What do radiology teams need after AI arrives? A 69-study review finds the evidence is thinnest after adoption
A systematic review found abundant evidence about perceived benefits, trust and short-term task performance, but far less about long-term adaptation, wellbeing, retention or job satisfaction. Its new framework is a research agenda—not proof that AI makes imaging work sustainable.
The short version
- What happened
- The review included 69 peer-reviewed empirical studies selected from seven databases, after 21,685 duplicate records were removed; it mixed perception surveys, qualitative research and AI-assisted reader studies.
- Who it affects
- Imaging professionals need protected learning time, clear override authority and a route to report failures after deployment—not only a launch-day demonstration.
- What remains unproven
- Systematic thematic review rather than a meta-analysis
Keep in mind
Systematic thematic review rather than a meta-analysis
69 heterogeneous empirical studies
English-language Q1/Q2 journal articles only
Search ended September 2025
Single-reviewer extraction and coding with secondary checks
No formal screening or coding agreement statistic
Theme frequencies are not effect sizes
Conceptual framework has not been tested as a causal pathway
Source analysis · 1 linked source. Primary analysed source date: 3 November 2026.
Keep following the evidence
Join The Impact Brief founding list for selected AI reports, their effect on people and the evidence limits. Read a sample first.
Read the full report · 9 min The evidence, methods, implications and limitations
The direct answer: training, control and support matter, but the long-term evidence is missing
Radiology departments cannot treat AI adoption as the finish line. A new systematic review of 69 peer-reviewed studies found that medical-imaging workers' responses repeatedly turned on six issues: perceived benefits, trust, readiness for change, autonomy and control, threats to professional identity, and willingness to use the technology. After a system enters practice, the literature points to training, task performance, mental strain and policy compliance as relevant forms of adaptation. Yet the review's most consequential finding is the imbalance in that evidence: researchers have studied acceptance and immediate performance far more often than retention, wellbeing, job satisfaction or sustained work engagement.
That distinction matters for staff and patients. A tool can shorten a reading task in a controlled reader study without reducing a department's backlog, improving working conditions or remaining safe when cases, staffing and software change. It can also increase accuracy on one test set while shifting responsibility, creating extra checking work or weakening confidence among less experienced staff. The review therefore does not establish that AI solves radiology shortages or makes a workforce sustainable. It organises what is known and shows what health services still need to measure after procurement and initial training.[1]
How the review selected and combined 69 different kinds of study
The authors searched Embase, PsycINFO, Medline, CINAHL, Web of Science, ScienceDirect and Scopus between 25 August and 1 September 2025. Eligible papers were English-language empirical articles published from January 2020 through September 2025 in journals ranked Q1 or Q2 by SCImago. They had to involve medical-imaging professionals interacting with or evaluating AI outputs. Purely technical model-development papers, reviews, conference papers, editorials, industry reports and studies without professional interaction were excluded. The journal-rank and language filters may raise an evidence-quality floor, but they also risk excluding relevant regional and practice-based knowledge.
Covidence removed 21,685 duplicate records before three reviewers screened titles and abstracts and then reviewed full texts. Formal inter-rater agreement was not calculated. One reviewer extracted data from the 69 included studies and another checked the extraction. One reviewer also performed the inductive thematic coding, with a second reviewer checking study-to-theme assignments and the wider team discussing the emerging structure. The team used the Mixed Methods Appraisal Tool for quality assessment. Those checks add discipline, but single-reviewer extraction and coding can still shape which concepts become prominent.[1]
The distinction that matters
What the review found—and what it could not establish
Counts show how often themes appeared across 69 heterogeneous studies; they are not pooled effect sizes.
Immediate performance
37 studies
Common, but not long-term sustainability
Training and skills
27 studies
Recurring implementation need
Mental strain
11 studies
Mixed experiences and limited evidence
Job satisfaction / engagement
5 / 3 studies
Evidence base remains thin
Short-term performance evidence was common; sustainable-work evidence was not
Thirty-seven of the 69 studies, or 54%, contained immediate task-level findings such as reading time, reporting time, sensitivity, diagnostic accuracy or lesion detection when professionals used AI assistance. The reviewers deliberately treated these as short-term performance outcomes rather than proof of workforce adaptation. That is an important qualification: frequency counts show how often a theme appeared, not the size or reliability of an effect. Studies could contribute to several themes, so the percentages do not sum to 100 and cannot be used as a meta-analysis.
Training and skills development appeared in 27 studies, or 39%. Thirteen studies discussed organisational policies, guidelines or regulatory compliance, while 11 examined mental health or work strain. Only five discussed job satisfaction and three discussed work engagement. Some papers associated AI with reduced repetitive work, lower strain or improved confidence; others described anxiety, overload, deskilling concerns and uncertainty about accountability. Because designs, settings and measures varied, the review could not calculate a pooled change in burnout, retention or productivity. The safest conclusion is that implementation conditions recur across the literature, not that any one condition guarantees a better workplace.[1]
Autonomy and identity are operational safety questions, not resistance to be managed away
Autonomy and control appeared in 26 studies, and threats to professional role or identity in 25. These themes include the ability to inspect, reject or correct an AI output, clarity about who remains accountable, fear of deskilling and concern that professional judgment will be devalued. Framing those reactions simply as reluctance would miss their safety function. A radiographer or radiologist who cannot understand when a system is outside its intended use may either over-trust it or spend so much time checking it that the promised efficiency disappears.
For managers, the practical response is not generic reassurance. Teams need protected training time, access to representative cases, a visible route for reporting failures, authority to override outputs and feedback about what happens after an incident. Job design should specify which decisions remain human, which tasks are delegated and how performance is monitored after updates. Procurement should include the cost of validation, education and workflow redesign rather than treating the licence price as the whole intervention. The review supports these questions as recurring concerns; it does not supply a validated implementation recipe or a universal training curriculum.[1]
The proposed framework is useful for measurement, not a proven causal pathway
The authors combine the themes into a conceptual route from adoption factors through post-adoption behaviours to workforce sustainability. In that model, perceived benefit, trust, readiness, autonomy, identity and intention affect adaptation; training, performance, mental burden and policy compliance then relate to work engagement and job satisfaction. The paper explicitly says these relationships have not been empirically tested together and should not be interpreted causally. A conceptual framework can help a service decide what to observe, but arrows on a diagram are not evidence that changing one factor will improve another.
A stronger test would follow workers and departments before and after a defined deployment, using comparable sites or a stepped rollout where feasible. It would measure not only attitudes but staffing, turnover, sick leave, overtime, reading quality, error review, time spent checking AI, access to training and patient outcomes. It would also report whether effects differ by profession, seniority, workload, modality and prior AI experience. Without those denominators and counterfactuals, a department cannot know whether apparent adaptation reflects the technology, local leadership, case mix or people simply absorbing extra work.[1]
What this means for workers, patients and health systems now
For imaging professionals, the review validates that implementation is a continuing work process, not a one-off competence check. People need time to learn how a specific tool fails, permission to challenge it and a role in redesigning the workflow. For patients, those conditions influence whether an alert is checked, whether responsibility is clear and whether staff have enough attention for communication. For health systems facing shortages, AI may support particular tasks, but capacity gains should be demonstrated in the real service rather than inferred from an accuracy result or a faster reading time in a study.
The evidence is also geographically and linguistically bounded by the review's English-only, Q1/Q2 eligibility rule and its search end date of September 2025. Workplaces with different regulation, resources, professional roles or training infrastructure may face different trade-offs. The review was funded by Australia's National Health and Medical Research Council, and the authors declared no conflicts of interest. Its most useful contribution is a disciplined agenda for post-deployment evaluation. It is not evidence that workers should accept an AI system because benefits are inevitable, or that professional concern is itself a barrier to progress.[1]
What would change the assessment
Confidence would increase with preregistered, multi-site longitudinal studies that follow real deployments for long enough to capture learning, software updates and staff turnover. Researchers should separate radiologists, radiographers, technicians and support staff; report sample sizes and attrition; compare exposed and unexposed teams; and link workforce measures to diagnostic quality and patient experience. Trials of protected training time, override design, audit feedback or workload policies could show which organisational changes actually help rather than merely correlate with positive attitudes.
This publication is current and consequential for implementation practice, but it is not an urgent safety alert, regulatory action or demonstrated change in care. It belongs in Work & Skills and the research feed, not the Breaking desk. The central conclusion remains deliberately modest: radiology teams consistently identify trust, competence, autonomy and organisational support as important, while the evidence needed to show durable workforce benefit is still scarce.[1]
Watch the background, then check the evidence
Illustrative photo · not a video frame
UNESCO · 11 min 27 sec · Background explainer
UNESCO Talks: Interview with 2024 Nobel laureate Professor Daron Acemoglu
YouTube loads only when you choose; playback starts when you press play.
Economist Daron Acemoglu discusses how choices about AI affect workers, inequality and society. Pair the interview with our comparison of task exposure, hiring and skills evidence.
Evidence check: An expert interview from 2024, offering a perspective on policy and work rather than a measured forecast of job losses.
Published 2024-10-14. Source checked 14 October 2026. Check captions on the publisher’s video page.
Read the AI and jobs guidePhoto: Dylan Gillis / Unsplash · Licence
Continue the story
Read next
Continue exploring training: compare the finding and its evidence limits.
Is AI easing worker shortages across OECD countries?
Not visibly in the OECD's new 38-country indicator. ICT, health and teaching professionals remain among the hardest workers to recruit, while the data cannot isolate how much AI caused any change.
Read the next reportExplore the full source and evidence record
What this means for people
- Imaging professionals need protected learning time, clear override authority and a route to report failures after deployment—not only a launch-day demonstration.
- Patients benefit only if AI-supported workflows preserve attention, accountability and diagnostic quality under real workload conditions.
- The review does not show that AI reduces shortages, burnout or turnover; those outcomes require longitudinal comparison in working services.
Global context
Workforce roles, regulation, training capacity and imaging demand differ across health systems. The recurring themes are useful questions for services everywhere, but the review's English-only, higher-ranked-journal sample cannot establish a universal implementation pathway.
What the evidence does not yet show
- Systematic thematic review rather than a meta-analysis
- 69 heterogeneous empirical studies
- English-language Q1/Q2 journal articles only
- Search ended September 2025
- Single-reviewer extraction and coding with secondary checks
- No formal screening or coding agreement statistic
- Theme frequencies are not effect sizes
- Conceptual framework has not been tested as a causal pathway
What to watch next
- Multi-site longitudinal evidence after real deployments
- Turnover, sick leave, overtime and workload alongside accuracy
- Effects of protected training, audit feedback and override design
- Differences by profession, seniority, modality and region
- Patient experience and safety outcomes linked to workforce measures
Living evidence record
Impact record IAI-0GKFUW5
Evidence stage
Studied
Confidence
Supported
Reporting basis
Source analysis
Independent or research support
Present
Record status
Monitoring
Last checked
4 November 2026
Source trail
1 direct source across 1 source type.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Single-source reporting disclosure
This record analyses one direct source. It can establish what JMIR AI published or reported, but it is not independent corroboration of every performance claim or predicted outcome. The confidence label will change only when broader evidence is added.
Evidence trail
Sources used for this report
Links checked 4 November 2026
Research question, topics and regions
How medical-imaging workers adopt and adapt to AI, and what the literature establishes about longer-term workforce sustainability
Topics: Radiology · Workforce adaptation · Training · Professional autonomy · Workplace wellbeing · AI implementation
Regions: Australia · Global
This report is labelled source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.