is a pressure-free case.
    Checked

    Back to text↩ 1

  16. C16Supported

    Bainbridge wrote that an operator will not monitor automatics effectively if they have been operating acceptably for a long period.

    Source
    Bainbridge, L. (1983). Ironies of automation. Automatica, 19(6), 775-779. Open source (full text read)
    Passage
    the operator will not monitor the automatics effectively if they have been operating acceptably for a long period
    Where
    section 1.1.3, printed page 776
    Limits
    An argument about process-control monitoring, made without the later vocabulary of complacency or automation bias. It does not establish the effect of reviewing an agent's overnight output. Passage copied from the IFAC-hosted PDF on 16 September 2026.
    Checked

    Back to text↩ 1

  17. C17Supported

    Parasuraman and Manzey's review describes automation complacency as occurring under multiple-task load, when manual tasks compete with the automated task for the operator's attention.

    Source
    Parasuraman, R., & Manzey, D. H. (2010). Complacency and Bias in Human Use of Automation: An Attentional Integration. Human Factors, 52(3), 381-410. Open source (abstract only)
    Passage
    Automation complacency occurs under conditions of multiple-task load, when manual tasks compete with the automated task for the operator’s attention.
    Where
    abstract, results
    Limits
    Abstract only. It does not show that monitoring alone is safe, nor that any particular team arrangement necessarily produces the effect.
    Checked

    Back to text↩ 1

  18. C18Supported

    Humlum and Vestergaard classified workers' descriptions of new AI-related tasks, and 35 per cent of that new AI-related work fell into AI quality review and AI ethics and compliance.

    Source
    Humlum, A., & Vestergaard, E. (2025, revised March 2026). Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI. NBER Working Paper 33777. Open source (full text read)
    Passage
    35% of new AI-related work is dedicated to AI Quality Review and AI Ethics & Compliance, suggesting that AI adoption shifts workers not only toward content generation but also toward supervisory roles overseeing AI outputs
    Where
    section 2.3, Figure 4 and Appendix B.3, March 2026 revision
    Limits
    A working paper, not peer reviewed. The figure is the authors' classification of self-described new tasks, not a share of working time.
    Checked

    Back to text↩ 1

  19. C19Supported

    Among chatbot users who reported saving time, 7.2 per cent selected more breaks and 5.8 per cent selected more leisure as expected uses of the savings, and the authors summarise this as fewer than 10 per cent.

    Source
    Humlum, A., & Vestergaard, E. (2025, revised March 2026). Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI. NBER Working Paper 33777. Open source (full text read)
    Passage
    By contrast, fewer than 10% report taking additional breaks or leisure, and about 30% spend more time on the same tasks they initially saved time on.
    Where
    section 2.3, footnote 9, printed page 14; numerical responses in Table E.3, row All, columns 9 and 10 and its sample note; expectation wording in Appendix H, question 15
    Limits
    Responses to a survey question about expected use of savings, not percentages of saved minutes. The categories are not exclusive shares of time. It does not show where the remaining time went.
    Checked

    Back to text↩ 1

  20. C20Supported

    The study estimates precise null effects of AI chatbot adoption on earnings and recorded hours in Danish registers, ruling out effects larger than 2 per cent two years after the launch of ChatGPT.

    Source
    Humlum, A., & Vestergaard, E. (2025, revised March 2026). Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI. NBER Working Paper 33777. Open source (full text read)
    Passage
    we estimate precise null effects on earnings and recorded hours at both the worker and workplace levels, ruling out effects larger than 2% two years after the launch of ChatGPT.
    Where
    abstract, March 2026 revision
    Limits
    Recorded hours are the hours held in Danish administrative registers. The registers do not observe unpaid work or recovery, and the result concerns this setting, not every account of longer hours.
    Checked

    Back to text↩ 1

  21. C21Supported

    In Vaccaro, Almaatouq and Malone's meta-analysis, human-AI combinations performed worse than the best of humans or AI alone, Hedges' g -0.23, 95 per cent confidence interval -0.39 to -0.07.

    Source
    Vaccaro, M., Almaatouq, A., & Malone, T. (2024). When combinations of humans and AI are useful: a systematic review and meta-analysis. Nature Human Behaviour, 8(12), 2293-2303. Open source (full text read)
    Passage
    human–AI combinations performed significantly worse than the best of humans or AI alone (Hedges’ g = −0.23; 95% confidence interval, −0.39 to −0.07)
    Where
    abstract
    Limits
    Against humans alone the same combinations gained. The corpus was searched to 30 June 2023 and measures task performance, not worker wellbeing. It does not experimentally establish an allocation policy.
    Checked

    Back to text↩ 1

  22. C22Supported

    Brynjolfsson, Li and Raymond analysed turnover patterns as a broad measure of how workers responded to AI implementation, because they could not directly observe factors such as stress and job satisfaction.

    Source
    Brynjolfsson, E., Li, D., & Raymond, L. (2025). Generative AI at Work. Quarterly Journal of Economics, 140(2), 889-942. Open source (full text read)
    Passage
    Although we cannot directly observe all these factors, we can analyze turnover patterns as a broad measure of how workers respond to AI implementation.
    Where
    section VII.C, printed page 934, author-hosted copy of the published version
    Limits
    One study of customer-support agents. Turnover is a coarse proxy that can move for reasons unrelated to wellbeing. Its measurement limits do not stand for the whole AI-at-work literature.
    Checked

    Back to text↩ 1

  23. C23Supported

    In METR's early-2025 randomised trial with 16 experienced developers on 246 tasks, allowing AI tools increased completion time by 19 per cent while developers estimated that AI had reduced it.

    Source
    Becker, J., Rush, N., Barnes, E., & Rein, D. (2025). Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity. arXiv:2507.09089. Open source (full text read)
    Passage
    Surprisingly, we find that allowing AI actually increases completion time by 19%--AI tooling slowed developers down.
    Where
    abstract
    Limits
    Early-2025 tools, mature repositories the developers knew well. A preprint. METR's February 2026 follow-up, discussed in a separate note, reports different estimates with wide intervals. It does not show that self-reports of verification load are universally unreliable.
    Checked

    Back to text↩ 1

  24. C24Supported

    A 2021 review of 21 technostress studies cautioned that studies may be subject to considerable conceptual overlap between exposure and outcome measures.

    Source
    Borle, P., Reichel, K., Niebuhr, F., & Voelter-Mahlknecht, S. (2021). How Are Techno-Stressors Associated with Mental Health and Work Outcomes? International Journal of Environmental Research and Public Health, 18(16), 8673. Open source (abstract only)
    Passage
    studies may be subject to considerable conceptual overlap between exposure and outcome measures.
    Where
    abstract
    Limits
    A methodological caution covering studies to June 2020 and no generative AI. Abstract depth: the publisher's full text returned 403 in review on 16 September 2026, and no earlier full-text read is recorded.
    Checked

    Back to text↩ 1

  25. C25Supported

    Where a standby period does not qualify in full as working time, time spent actually working during it still counts as working time.

    Source
    C-344/19, D.J. v Radiotelevizija Slovenija, 9 March 2021, paragraphs 37 to 40, especially 38. Open source (full text read)
    Passage
    only the time linked to the provision of work actually carried out during that period constitutes ‘working time’
    Where
    paragraph 38
    Limits
    The whole-period test concerns imposed constraints on free time. This later EU judgment is not binding on UK courts and does not decide an AI-notification dispute.
    Checked

    Back to text↩ 1

  26. C26Supported

    In classifying a whole standby period, constraints that are the consequence of natural factors or of the worker's own free choice may not be taken into account.

    Source
    C-344/19, D.J. v Radiotelevizija Slovenija, 9 March 2021, paragraph 40. Open source (full text read)
    Passage
    the consequence of natural factors or of his or her own free choice, may not be taken into account.
    Where
    paragraph 40
    Limits
    Concerns the constraints used in classifying the whole standby period. It does not convert work actually done into rest, and it does not decide who imposed a constraint in any particular arrangement.
    Checked

    Back to text↩ 1

  27. C27Supported

    Employers cannot establish standby periods so long or so frequent that they constitute a risk to workers' safety or health, irrespective of whether those periods are classified as rest.

    Source
    C-344/19, D.J. v Radiotelevizija Slovenija, 9 March 2021, paragraph 65. Open source (full text read)
    Passage
    employers cannot establish periods of stand-by time that are so long or so frequent that they constitute a risk to the safety or health of workers, irrespective of those periods being classified as ‘rest periods’
    Where
    paragraph 65
    Limits
    Concerns employer-established standby. No case search has been done on its application to automated output, so nothing is claimed about that.
    Checked

    Back to text↩ 1

  28. C28Supported

    Guidance telling employers they need not ensure workers can actually take rest undermines the rest rights.

    Source
    Case C-484/04, Commission v United Kingdom, at [43] to [44]. Open source (full text read)
    Passage
    clearly liable to render the rights enshrined in Articles 3 and 5 of that directive meaningless
    Where
    paragraph 44
    Limits
    The case concerned UK guidance on daily and weekly rest. Paragraph 43 recognises limits to the employer's responsibility, including that employers generally need not force workers to claim rest. It did not decide an AI-notification dispute.
    Checked

    Back to text↩ 1

  29. C29Supported

    Ontario's written-policy provisions define disconnecting from work to include not sending or reviewing work-related messages, and the ministry's manual states that the provision does not create a right to disconnect.

    Source
    Ontario Ministry of Labour, Immigration, Training and Skills Development. Employment Standards Act Policy and Interpretation Manual, Part VII.0.1, sections 21.1.1 and 21.1.2. Open source (full text read)
    Passage
    This provision does not create a “right to disconnect from work” for employees.
    Where
    section 21.1.2; the definition at section 21.1.1
    Limits
    These written-policy provisions do not themselves create a right to disconnect. They do not decide the note's AI-notification scenario.
    Checked

    Back to text↩ 1

  30. C30Supported

    Article 12(3) prohibits digital labour platforms from using automated monitoring or decision systems in ways that put undue pressure on platform workers or endanger their safety or physical or mental health.

    Source
    Directive (EU) 2024/2831 on improving working conditions in platform work, Articles 2(1)(d), 12(3) and 29(1). Open source (full text read)
    Passage
    puts undue pressure on platform workers or otherwise puts at risk the safety and physical and mental health of platform workers
    Where
    Article 12(3)
    Limits
    Platform workers are defined through an employment relationship under Article 2. The transposition deadline is 2 December 2026; national implementation needs checking before asserting present enforceability. This is not a general UK duty for every employer using AI.
    Checked

    Back to text↩ 1

  31. C31Supported

    The Directive applies to digital labour platforms organising platform work performed in the Union.

    Source
    Directive (EU) 2024/2831, Article 1(3). Open source (full text read)
    Passage
    This Directive applies to digital labour platforms organising platform work performed in the Union, irrespective of their place of establishment or of the law otherwise applicable.
    Where
    Article 1(3)
    Limits
    Fixes the scope of this Directive only. It does not rule out protection under other instruments.
    Checked

    Back to text↩ 1

  32. C32Supported

    HSE's Management Standards define Demands as including workload, work patterns and the work environment.

    Source
    Health and Safety Executive (2019). Tackling work-related stress using the Management Standards approach: A step-by-step workbook (WBK01). Open source (full text read)
    Passage
    Demands Includes issues such as workload, work patterns and the work environment.
    Where
    the six Management Standards
    Limits
    A definition from one workbook. It is not the indicator questionnaire or the whole of HSE's guidance, and the absence of particular words from it says nothing about whether its concepts apply to machine-set pace.
    Checked

    Back to text↩ 1

  33. C33Supported

    In a survey of 319 knowledge workers, Lee and colleagues report that for knowledge and comprehension tasks, perceived effort shifts from information gathering to information verification when using generative AI.

    Source
    Lee, H.-P., et al. (2025). The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers. Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems. Open source (full text read)
    Passage
    for Knowledge and Comprehension, the effort shifts from information gathering to information verification
    Where
    section 5.2, introduction to the shifts; section 5.2.1; author-hosted PDF
    Limits
    Reported experience from a survey with 936 task examples, not a timed measure of verification work. It does not establish a numerical increase in checking time or a causal effect on critical thinking.
    Checked

    Back to text↩ 1

  34. C34Supported

    SaaStr's founder wrote that its three people work harder than its 20-person team did in 2020, while running more than twenty agents.

    Source
    Lemkin, J. (2026). Dear SaaStr: Should I Run 20+ AI Agents the Way SaaStr Does? SaaStr. Open source (full text read)
    Passage
    The three people at SaaStr right now work harder than our 20-person team did in 2020.
    Where
    The Part Nobody Tells You
    Limits
    Grey literature with a commercial interest: a media and events business, a Replit customer featured in a published Replit case study, promoting its own operating model. One organisation's self-report, with no measured hours, workload, detachment or health outcome and no comparison group.
    Checked

    Back to text↩ 1

  35. C35Supported

    A 2025 review of fatigue management for on-call workers covered 17 original studies of management strategies.

    Source
    Bumpstead, H., et al. (2025). How should we manage fatigue in on-call workers? A review of guidance materials and a systematic review of the evidence-base. Sleep Medicine Reviews, 79, 102012. doi:10.1016/j.smrv.2024.102012. Open source (full text read)
    Passage
    Seventeen studies were included in the final systematic review.
    Where
    section 5.1, page 5; scope in section 4.1 and study characteristics in section 5.2
    Limits
    The review concerns fatigue-management strategies for on-call workers. It does not establish what notification batching does to recovery from agent work.
    Checked

    Back to text↩ 1

Back to the noteNobody to hand it to