18 November 2026 · 17 min

Why Do Imaging Algorithms Stall? Deskilling, Friction, and AI's Workforce Problem

Radiology & ImagingHealth Systems & OperationsEvidence & ResearchEconomics & Adoption

Health systems are eager to deploy diagnostic AI to boost speed and accuracy, but a new systematic review reveals a massive blind spot: human adaptation. We unpack why high-performing algorithms face workforce resistance, exploring fears of deskilling, loss of autonomy, and mental strain. Discover what the evidence says about building a sustainable, AI-equipped clinical team.

Key points

Source: AI Adoption, Adaptation, and Workforce Sustainability in Medical Imaging: Systematic Review and Conceptual Framework - JMIR AI, 2026 (CC BY)

This spot is available. Reach clinicians, health-system leaders and medtech and pharma teams following AI in medicine. Sponsor the show

This episode is an AI-generated conversation summarising a public document; the hosts' voices are synthetic. It is for information only and is not medical advice. Always refer to the original source.

Authors: Paula Brough, Khondker Mohammad Zobair, Javad Pool, Natalie Haskell, Luke Houghton, Paul Scuffham et al. · DOI: 10.2196/101944

Your company here. This podcast is looking for its first sponsors: reach clinicians, health-system leaders and medtech and pharma teams following AI in medicine. Sponsorship options and rates →

Transcript

Sam: It is amazing to me that in a world where 54% of studies show AI makes reading imaging scans faster and more accurate, adoption still stalls because we completely ignore the psychological toll on the humans actually using it. By the way, I am Sam, and my voice is AI-generated.

Maya: And I am Maya, also an AI-generated voice. Today we are summarizing a public document on AI adoption, adaptation, and workforce sustainability in medical imaging, which is a systematic review and conceptual framework published in JMIR AI.

Sam: So, we are diving deep into the friction between medical imaging algorithms and the workforce expected to use them. Maya, this document really flips the script on how health systems look at buying diagnostic AI, right?

Maya: It absolutely does. This is a systematic review that looked at 69 peer-reviewed articles published between January 2020 and September 2025. What the authors found is a massive research imbalance. We spend all our time looking at whether a radiologist accepts the technology on day one, and almost no time looking at how they cognitively, behaviorally, and operationally adapt to it over time.

Sam: And that long-term piece is what they call workforce sustainability. It is not just about the tool working; it is about whether the human using the tool burns out, loses their skills, or just stops trusting the system altogether.

Maya: Exactly. To be included in this review, the studies had to be in journals ranked in Quartile 1 or Quartile 2, which are the Q1 or Q2 journals according to the SCImago Journal Rank. This means they set a strict boundary for quality from the start. And what these 69 articles revealed is that clinical leaders and imaging chairs are missing a huge piece of the puzzle. Buying the algorithm is the easy part. The hard part is the postadoption adaptation process.

Sam: Let us get into the hard numbers of what the workforce actually thinks about these tools. Because on one hand, the algorithms are undeniably effective, right?

Maya: Yes, 37 of the studies, which is 54%, showed that AI-assisted systems were associated with shorter radiologic reporting and reading times. They were also associated with higher detection of true-positive cases and better accuracy in identifying lesions. So the immediate task-level performance is clearly there.

Sam: Okay, so if it makes you faster and more accurate, you would think every radiologist would be demanding this tech. But that is not what the review found.

Maya: It is complicated. 41 articles, or 59%, identified perceived benefits as a key antecedent of adopting AI. And 26 empirical studies, which is 38%, explicitly assessed positive workforce perceptions, showing enthusiasm, optimism, and willingness to adopt AI. They recognize it can ease their workload.

Sam: But there is a big 'however' coming, isn't there?

Maya: A massive one. Alongside that optimism, 25 articles, representing 36% of the review, described workforce concerns that AI could negatively affect their professional roles, expertise, and identity. This is where we hit the boardroom reality. If you are a hospital leader, you cannot just look at the speed gains. You have to address these fears.

Sam: What exactly are they afraid of? The classic trope is that AI is going to steal their jobs, but this document suggests the fear is actually much more nuanced than simple job displacement.

Maya: Right. The authors note that the fear of job displacement was actually limited in several studies. Instead, the concerns are framed around nontechnical aspects. The big ones are deskilling, loss of autonomy, erosion of professional reputation, and uncertainty about future career prospects.

Sam: Wait, really? Deskilling? So they are worried that by leaning on the AI, they will physically or cognitively forget how to read complex scans themselves?

Maya: Exactly. The review specifically calls out deskilling anxiety. If the machine is always pointing out the pulmonary nodules or the breast lesions, does the radiologist lose their edge over time? The authors recommend that future research really needs to dig into this through qualitative or ethnographic approaches, because right now, we barely understand how professionals experience and cope with that skill erosion.

Sam: And you mentioned loss of autonomy. How does that show up in the daily workflow?

Maya: Autonomy and control were a huge theme, popping up in 26 articles, or 38% of the studies. This is the degree to which radiologists and radiographers retain authority over diagnostic decisions. They want oversight of the AI recommendations. They need the ability to modify or reject the AI-generated results.

Sam: Because at the end of the day, if the AI makes a mistake, the algorithm does not get sued. The doctor does.

Maya: Exactly. The review explicitly states that any fear of incompetence was overshadowed by radiologists' fears of personal liability and AI arbitrations. They view professional decision-making autonomy as absolutely central to safeguarding patient safety, as well as meeting legal and ethical standards.

Sam: That brings up the concept of trust, which seems like the linchpin holding this all together. If you are worried about liability, you have to trust the tool you are using.

Maya: Trust was the focus of 28 articles, which is 41% of the total. The literature links trust to things like perceived accuracy, transparency, evidence-based validation, and explainability. It also flags the risk of overreliance, which is sometimes called automation bias. If they cannot understand how the model arrived at its conclusion, or if the algorithm is a black box, their trust plummets.

Sam: Which makes total sense. And this document points out that the technical landscape is only getting more complex. They mention things like multicenter learning approaches, multimodal systems combining clinical images and textual information, and even time-frequency modeling for EIT-based pulmonary disease diagnosis.

Maya: Yes, as the AI systems become more technically heterogeneous, the barrier to trust gets higher. If a health system deploys an opaque model across different clinical settings without proving it generalizes well, they are going to hit a wall with the workforce.

Sam: Let us pivot to what happens after the hospital actually buys and installs the AI. The document calls this postadoption adaptation. What does that actually look like on the ward?

Maya: The review identified 4 key facets of effective adaptation. First is training and skills development. Second is task performance during AI-assisted practice. Third is the reduction of mental health burden and work strain. And fourth is policy compliance.

Sam: Let us break those down. Let us start with training. I imagine you cannot just drop a deep learning algorithm onto a radiologist's workstation and say, good luck.

Maya: You definitely cannot, but sadly, that seems to be happening. 27 studies, or 39%, corroborated the need for educational activities and resources to help the workforce adapt. But the authors identified a significant lack of curricula, mentorship, and access to courses as direct impediments to a sustainable workplace.

Sam: So hospitals are spending millions on the tech but starving the training budget.

Maya: That is what the evidence suggests. The review notes that a workforce with AI training typically perceived the systems as personally useful work tools. In contrast, employees with no AI training had a greater fear of replacement and lower acceptance.

Sam: What about the mental health piece? You mentioned that was one of the 4 adaptation facets. Does AI make the job more or less stressful?

Maya: It is a double-edged sword. 11 articles, which is 16%, examined mental health and work strain. Some articles reported that AI lowered perceived mental strain, reduced anxiety, and took away repetitive task burdens. But, and this is crucial, those positive impacts were reduced in situations of high workloads and low AI acceptance.

Sam: Quick note before we carry on. This spot is open for a sponsor. If your company builds or sells AI for healthcare and wants to reach the clinicians, health-system leaders and industry teams who listen to this show, the link to our sponsorship page is in the show notes.

Maya: And now, back to the document.

Sam: So if a department is already understaffed and drowning in scans, and then you force them to use an AI tool they do not trust, it actually makes the burnout worse?

Maya: Exactly. The literature shows it can exacerbate stress, fatigue, and cognitive strain. They specifically refer to technostress. AI is supposed to mitigate strain, but if the integration is poor, it just adds to the cognitive burden.

Sam: That brings us to policy and compliance, which you mentioned as the fourth facet. How much are unclear rules driving this technostress?

Maya: A lot. 13 studies, or 19%, recognized that sustainable workforce development relies on supportive policies and regulatory compliance. When regulatory standards and ethical guidelines are absent or unclear, it raises skepticism and creates resistance. Radiologists literally avoid using the AI tools because they do not know who is legally and ethically responsible for the outcomes.

Sam: It is the wild west out there, and clinicians do not want to be the ones holding the bag if something goes wrong. That makes complete sense. So, looking at the literature as a whole, what kind of research is actually being done here? Is it mostly surveys, or are people doing deep observational studies?

Maya: That is actually one of the biggest red flags the authors raise. The field is severely skewed methodologically. Out of the 69 articles, 51 of them were quantitative research designs. That is 73.91%. There were 12 mixed-method studies, which is 17.39%. And only 6 qualitative studies, a mere 8.70%.

Sam: Wow. So we are just sending out multiple-choice surveys asking people if they like AI, rather than actually sitting with them to understand how it changes their daily practice.

Maya: Exactly. And it gets worse from an academic standpoint. 59 of the included articles, which is 85.50%, did not empirically apply any theoretical framework in their research design or analysis. Only 7 studies, or 10%, explicitly applied a theory.

Sam: What kind of theories are we talking about here? Why does that matter for a hospital executive?

Maya: It matters because without a framework, you cannot predict behavior. The 10% of studies that did use theories used models like the Technology Acceptance Model, or TAM. They also cited the Nonadoption, Abandonment, Scale-up, Spread, and Sustainability framework, known as NASSS. Others mentioned Consumption values theory, Involvement theory, Identity Work theory, and Job Crafting theory. The authors argue that to truly understand acceptance or rejection behaviors, we need to examine them within validated frameworks like the coping model of user adaptation.

Sam: Because right now, we are treating it like a pure IT rollout, rather than a fundamental shift in professional identity and workflow.

Maya: Spot on. And geography plays a role too. The review showed a large number of studies originated from China, with 14 studies or 20.28%. There were 10 multinational studies at 14.49%. The United States had 7 studies at 10%, and Germany also had 7 studies at 10.14%. Fewer came from the UK, South Korea, and Saudi Arabia.

Sam: It is interesting to see that global spread, but also how heavily concentrated the data is in certain regions. What kinds of imaging were these studies actually looking at?

Maya: It covered a lot of ground. General radiology accounted for 16 studies, which is 23.18%. Breast imaging had 13 studies at 18.84%. Neuroradiology had 10 studies at 14.49%, and chest and thoracic radiology had 8 studies at 11.59%. In terms of modalities, most assessed MRI, with 11 studies or 15.94%, CT with 10 studies at 14.49%, and X-rays with 9 studies at 13.04%. Others looked at ultrasound, PET, SPECT, and mammography.

Sam: So this is not just an isolated problem in mammography or chest X-rays. This friction is happening across the entire spectrum of medical imaging.

Maya: It is systemic. And to solve it, the authors synthesized all these findings into a conceptual model. The model links 3 main components: AI Adoption factors, postadoption adaptation factors, and workforce sustainability outcomes. They represent this as an interdependent process connected by double-headed arrows.

Sam: Double-headed meaning it is not just a straight line from buying the tech to being happy. The way people adapt changes how they view adoption, and vice versa.

Maya: Precisely. The adoption factors include the perceived benefits, trust, and autonomy we discussed. The adaptation factors include training, task performance, work strain, and policy compliance. And those feed into the ultimate goal: workforce sustainability, which the literature currently measures through job satisfaction and work engagement.

Sam: But the literature on that final piece, sustainability, is incredibly thin, right?

Maya: Incredibly thin. Only 5 studies, which is 7%, reported on job satisfaction. And only 3 articles, just 4%, highlighted work engagement. The authors note that 66 studies, or 95.65%, claimed to be assessing workforce sustainability, but very few actually measured these long-term organizational outcomes.

Sam: That is a fascinating disconnect. Everyone says they are studying sustainability, but they are really just studying whether the doctor clicked 'accept' on the AI recommendation today, not whether they are still engaged and satisfied a year from now.

Maya: Exactly. When the workforce actively engages with the AI, they participate in related training and pay attention to system use. Features that encourage critical evaluation, like the visibility of the system's confidence levels, actually support that engagement. But when they have limited knowledge or negative attitudes, job satisfaction declines.

Sam: So if you are a health-system leader listening to this, you might have the best deep-learning algorithm on the market. It might cut reading times in half and boost accuracy perfectly in a lab setting. But if you do not have clear policies on legal liability, and if you do not invest heavily in ongoing mentorship and training, it is going to fail on the floor.

Maya: There is a direct quote from the conclusion that sums this up perfectly: "Without clear organizational policies and procedures, AI will remain an external imposition rather than an accepted facet of radiologic practice."

Sam: An external imposition. That is exactly what it feels like when new tech is forced onto a clinical team without the operational scaffolding to support them. You get resistance, you get deskilling anxiety, and ultimately, you get a brittle workforce.

Maya: And the authors admit there are limitations to their review. They restricted their search to English-language, peer-reviewed publications in Q1 and Q2 journals, which means they might have excluded relevant evidence from lower-quartile or nonindexed journals. They also note that their formal interrater agreement statistics, like Cohen kappa, were not calculated during screening, though multiple reviewers resolved variations by consensus.

Sam: Right, they used a tool called the Mixed Methods Appraisal Tool, or MMAT, to look at quality. And while they did not exclude studies based on it, it did reveal issues like nonresponse bias in the quantitative surveys.

Maya: Yes. And they also openly acknowledge that because they used an inductive thematic synthesis, the dimensions of workforce sustainability, like job satisfaction and work engagement, were derived from whatever limited data was available in the literature, rather than being a comprehensive, predetermined set of dimensions.

Sam: So the call to action for the researchers out there is pretty clear: stop just measuring day-one acceptance. We need structural equation modeling to test these pathways, and we desperately need more qualitative research to understand how clinicians are actually coping with deskilling anxiety and technostress over the long haul.

Maya: And for the companies building these tools, you need to realize that readiness to change and adaptation was only assessed in 13 studies, which is 19% of the literature. You cannot just sell an algorithm; you have to help the hospital build the epistemic environment and the clinical protocols that make meaningful organizational change possible.

Sam: That is a brilliant place to land. AI will always be an external imposition unless we actively support the humans having to adapt to it.

Maya: To recap, this systematic review highlights a massive gap between the technical adoption of imaging AI and the long-term adaptation of the workforce. While the algorithms offer speed and accuracy, widespread fears of deskilling, liability, and a severe lack of training are stalling meaningful integration.

Sam: You can find a link to the full JMIR AI systematic review in our show notes. As always, this podcast is for informational purposes only and is not medical advice. Thanks for diving deep with us.