Each entry gives the study, how it was done, what it found, what it supports and, as plainly, what it does not. Grades describe what kind of evidence a source is; how to read them is on the main page. Every entry has a permanent link and a copyable citation.
Peer-reviewed#
Cook, R. I., Render, M. and Woods, D. D. (2000). Gaps in the continuity of care and progress on patient safety
BMJ, 320(7237), 791-794
- Method
- Conceptual paper in the education and debate section, developed against two publicly documented US clinical accidents.
- Finding
- Gaps are defined as discontinuities in care, appearing as losses of information or momentum or interruptions in delivery. Most gaps are anticipated and bridged by practitioners, so invisibly that neither outsiders nor insiders recognise the activity as distinct work. Bridging is not elimination: some bridges are frail and easily undone. Accidents occur when conditions overwhelm the mechanisms practitioners use to detect and bridge gaps, from which the authors conclude that efforts to forestall errors by isolating practitioners from the system will misfire. Their worked example is the division of nursing work with less credentialed technicians, which delivers a real economic benefit and restricts the nurse's ability to anticipate gaps.
- What it supports
- That the boundary between steps in a process is a named object with its own failure modes, and that the work of bridging it is invisible to every measure of output. The nursing example transfers directly to delegating work to agents.
- What it does not support
- Anything quantified. There is no sample, no rate and no measurement, and nothing in it concerns automation or AI. It is a framing paper.
Peer-reviewed#
Drew, B. J., Harris, P., Zegre-Hemsey, J. K., Mammone, T., Schindler, D., Salas-Boni, R. et al. (2014). Insights into the Problem of Alarm Fatigue with Physiologic Monitor Devices: A Comprehensive Observational Study of Consecutive Intensive Care Unit Patients
PLoS ONE, 9(10), e110274. DOI 10.1371/journal.pone.0110274, published 22 October 2014
- Method
- Observational study of consecutive adults in five adult intensive care units at the University of California, San Francisco, over the 31 days of March 2013. All monitor data, seven ECG leads, pressure, SpO2 and respiration waveforms, user settings and alarms, were stored for 461 patients. Nurse scientists annotated 12,671 arrhythmia alarms against a defined protocol, with inter-rater agreement of 95 per cent on true against false and a Cohen kappa of 0.86. Funded by GE Healthcare.
- Finding
- 2,558,760 unique alarms occurred in 31 days: 1,154,201 arrhythmia, 612,927 parameter and 791,632 technical. 381,560 were audible, an audible alarm burden of 187 per bed per day. 88.8 per cent of the 12,671 annotated arrhythmia alarms were false positives, and 93 per cent of the 168 true ventricular tachycardia alarms were not sustained long enough to warrant treatment.
- What it supports
- That a safety control firing at this rate trains the person holding it to ignore it, and that the training is rational rather than negligent. It is the best measured case anywhere of the base-rate problem that makes a stop control nominal.
- What it does not support
- Anything about AI. The 88.8 per cent applies only to the 12,671 annotated arrhythmia alarms, NOT to all 2.56 million alarms and not to clinical alarms in general; a page stating that 88.8 per cent of clinical alarms are false has misread it. Single centre, one month, five units, industry funded.
Statutory investigation#
National Transportation Safety Board (2014). Descent Below Visual Glidepath and Impact With Seawall, Asiana Airlines Flight 214, Boeing 777-200ER, HL7742, San Francisco, California, July 6, 2013
NTSB/AAR-14/01, adopted 24 June 2014
- Method
- Statutory accident investigation with access to flight recorders, crew interviews, manufacturer documentation and operator training records. N is one.
- Finding
- The pilot flying selected a mode that caused the autoflight system to climb, then disconnected the autopilot and moved the thrust levers to idle, which put the autothrottle into HOLD, a mode in which it does not control airspeed. The board records that neither the pilot flying, the pilot monitoring, nor the observer noted the change in autothrottle mode to HOLD. Contributing factors in the probable cause include the complexities of the autothrottle and autopilot flight director systems being inadequately described in the manufacturer's documentation and the operator's training, which increased the likelihood of mode error. The board attributes insufficient airspeed monitoring in part to automation reliance, and notes the operator's automation policy emphasised full use of automation and did not encourage manual flight in line operations.
- What it supports
- That a correct automation state change can constitute a failed handoff. Nothing malfunctioned: authority moved from machine to crew and the crew were not told in a way that reached them.
- What it does not support
- Any frequency of mode confusion. One accident, three crew, one aircraft type. It cannot support a rate and is used here as a demonstration of a mechanism.
Peer-reviewed#
Starmer, A. J., Spector, N. D., Srivastava, R., West, D. C., Rosenbluth, G., Allen, A. D. et al. for the I-PASS Study Group (2014). Changes in Medical Errors after Implementation of a Handoff Program
New England Journal of Medicine, 371(19), 1803-1812
- Method
- Prospective systems-based intervention study across nine paediatric residency programmes in the US and Canada, January 2011 to May 2013, in three staggered waves with season-matched six-month pre and post periods. 875 consenting residents, 10,740 patient admissions. Active surveillance five days a week, incidents classified by two blinded physician reviewers. Not randomised, no control group.
- Finding
- The medical-error rate fell 23 per cent, from 24.5 to 18.8 per 100 admissions, and preventable adverse events fell 30 per cent, from 4.7 to 3.3 per 100 admissions, both P<0.001. Near misses and non-harmful errors fell 21 per cent. Non-preventable adverse events did not change, 3.0 against 2.8, P=0.79. Oral handoff duration did not change, 2.4 against 2.5 minutes per patient, P=0.55. Error rates did not change significantly at three of the nine sites, although written and oral handoff processes improved at all nine.
- What it supports
- That a handoff can be made substantially safer without being made longer, by changing what is transferred rather than how much time is spent transferring it. The unchanged non-preventable rate is the strongest internal evidence the effect is real.
- What it does not support
- Causation, by the authors' own statement, and not which element of the bundle did the work. Paediatric inpatient units only; generalisation to other specialties is untested. Reviewer agreement was moderate, kappa 0.47 for error classification. Data collectors could not be blinded to period. Funding included an unrestricted medical education grant from Pfizer.
Compiled review#
The Joint Commission (2017). Sentinel Event Alert 58: Inadequate hand-off communication
The Joint Commission, Issue 58, 12 September 2017
- Method
- Advisory bulletin aggregating third-party findings. No original data, no sample, no denominator.
- Finding
- States that inadequate hand-off communication contributes to adverse events including wrong-site surgery, delay in treatment, falls and medication errors. Reports, from cited third parties, that communication failures were responsible at least in part for 30 per cent of US malpractice claims over five years, 1,744 deaths and 1.7 billion dollars in costs; that a typical teaching hospital may experience more than 4,000 hand-offs a day; and that 69 per cent of clinical learning environments had no standardised hand-off process against 20 per cent with some standardisation. Its only quantified outcome evidence is the I-PASS trial.
- What it supports
- That handoff has been named as a systemic failure point by a national accreditation body, and that the profession's own quantified evidence for it is thinner than the advisory framing implies.
- What it does not support
- The most quoted claim attached to it. The statement that 80 per cent of serious medical errors involve miscommunication during handoff does not appear anywhere in this alert; the full six pages were read on 4 September 2026. The nearest real figure is Starmer and colleagues citing a Joint Commission statistics page for two of every three sentinel events involving communication failures, which is a different denominator and a broader category. Every figure in the alert is footnoted to a document not read here.
Peer-reviewed#
Weber, D. E., MacGregor, S. C., Provan, D. J. and Rae, A. (2018). 'We can stop work, but then nothing gets done.' Factors that support and hinder a workforce to discontinue work for safety
Safety Science, 108, 149-160. Safety Science Innovation Lab, Griffith University
- Method
- Qualitative study. Ten focus groups with workers in a range of roles in the liquefied petroleum gas industry, examining an explicit organisational Authority to Stop an Unsafe Task.
- Finding
- Stopping work for safety was reported as challenging at the sharp operational end despite the authority existing and carrying no formal penalty. The authors conclude that stopping an unsafe task 'does not solely hinge on the willingness of individual workers to stop, but also depends on contextual factors surrounding the stop work decision'. A participant's account supplies the title: the authority exists, using it produces no drama, and then nothing gets done, so the work resumes as before.
- What it supports
- That granting an authority is not the same as making it usable, and that the decisive factor is what happens to the work and the worker afterwards rather than the existence of the permission.
- What it does not support
- Any rate or frequency. Ten focus groups in one industry in one country, qualitative by design, with no measurement of how often stops occurred or should have.
Peer-reviewed#
Dauth, W., Findeisen, S., Suedekum, J. and Woessner, N. (2021). The Adjustment of Labor Markets to Robots
Journal of the European Economic Association, 19(6), 3104-3153
- Method
- German administrative worker and plant data, 1994 to 2014, with a shift-share instrument for robot exposure.
- Finding
- Incumbent workers largely kept their jobs and moved into new, higher-quality tasks within their original plants. The cost fell instead on young labour-market entrants, who shifted away from vocational manufacturing training towards university.
- What it supports
- That the damage from automation falls on skill FORMATION rather than skill possession. Twenty years of German manufacturing data making the missing-rungs argument before anyone applied it to knowledge work.
- What it does not support
- That generative AI will behave like industrial robots. The technologies and the tasks differ substantially.
Peer-reviewed#
Wong, A., Otles, E., Donnelly, J. P. et al. (2021). External Validation of a Widely Implemented Proprietary Sepsis Prediction Model in Hospitalized Patients
JAMA Internal Medicine, 181(8), 1065-1070. DOI 10.1001/jamainternmed.2021.2626
- Method
- Retrospective external validation cohort study. 27,697 patients aged 18 or over across 38,455 hospitalisations at Michigan Medicine, 6 December 2018 to 20 October 2019. Sepsis occurred in 7 per cent of hospitalisations.
- Finding
- The Epic Sepsis Model achieved a hospitalisation-level area under the curve of 0.63 (95% CI 0.62-0.64), against the 0.76-0.83 cited by its developer. At the alerting threshold in clinical use, sensitivity was 33 per cent, specificity 83 per cent, positive predictive value 12 per cent. It did not identify 1,709 of 2,552 septic hospitalisations (67 per cent), 60 per cent of whom received timely antibiotics anyway, while crossing the alert threshold in 18 per cent of all hospitalisations (6,971 of 38,455), requiring eight patients to be evaluated per case of sepsis found.
- What it supports
- That a proprietary clinical prediction model deployed at national scale can perform far below its developer's stated figures when validated independently, and that nobody had checked. The authors' own conclusion is that widespread adoption despite poor performance raises fundamental concerns about sepsis management nationally.
- What it does not support
- That all clinical prediction models fail, or that this model performs identically elsewhere. It is one model at one academic health system, and the authors note the theoretical alert burden does not account for real-world trigger criteria and lockouts.
Peer-reviewed#
Kanazawa, K., Kawaguchi, D., Shigeoka, H. and Watanabe, Y. (2022). AI, Skill, and Productivity: The Case of Taxi Drivers
NBER Working Paper 30612; published in Management Science, 72(2), 1376-1388 (2026)
- Method
- Driver-level data from a Japanese taxi fleet through the rollout of an AI demand-prediction system.
- Finding
- Productivity gains accrued almost entirely to LOW-skilled drivers, narrowing the gap between best and worst by 14 percent.
- What it supports
- That the novice-boost pattern found in customer support and software also appears in manual frontline work.
- What it does not support
- Included deliberately as a disconfirming case for a tidy story. Anyone arguing that AI levels up white-collar workers while degrading frontline ones has to explain this result, which runs the other way.
Peer-reviewed#
Kesavan, S., Lambert, S. J., Williams, J. C. and Pendem, P. K. (2022). Doing Well by Doing Good: Improving Retail Store Performance with Responsible Scheduling Practices at the Gap, Inc.
Management Science, 68(11), 7818-7836
- Method
- Randomised field experiment across 28 Gap stores in San Francisco and Chicago over nine months, November 2015 to August 2016, analysed as intent-to-treat.
- Finding
- Restoring schedule predictability and worker control raised productivity 5.1 percent, with sales up 3.3 percent and labour hours down 1.8 percent.
- What it supports
- That algorithmic optimisation of scheduling is not merely harsh but operationally counterproductive: giving humans back control improved the numbers the algorithm was optimising.
- What it does not support
- Anything directly about generative AI. It concerns algorithmic scheduling, which is an older and different technology.
Peer-reviewed#
Lang, K., Josefsson, V., Larsson, A.-M. et al. (2023). Artificial intelligence-supported screen reading versus standard double reading in the Mammography Screening with Artificial Intelligence trial (MASAI): a clinical safety analysis
The Lancet Oncology, 24(8), 936-944. DOI 10.1016/S1470-2045(23)00298-X
- Method
- Planned interim safety analysis of a randomised, controlled, non-inferiority, single-blinded screening accuracy trial. 80,033 women aged 40-80 screened at four sites in southwest Sweden between April 2021 and July 2022, randomised 1:1 to AI-supported screen reading or standard double reading by two radiologists.
- Finding
- Cancer detection was six per 1,000 screened women with AI support against five per 1,000 with standard double reading, 41 more cancers detected. The false-positive rate was 1.5 per cent in both arms. Screen readings fell from 83,231 in the control arm to 46,345 in the AI arm, a 44 per cent reduction in screen-reading workload.
- What it supports
- That a triage-plus-detection-support workflow with a radiologist retaining the recall decision can hold detection while roughly halving reading volume, in a randomised population-based programme.
- What it does not support
- Patient benefit, which the interim analysis was not designed to test. It is also one mammography device, one AI system, one country, and moderately to highly experienced readers, which the authors state as limits on generalisability.
Peer-reviewed#
Goh, E., Gallo, R., Hom, J. et al. (2024). Large Language Model Influence on Diagnostic Reasoning: A Randomized Clinical Trial
JAMA Network Open, 7(10), e2440969. DOI 10.1001/jamanetworkopen.2024.40969
- Method
- Single-blind randomised clinical trial, 29 November to 29 December 2023. 50 US-licensed physicians (26 attendings, 24 residents) in family medicine, internal medicine or emergency medicine, randomised to GPT-4 or to conventional resources, working through up to six clinical vignettes, graded blind against a validated diagnostic reasoning rubric. 244 cases completed.
- Finding
- Median diagnostic reasoning score per case was 76 per cent (IQR 66-87) with the LLM and 74 per cent (IQR 63-84) with conventional resources: an adjusted difference of 2 percentage points (95% CI -4 to 8, p=0.60). Median time per case was 519 seconds against 565, adjusted difference -82 seconds (95% CI -195 to 31, p=0.20). The LLM alone scored a median 92 per cent, 16 percentage points above the conventional-resources group (95% CI 2-30, p=0.03).
- What it supports
- That adding a capable model to a physician did not, in this trial, improve diagnostic reasoning or save time, while the same model working alone outperformed both groups of clinicians. The performance of a human-plus-model pairing cannot be inferred from the performance of either part.
- What it does not support
- That LLMs should diagnose autonomously, which the authors explicitly reject. Six curated vignettes exclude history-taking, examination, context and time, which is most of clinical reasoning. Participants received no prompt-engineering training, and the authors offer prompting and interaction design as explanations.
Working paper#
Lee, Y. S., Iizuka, T. and Eggleston, K. (2024). Robots and Labor in Nursing Homes
NBER Working Paper 33116
- Method
- Original facility-level panel of Japanese nursing homes, using regional robot subsidies as an instrument for adoption.
- Finding
- Robot adoption RAISED employment and improved retention, most strongly for non-regular staff, reallocated worker effort towards direct care, and improved quality: less use of physical restraint and fewer pressure ulcers.
- What it supports
- That automation can absorb routine physical work and upgrade the human job rather than hollow it. The cleanest counter-case in this evidence base, drawn from care work rather than knowledge work.
- What it does not support
- That this generalises. Japanese long-term care faces acute labour shortage, so robots substituted for vacancies rather than for people, which is a very particular condition.
Peer-reviewed#
Patel, V. R., Liu, M., Worsham, C. M. and Jena, A. B. (2024). Alzheimer's disease mortality among taxi and ambulance drivers: population based cross sectional study
The BMJ, 387, Christmas issue, 17 December 2024. DOI 10.1136/bmj-2024-082194. PROVENANCE: bmj.com could not be reached from either fetcher on 5 September 2026, so the figures below are taken from BMJ Group's own press release for the paper and the Science Media Centre briefing, both read at source that day, rather than from the full text. Re-read and confirm at bmj.com when access allows.
- Method
- Population-based cross-sectional study of US death certificates from the National Vital Statistics System, 1 January 2020 to 31 December 2022, covering 443 occupations and nearly 9 million deaths with occupational information. Usual occupation is the one in which the decedent spent most of their working life. Adjusted for age at death and sociodemographic factors. Published in the BMJ's Christmas issue, which is peer reviewed and deliberately light in subject matter.
- Finding
- 3.9 per cent of deaths (348,328) had Alzheimer's disease listed as a cause. Among 16,658 taxi drivers, 171 did (1.03 per cent); among 1,348 ambulance drivers, 10 did (0.74 per cent). After adjustment, taxi and ambulance drivers had the lowest proportion of any occupation examined (1.03 and 0.91 per cent) against 1.69 per cent for the general population. The pattern did not appear among bus drivers or pilots, who follow predetermined routes, nor for other dementias.
- What it supports
- That an occupational association exists at national scale, and that it is specific to route-generating rather than route-following driving. The authors' own summary is the right one: 'We view these findings not as conclusive, but as hypothesis generating.'
- What it does not support
- That navigation protects anyone. Three limitations do most of the damage, two of them raised by Tara Spires-Jones for the Science Media Centre. The drivers died at around 64 to 67 against 74 for other occupations, and Alzheimer's onset is typically after 65, so some may not have lived long enough to develop it. Women were 10 to 22 per cent of the drivers against 48 per cent elsewhere, on a disease women are more likely to develop. And no brain imaging was involved, so the hippocampal mechanism is hypothesis rather than measurement. Robert Howard, reviewing it alongside her, judged it premature to suggest drivers turn off their satnavs to prevent dementia.
Peer-reviewed#
Yu, F., Moehring, A., Banerjee, O., Salz, T., Agarwal, N. and Rajpurkar, P. (2024). Heterogeneity and predictors of the effects of AI assistance on radiologists
Nature Medicine, 30(3), 837-849
- Method
- 140 radiologists, 15 chest X-ray diagnostic tasks, roughly 5,190 observations, randomised AI assistance, with empirical-Bayes shrinkage to separate genuine individual differences from noise.
- Finding
- The effect of AI assistance diverged sharply between radiologists, from strongly positive to strongly negative. Experience, subspecialty and prior familiarity with AI all failed to predict who would benefit, and lower performers did not consistently gain.
- What it supports
- That the effect of AI assistance on expert performance is individual and currently unpredictable, so a policy of giving everyone the tool will help some professionals and harm others with no way to tell in advance which.
- What it does not support
- That AI assistance is bad on average, or that the pattern holds outside diagnostic imaging.
Peer-reviewed#
Budzyn, K., Romanczyk, M., Kitala, D. and 18 others (2025). Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy: a multicentre, observational study
The Lancet Gastroenterology and Hepatology, 10(10), 896-903. DOI 10.1016/S2468-1253(25)00133-5. A CORRECTION EXISTS and was added to this entry on 23 September 2026: Correction to Lancet Gastroenterol Hepatol 2025; 10: 896-903, volume 10, issue 11, page e12, DOI 10.1016/s2468-1253(25)00294-8, PMID 40946709, confirmed in the Crossref record today. ITS CONTENT HAS NOT BEEN READ: thelancet.com is gated and PubMed returns a CAPTCHA to both fetchers, so nobody here knows what it changes. Recorded in the pattern of the Bastani correction of 12 September 2026, so that a reader who finds a correction notice does not assume the paper was retracted, and so that nobody here assumes it was left unamended. Somebody with journal access should read it and report back.
- Method
- Retrospective observational study nested in the ACCEPT trial, four Polish endoscopy centres. 1,443 colonoscopies performed WITHOUT AI assistance (795 before and 648 after AI was introduced) by 19 endoscopists averaging 27.6 years of experience, ranging from 8 to 39 years.
- Finding
- Adenoma detection rate in unassisted colonoscopy fell from 28.4 percent before AI exposure to 22.4 percent after, a drop of 6.0 percentage points (p=0.0089; adjusted odds ratio 0.69).
- What it supports
- Measured deskilling in highly experienced professionals, in unassisted performance, within months of routine AI exposure. The strongest direct evidence that capability degrades when a tool takes over the judgement, rather than merely a plausible mechanism.
- What it does not support
- Causation with certainty: it is observational, not randomised, and other changes over the period cannot be fully excluded. It is also one procedure in one country, and detection rate is a proxy for skill rather than skill itself.
Peer-reviewed#
Heinz, M. V., Mackin, D. M., Trudeau, B. M. et al. (2025). Randomized Trial of a Generative AI Chatbot for Mental Health Treatment
NEJM AI, 2(4). DOI 10.1056/AIoa2400802
- Method
- Randomised controlled trial, N=210 adults recruited by national advertising, with major depressive disorder, generalised anxiety disorder or clinically high risk for feeding and eating disorders. Four-week intervention with Therabot, an expert-fine-tuned generative chatbot built on a hand-curated CBT corpus, against a WAITLIST control, with four weeks of follow-up.
- Finding
- Statistically significant symptom reductions across all three diagnostic groups at four and eight weeks. 95 per cent of participants engaged, averaging 260 messages and 6.18 hours over four weeks. Therapeutic alliance was comparable to outpatient psychotherapy (mean WAI 3.59). Expressions of suicidal ideation required staff intervention 15 times, and inappropriate responses required correction 13 times.
- What it supports
- That a purpose-built, expert-curated generative chatbot can produce measurable symptom improvement over four weeks, and that users form a working alliance with it.
- What it does not support
- That it is safe unsupervised, or that the effect is the chatbot rather than attention and expectation. The control was a waitlist rather than an active comparator, so the treatment effect cannot be separated from the effect of receiving something. The developer confirmed to the FDA advisory committee that humans reviewed all messages in near real time and conducted risk assessments where needed, so the safety record describes a supervised system rather than an autonomous one.
Peer-reviewed#
Hennum Nilsson, K., Bodin, T., Strauss, P., Matilla-Santander, N., Badarin, K., Brulin, E. and Hakansta, C. (2025). Algorithmic management is associated with psychological distress, musculoskeletal pain, and occupational accidents: a cross-sectional study in logistics
International Archives of Occupational and Environmental Health, 98
- Method
- Online survey of Swedish logistics workers, recruitment and data collection February to July 2024, 978 respondents (592 drivers, 378 warehouse), using an eleven-item algorithmic-management exposure scale covering task allocation, surveillance and performance monitoring, with models adjusted for age, gender, country of birth, education, income, company size, tenure, employment type, contracted hours and union membership. Participants were recruited through paid social-media campaigns on Facebook and Instagram, with two trade unions also circulating the study to members, so this is a self-selected convenience sample rather than a representative one. AUTHOR LIST CORRECTED 23 September 2026: this entry previously read Nilsson, A. et al. The first author is Karin Hennum Nilsson and there are seven authors. PMC metadata lists only the corresponding author, Carin Hakansta, the trap already recorded for the Nature Medicine triad on 5 September 2026.
- Finding
- Higher exposure to algorithmic management was associated with psychological distress at an adjusted prevalence ratio of 2.12 (95% CI 1.49 to 3.02), occupational accidents at 1.92 (1.22 to 3.01), headaches at 1.68 (1.09 to 2.58) and musculoskeletal pain at 1.54 (1.23 to 1.92). Stratified analyses gave stronger associations for drivers, particularly on psychological distress, headaches and sleep disturbance; warehouse workers were less consistent. All four figures read at source on 23 September 2026.
- What it supports
- That where AI meets frontline work the output is measured in bodies rather than in output quality, a category almost entirely absent from white-collar productivity research.
- What it does not support
- Causation: it is cross-sectional and self-reported. Workers under strain may also perceive management as more algorithmic.
Peer-reviewed#
Lukac, P. J., Turner, W., Vangala, S., Chin, A. T., Khalili, J., Shih, Y.-C. T., Sarkisian, C., Cheng, E. M. and Mafi, J. N. (2025). Ambient AI Scribes in Clinical Practice: A Randomized Trial
NEJM AI, 2(12). DOI 10.1056/AIoa2501000. Preprint: medRxiv 2025.07.10.25331333
- Method
- Parallel three-arm pragmatic randomised clinical trial at one US health system. 238 outpatient physicians across 14 specialties randomised 1:1:1 by covariate-constrained randomisation to Microsoft DAX, Nabla or usual care, 4 November 2024 to 3 January 2025, with the second intervention month compared to baseline.
- Finding
- Time writing a note fell by an estimated 18 seconds in the control arm, 23 seconds in the DAX arm and 41 seconds in the Nabla arm. Only Nabla differed significantly from control (-9.5 per cent, 95% CI -17.2 to -1.8, p=0.02); DAX did not (-1.7 per cent, 95% CI -9.4 to +5.9, p=0.66). Scribe users improved on Mini-Z burnout (+2.76, p<0.001), task load (-35.8, p=0.01) and work exhaustion (-0.27, p=0.01), with no significant difference between the two products on any psychometric. Roughly 15 per cent of physicians given a tool never used it.
- What it supports
- That ambient documentation produces a small, product-dependent time saving and a more consistent wellbeing effect, measured against a randomised control rather than against a before-and-after.
- What it does not support
- A general time saving for ambient AI. One health system, majority female sample, a two-month contract-limited window, and the authors flag that the electronic record's own time metrics do not count editing done inside the scribe platform, so reported savings may be overstated. They note this limitation affects all studies using those metrics.
Peer-reviewed#
Moore, J., Grabb, D., Agnew, W., Klyman, K., Chancellor, S., Ong, D. C. and Haber, N. (2025). Expressing stigma and inappropriate responses prevents LLMs from safely replacing mental health providers
Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency (FAccT). DOI 10.1145/3715275.3732039
- Method
- Mapping review of therapy guides used by major medical institutions to identify the requirements of a therapeutic relationship, followed by several experiments testing current models, including gpt-4o, against those requirements in naturalistic therapy settings.
- Finding
- Models expressed stigma towards people with mental health conditions and responded inappropriately to common and critical presentations, including encouraging delusional thinking, which the authors attribute to sycophancy. The pattern persisted in larger and newer models.
- What it supports
- That the failures are structural rather than a matter of model size, and that current safety training does not remove them.
- What it does not support
- How often this occurs in real use, or what happens with purpose-built clinical systems rather than general-purpose models. These are constructed scenarios, not observed patient interactions.
Institutional survey#
Boston Consulting Group; Vinciane Beauchene, Sylvain Duranton, David Martin, Vanessa Lyon, Jeff Walters (2026). AI at Work: Why Strategy Matters More Than Tools
BCG, June 2026 (fourth edition)
- Method
- Survey of approximately 12,000 frontline employees, managers and leaders in more than a dozen markets; fieldwork dates are not stated on the publication page. Fourth annual wave, with year-on-year comparisons to 2025.
- Finding
- 74% of frontline employees describe themselves as AI users, up 23 percentage points from 2025. Among frontline regular users, 42% report saving at least a workday per week, but 66% receive limited or no guidance on what to do with the saved time. 30% say their organisation has integrated AI agents into workflows, against 13% a year earlier, and more than six in ten believe agents could do at least half of their job within three years. 72% say the skills expected of them have shifted and 36% feel they have received adequate upskilling. Organisations using AI to reshape workflows end to end or invent new business models rose to 42% from 22%. Employees with a clear strategy but limited tool access report better outcomes than those with strong access but no direction.
- What it supports
- That frontline self-reported use has caught up with managers, and that direction and training lag access to tools.
- What it does not support
- Time saved and outcomes are self-reported and unverified; a 'workday per week' is a perception. BCG sells AI strategy work. Sampling frame and country weighting are not published on the page.
Qualitative field study#
Ehsan, U., Passi, S., Saha, K., McNutt, T., Riedl, M.O. and Alcorn, S. (2026). From Future of Work to Future of Workers: Addressing Asymptomatic AI Harms to Foster Dignified Human-AI Interaction
Proceedings of the CHI Conference on Human Factors in Computing Systems (CHI '26), ACM. DOI 10.1145/3772318.3791081. Preprint arXiv:2601.21920, 29 January 2026
- Method
- Twelve months of situated fieldwork through the first year of routine use of a commercial AI-assisted radiotherapy treatment-planning system across a five-site North American hospital group. 42 participants: 15 radiation oncologists, 12 medical physicists, 7 dosimetrists, 8 administrators, with 2 to 24 years of experience. 52 think-aloud sessions, 24 interviews staged across months 2 to 11, five participatory workshops, 63 hours transcribed, analysed by grounded theory.
- Finding
- Measured operational gains and reported capability loss ran together. Planning cycles shortened by roughly 15 per cent and confidence rose, while by month nine several dosimetrists said their unaided proficiency had worsened over the year, one saying they had grown slower without the tool. The authors name two things the research has nowhere else: intuition rust, the gradual dulling of expert judgement beneath intact output, and identity commoditisation, the erosion of professional dignity as practitioners describe becoming AI babysitters, button-pushers and bystanders in their own practice. They call these harms asymptomatic because the organisation's own measures showed only the improvement.
- What it supports
- That the instruments an organisation uses to judge an AI deployment can register the gain and be structurally blind to the cost, and that practitioners perceive the loss long before any dashboard does. It is the best available account of what deskilling feels like from inside a high-stakes clinical specialty, and the only source here on occupational identity.
- What it does not support
- Any measured deskilling. The skill claims are self-reported, plus one informal unaided exercise with a couple of dosimetrists during a workshop. There is no controlled comparison, no pre and post measurement and no quantified skill outcome, so it cannot be set beside Budzyn as a second measured result. IMPORTANT, this study is widely misdescribed in secondary summaries as radiologists using generative AI. It is neither. The participants plan radiotherapy treatment rather than interpret images, and the system is optimisation-based, not generative. Do not repeat that description.
Peer-reviewed#
Gommers, J., Lang, K., Hofvind, S. et al. (2026). Interval cancer, sensitivity, and specificity comparing AI-supported mammography screening with standard double reading without AI in the MASAI study
The Lancet. DOI 10.1016/S0140-6736(25)02464-X. Published 29 January 2026
- Method
- Full results of a randomised, controlled, non-inferiority, single-blinded, population-based screening-accuracy trial. Over 100,000 women screened at four Swedish sites between April 2021 and December 2022, with two years of follow-up.
- Finding
- Interval cancers fell from 1.76 per 1,000 women (93/52,872) in the control arm to 1.55 per 1,000 (82/53,043) in the AI arm, a 12 per cent reduction. There were 16 per cent fewer invasive (75 v 89), 21 per cent fewer large (38 v 48) and 27 per cent fewer aggressive-subtype (43 v 59) interval cancers. Cancers detected at screening rose from 74 per cent (262/355) to 81 per cent (338/420) of all cases. False positives were 1.5 per cent in the intervention arm and 1.4 per cent in the control arm.
- What it supports
- That AI-supported screen reading, with a radiologist retaining the recall decision, reduced interval cancers over two years in a randomised population-based programme. The only outcome-level randomised evidence of this kind in clinical AI.
- What it does not support
- Generalisation beyond one country, one mammography device, one AI system and experienced readers, all stated as limitations by the authors. Mortality was not an endpoint, and cost-effectiveness was not assessed. The first author's own summary states the design does not support replacing radiologists, since at least one still reads every case.
Working paper#
Hosseini Maasoum, S. M. and Lichtinger, G. (2026). Generative AI as Seniority-Biased Technological Change: Evidence from U.S. Resume and Job Posting Data
SSRN working paper, DOI 10.2139/ssrn.5425555. Harvard University. First version 31 August 2025; version read is dated 25 May 2026 on its own masthead, obtained from the author's site. SSRN records the current revision as 6 June 2026, 109 pages. NOT peer-reviewed, no journal or NBER placement. Main text read in full at source 9 September 2026; the supplemental appendix, A.1 to A.23, was NOT read
- Method
- Revelio Labs resume data via WRDS, LinkedIn-derived, merged with Revelio job postings. 281,111 firms, 161,192,969 positions from 2015, roughly 66 million unique workers, and 198,773,384 postings from 2021. Firms included if they recorded at least 20 new positions between January 2021 and March 2025; HR intermediaries and the roughly 800 largest firms excluded. Juniors are Revelio's Entry and Junior seniority levels, seniors Associate and above. Adoption is identified from job postings for people hired to integrate generative AI into the firm, by keyword flag then LLM classification: 131,845 of 198.8 million postings, 0.066 per cent, giving 10,433 adopter firms, 3.71 per cent of firms and 16 per cent of employment. Four designs against non-adopting controls: difference-in-differences, triple-difference with firm-by-time and industry-by-seniority-by-time fixed effects, a triple-difference by occupational exposure, and a staggered event study using Callaway and Sant'Anna. Errors clustered at firm level. A separate task-level analysis maps 355,013 postings from 1,000 firms onto O*NET tasks and tests how task bundles change.
- Finding
- Junior employment at adopting firms fell about 9 per cent relative to non-adopters six quarters after diffusion, and 8 per cent eight quarters after adoption in the staggered design, while senior employment showed no comparable break. The fall concentrates in exposed occupations: high-exposure junior employment contracts roughly 7 log points against low-exposure between 2022Q4 and 2025Q1, while the senior coefficient continues to rise. The decomposition attributes the fall to hiring, not exits. Junior hiring fell by 4.006 per firm-quarter (SE 0.221) against a pre-period mean of 5.059, which the authors describe as about an 80 per cent reduction. Junior separations ALSO fell, by 1.089 (SE 0.163), roughly a quarter of the hiring decline. Promotion rates rose slightly, by 0.033 percentage points (SE 0.014). At task level, a one standard deviation increase in a task's exposure is associated with roughly a four percentage point larger contraction in junior task bundles at adopters, with the senior interaction running the other way.
- What it supports
- The first firm-level, within-firm measurement of what happens to junior employment when a firm adopts generative AI, at a scale no other study has. Pre-trends are flat back to 2015, covering an earlier tightening cycle. A postings placebo points to falling demand rather than a labour supply shock. The decisive contribution is the mechanism: this is a hiring story, not a redundancy story. Separations fell. Firms stopped opening the door rather than pushing people out of it, which is why the effect is invisible in unemployment figures and visible only to people who cannot get in. The task-level test shows exposed tasks being removed from junior job descriptions specifically, which is the missing-rungs claim measured directly rather than inferred.
- What it does not support
- Causation. The authors write throughout that adoption IS ASSOCIATED WITH the decline, call the evidence suggestive, and state in their conclusion that unobserved confounders may remain and that the window, 2023 to 2025, is short. It is an unrefereed working paper: no journal, no NBER number, and the headline employment estimates are read off event-study figures with no printed standard errors, so only the flows and task tables carry reportable precision. The data are LinkedIn profiles, which tilt towards managerial, professional and high-information work; the authors now show that tilt is stable across the period and absorbed by fixed effects, which protects the comparison but does not make the sample representative of the labour force. The adoption measure misses informal use inside firms, which the authors say should bias estimates towards zero. They cannot test the mechanism they propose, which is that firms cut junior hiring in anticipation rather than because tasks were already automated. And they flag the missing intercept problem themselves: this does not aggregate to an economy-wide claim without further assumptions. Nothing here shows that the juniors not hired were worse off, or that the capability they would have built is gone rather than delayed.
Where this evidence is used
The research that rests most on this section:
- Does AI weaken human judgement?
- How will AI change medicine?
- How do humans and agents divide work across a process?
- What happens to medical and legal training?
- Should we trust AI over human experts?
- How do you design a stop button people will actually use?
- The Missing Rungs
- Will AI replace entry-level jobs?
The other sections
- Judgement, thinking and offloading 79
- Learning, practice and expertise 92
- Human and AI collaboration 44
- Empathy, creativity and what stays human 36
- Work, jobs and the labour market 94
- The seven human capabilities 38
- The institutional record 180
- Professions and sectors 25
- The international picture 29
The whole base is also machine-readable at evidence.json under CC BY 4.0.
Reference · SS-2026-444 · How I use AI here
Cite this page
Hirji, R. (2026). Frontline, clinical and physical work. The SuperSkills evidence base, SS-2026-444. https://thesuperskills.com/research/evidence/frontline-and-clinical-work. Last reviewed 3 November 2026.
An evidence review by Rahim Hirji, not peer-reviewed research. For a material claim, cite the underlying study as well; every study here carries its own permanent link.
How citations and IDs work