Radiology staff perspectives after decommissioning an artificial intelligence (AI) tool for expedited lung cancer triage

Classified as probably not a substantial claim.

Recent frameworks examining automation withdrawal in high-reliability organisations emphasise that decommissioning should be recognised as an active stage in the technology lifecycle rather than an administrative end point.

“Recent frameworks examining automation withdrawal in high-reliability organisations emphasise that decommissioning should be recognised as an active stage in the technology lifecycle rather than an administrative end point 14 .”

• cites [14] Chambers DA, Glasgow RE, Stange KC (2013) The dynamic sustainability framework: addressing the paradox of sustainment amid ongoing change. Implement Sci 8:117. https://doi.org/10.1186/1748-5908-8-117 sourceCould not verify

  The Abstract’s Discussion and the Discussion section on DSF implications describe ongoing adaptation and sustainment of healthcare interventions, not their withdrawal or decommissioning. Searches for automation withdrawal, high-reliability organisations, decommissioning, and intervention termination found no passage addressing the claimed lifecycle stage. (LLM, conf=0.76)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

Relatively few studies have examined decommissioning of functional digital tools in healthcare.

“Relatively few studies have examined decommissioning of functional digital tools in healthcare.”

• cites [12] Ahmed MI, Spooner B, Isherwood J, Lane M, Orrock E, Dennison A (2023) A systematic review of the barriers to the implementation of artificial intelligence in healthcare. Cureus 15(10):e46454. 10.7759/cureus.46454 sourceCould not verify

  The Review discusses barriers to implementing and maintaining AI tools, not the decommissioning of functional digital tools. Searches for “decommissioning” and for discontinuing or retiring working digital health technologies found no passage assessing how many studies have examined that topic. Page numbers were unavailable in our copy. (LLM, conf=0.77)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Such withdrawal can disrupt established workflows, reduce confidence in the stability of digital initiatives, reduce trust in managerial decisions and create uncertainty about the longevity or scope of future innovations.

“Such withdrawal can disrupt established workflows, reduce confidence in the stability of digital initiatives, reduce trust in managerial decisions and create uncertainty about the longevity or scope of future innovations.”

• cites [6] Hilty DM, Armstrong CM, Smout SA, Crawford A, Maheu MM, Drude KP et al (2022) Findings and guidelines on provider technology, fatigue, and well-being: scoping review. J Med Internet Res 24(5):e34451. 10.2196/34451 sourceCould not verify

  The Introduction (Background) discusses integrating technology into workflows, and the Results (Organizational Responses) discusses adjusting workflows through change management. Neither addresses the withdrawal of a digital initiative or its claimed effects on confidence, managerial trust, or uncertainty about future innovations. Searches for withdrawal, discontinuation, stability, and trust did not locate a passage supporting those effects. (LLM, conf=0.74)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Automation bias may creep into practice, potentially leading to deskilling within the profession with long-term AI usage, an already recognised phenomenon with AI use in radiology.

“Whilst this aspect was not fully explored in our work, it highlights how 'automation bias' may creep into practice, potentially leading to deskilling within the profession with long term AI usage, an already recognised phenomenon with AI use in radiology 11 .”

• cites [11] Savardi M, Signoroni S, Benini S, Vaccher F, Alberti M, Ciolli P et al (2025) Upskilling ot deskilling? Measurable role of an AI-supported training for radiology residents: a lesson from the pandemic. Insights into Imaging 16(1):23. 10.1186/s13244-024-01893-4 sourceSupported

Probably not a substantial claim: may not pass S3 (load-bearing).

Repeated cycles of short-term pilots could reduce staff willingness to engage with future digital investment projects.

“Staff witness care quality decline for administrative purposes outside their control which could mean that repeated cycles of short-term pilots could reduce their willingness to engage with future digital investment projects.”

• cites [6] Hilty DM, Armstrong CM, Smout SA, Crawford A, Maheu MM, Drude KP et al (2022) Findings and guidelines on provider technology, fatigue, and well-being: scoping review. J Med Internet Res 24(5):e34451. 10.2196/34451 sourceCould not verify

  The Introduction (Background) discusses technology-related fatigue, and the Results discuss provider engagement, but neither links repeated short-term pilots to willingness to join future digital projects. Searches for pilots, short-term projects, and change fatigue found no passage addressing that link. (LLM, conf=0.72)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

When a tool perceived as clinically useful is withdrawn for reasons unrelated to performance, some staff may experience this as the loss of a valuable component of their daily work (moral distress or moral injury), stemming from dissonance between sense of purpose and actual practice.

“When a tool perceived as clinically useful is withdrawn for reasons unrelated to performance, some staff may experience this as the loss of a valuable component of their daily work (sometimes referred to in the literature as 'moral distress or moral injury' 13 , stemming from the dissonance between sense of purpose and actual practice).”

• cites [13] Cartolovni A, Stolt M, Scott P, Suhonen R (2021) Moral injury in healthcare professionals: A scoping review and discussion. Pubmed 28(5):590–602. 10.1177/0969733020966776 sourcePartly supported

  The Discussion and concept contextualization section supports moral distress as a response to impeded professional goals or thwarted moral action, which fits the cited explanation of dissonance between purpose and practice. It does not address withdrawal of a clinically useful tool, and it distinguishes moral injury from distress by linking injury to potentially morally injurious events and more lasting consequences. Thus, the source supports the conceptual analogy, but not the specific tool-withdrawal scenario or its characterization as moral injury. (LLM, conf=0.85)

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing).

Responses are subject to self-selection and social desirability biases.

“As with all survey studies, responses are subject to self-selection and social desirability biases 15 .”

• cites [15] Campbell S, Greenwood M, Prior S, Shearer T, Walkem K, Young S et al (2020) Purposive sampling: complex or simple? Research case examples. J Res Nurs 25(8). 10.1177/1744987120927206 sourcePartly supported

  The Introduction (page unavailable) describes voluntary participation and says that women in one case study self-select, supporting a concern about who responds. It does not address social desirability bias or establish that both biases apply to all survey studies. (LLM, conf=0.82)

  Experimental — accuracy not yet measured

May combine two claims.

Broader digital health and implementation science literature has described phenomena such as non-adoption, abandonment and limited sustainability, but often focuses on technologies that fail to embed rather than those that are actively withdrawn after successful pilot use.

“Broader digital health and implementation science literature has described phenomena such as non-adoption, abandonment and limited sustainability, but often focuses on technologies that fail to embed rather than those that are actively withdrawn after successful pilot use 12 .”

• cites [12] Ahmed MI, Spooner B, Isherwood J, Lane M, Orrock E, Dennison A (2023) A systematic review of the barriers to the implementation of artificial intelligence in healthcare. Cureus 15(10):e46454. 10.7759/cureus.46454 sourcePartly supported

  Split into 2 atoms — 0 confirmed, 2 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

  Experimental — accuracy not yet measured

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Several radiographers described the workflow as calmer and less fragmented after decommissioning, attributing this to removal of real-time alerts that previously increased interruptions, time pressure and emotionally challenging patient conversations.

“Several radiographers described the workflow as calmer and less fragmented after decommissioning, attributing this to the removal of real-time alerts that previously increased interruptions, time pressure and emotionally challenging patient conversations.”

• AssumptionRests on an assumption: Themes identified from free-text responses represent the views of the respondent population.

May combine two claims.

Perceived logistical burden regarding AI use increased after AI removal, with 35% agreeing or strongly agreeing that AI had caused departmental logistical problems, similar to pre-implementation (33.3%) and higher than late post-implementation (26.4%).

“Perceived logistical burden regarding AI use increased after AI removal, with 35% of respondents agreeing or strongly agreeing that AI had caused departmental logistical problems, a proportion similar to preimplementation expectations (33.3%), and higher than at late post-implementation (26.4%)”

• AssumptionRests on an assumption: Differences in survey responses across timepoints reflect the effect of AI implementation or decommissioning rather than differences in respondent samples or other temporal factors.
• AssumptionRests on an assumption: The survey items validly measure staff attitudes and perceptions across the four timepoints.

May combine two claims.

12.5% of respondents reported reduced satisfaction at work after decommissioning, an increase from 5.7% at late post-implementation.

“with an identical proportion (12.5%) reporting reduced satisfaction at work after decommissioning (an increase from 5.7% at late post-implementation)”

• AssumptionRests on an assumption: Differences in survey responses across timepoints reflect the effect of AI implementation or decommissioning rather than differences in respondent samples or other temporal factors.
• AssumptionRests on an assumption: The survey items validly measure staff attitudes and perceptions across the four timepoints.

May combine two claims.

Personal comfort with AI use declined after decommissioning; only 25.6% agreed or strongly agreed they would be comfortable with AI diagnosing their own chest radiograph, compared to 47.2% at late post-implementation and 48.3% at early post-implementation.

“Only 25.6% of respondents agreed or strongly agreed that they would be comfortable with AI diagnosing their own chest radiograph, compared to 47.2% at late post-implementation, and 48.3% at early post-implementation.”

• Also stated in: § Discussion · plain

  “Notably, personal comfort with AI applied to one's own imaging (if one were a patient) fell to its lowest level after withdrawal, despite having been high during early active deployment.”

• AssumptionRests on an assumption: Differences in survey responses across timepoints reflect the effect of AI implementation or decommissioning rather than differences in respondent samples or other temporal factors.
• AssumptionRests on an assumption: The survey items validly measure staff attitudes and perceptions across the four timepoints.

May combine two claims.

Among reporters at decommissioning (n=5), two (40%) expressed disappointment at no longer using the AI tool, while only one (20%) admitted feeling reliant on the tool for their practice.

“Whilst two reporters (40%) expressed disappointment at no longer using the AI tool, only one (20%) admitted to feeling reliant on the tool for their practice.”

• AssumptionRests on an assumption: The survey items validly measure staff attitudes and perceptions across the four timepoints.

May combine two claims.

Support for human-in-the-loop models strengthened after AI removal, with 92.3% of respondents agreeing that radiologists or reporting radiographers should continue to review chest radiographs with AI acting as triage aid (compared to 86.7%-89.7% at prior survey phases).

“Support for human-in-the-loop models strengthened after AI removal, with 92.3% of respondents agreeing that radiologists or reporting radiographers should continue to review chest radiographs with AI acting as triage aid (compared to 86.7% -89.7% agreement at prior survey phases)”

• AssumptionRests on an assumption: Differences in survey responses across timepoints reflect the effect of AI implementation or decommissioning rather than differences in respondent samples or other temporal factors.
• AssumptionRests on an assumption: The survey items validly measure staff attitudes and perceptions across the four timepoints.

Classified as probably not a substantial claim. May combine two claims.

Concerns raised following AI withdrawal included loss of a rapid triage mechanism, potential for subtle abnormalities to be missed without AI support, and disappointment that a clinically useful tool had been removed.

“Concerns raised following AI withdrawal included the loss of a rapid triage mechanism, the potential for subtle abnormalities to be missed without AI support, and disappointment that a tool perceived as clinically useful had been removed.”

• AssumptionRests on an assumption: Themes identified from free-text responses represent the views of the respondent population.

May combine two claims.

12.5% of respondents did not think there were any drawbacks to the AI tool, reduced from 26.4% at late post-implementation.

“12.5% of respondents did not think there were any drawbacks to the AI tool (reduced from 26.4% at late postimplementation)”

• AssumptionRests on an assumption: Differences in survey responses across timepoints reflect the effect of AI implementation or decommissioning rather than differences in respondent samples or other temporal factors.
• AssumptionRests on an assumption: The survey items validly measure staff attitudes and perceptions across the four timepoints.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

Some respondents viewed decommissioning as a step backwards in digital innovation and a reduction in learning opportunities previously afforded by AI-highlighted findings.

“Some respondents (both reporters and non-reporters) viewed decommissioning as a step backwards in digital innovation and a reduction in learning opportunities previously afforded by AI-highlighted findings.”

• AssumptionRests on an assumption: Themes identified from free-text responses represent the views of the respondent population.

May combine two claims.

Removal of the AI tool was associated with fewer same-day escalations and a shift away from reactive pathway coordination, which was perceived as increasing the time to follow-up for some patients.

“their responses suggested that removal of the AI tool was associated with fewer same-day escalations and a shift away from reactive pathway coordination, which was perceived as increasing the time to follow-up for some patients.”

• AssumptionRests on an assumption: The survey items validly measure staff attitudes and perceptions across the four timepoints.

May combine two claims.

Autonomous AI diagnosis remained unfavourable, with 75% disagreement for an AI to report chest radiographs independently at decommissioning, similar to post-implementation phases (75.9% and 72.5%) and higher than pre-implementation (53.3%).

“In contrast, autonomous AI diagnosis remained unfavourable, with 75% disagreement for an AI to report chest radiographs independently at decommissioning, similar to levels of disagreement during post-implementation phases (75.9% and 72.5%) and higher than at preimplementation stage (53.3%).”

• AssumptionRests on an assumption: Differences in survey responses across timepoints reflect the effect of AI implementation or decommissioning rather than differences in respondent samples or other temporal factors.
• AssumptionRests on an assumption: The survey items validly measure staff attitudes and perceptions across the four timepoints.

May combine two claims.

Increased workload for radiology staff was the most frequently cited drawback after decommissioning (35.0%), similar to early implementation (41.4%), followed by increased departmental costs (22.5%).

“Increased workload for radiology staff was the most frequently cited drawback after decommissioning (35.0%) (similar to early implementation levels at 41.4%), followed by increased departmental costs (22.5%).”

• AssumptionRests on an assumption: Differences in survey responses across timepoints reflect the effect of AI implementation or decommissioning rather than differences in respondent samples or other temporal factors.
• AssumptionRests on an assumption: The survey items validly measure staff attitudes and perceptions across the four timepoints.

May combine two claims.

Reporters valued AI as a decision support tool and a safety net during image interpretation.

“Reporters, in contrast, valued AI as a decision support tool and a safety net during image interpretation.”

• AssumptionRests on an assumption: The survey items validly measure staff attitudes and perceptions across the four timepoints.
• AssumptionRests on an assumption: Themes identified from free-text responses represent the views of the respondent population.

May combine two claims.

Faster access to follow-up care and time saving for patients and the department remained consistently the most frequently selected benefits across time.

“Whilst the proportion of responses changed over the different survey phases, these two factors remained consistently the most frequently selected benefits across time”

• AssumptionRests on an assumption: Differences in survey responses across timepoints reflect the effect of AI implementation or decommissioning rather than differences in respondent samples or other temporal factors.
• AssumptionRests on an assumption: The survey items validly measure staff attitudes and perceptions across the four timepoints.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

One reporter expressed growing reliance upon the AI tool for work and disappointment at removal.

“One reporter expressed growing reliance upon the tool for work and disappointment at removal.”

• AssumptionRests on an assumption: The survey items validly measure staff attitudes and perceptions across the four timepoints.

Classified as probably not a substantial claim. May combine two claims.

Some operational burden eased after decommissioning, but benefits associated with the AI pathway, such as more rapid escalation to CT, were no longer present.

“Others noted that while some operational burden had eased, benefits associated with the AI pathway -such as more rapid escalation to CT -were no longer present.”

• AssumptionRests on an assumption: Themes identified from free-text responses represent the views of the respondent population.

May combine two claims.

Perceptions of AI's influence for patient benefit remained stable after withdrawal; 70% agreed or strongly agreed that AI use had improved patient care, consistent with earlier phases (65.5-67.9% agreement).

“At decommissioning, 70% of respondents agreed or strongly agreed that AI use had improved patient care, consistent with earlier survey phases (ranging between 65.5-67.9%agreement).”

• AssumptionRests on an assumption: Differences in survey responses across timepoints reflect the effect of AI implementation or decommissioning rather than differences in respondent samples or other temporal factors.
• AssumptionRests on an assumption: The survey items validly measure staff attitudes and perceptions across the four timepoints.

Classified as probably not a substantial claim. May combine two claims.

For non-reporting radiographers, AI withdrawal was perceived as reducing emotional and administrative burden, although others expressed concern that removal of AI might result in slower diagnostic escalation for patients.

“For these staff, AI withdrawal was perceived as reducing emotional and administrative burden although others expressed concern that the removal of AI might result in slower diagnostic escalation for patients.”

• AssumptionRests on an assumption: Themes identified from free-text responses represent the views of the respondent population.

May combine two claims.

When asked to retrospectively identify main benefits of the AI tool, 40.0% (16/40) chose faster access to follow-up care and 32.5% (13/40) selected time saving for patients and the department.

“When asked to retrospectively identify the main benefits of the AI tool, almost half of the respondents (16/40, 40.0%) chose faster access to follow-up care, and about a third (13/40, 32.5%) selected time saving for patients and the department.”

• AssumptionRests on an assumption: The survey items validly measure staff attitudes and perceptions across the four timepoints.

May combine two claims.

Previous analysis across pre-implementation, early and late implementation phases closely mirrored the Gartner Hype Cycle pattern, with early enthusiasm, a peak of frustration as logistical challenges emerged, and later partial adaptation.

“Our previous analysis across the pre-implementation, early and late implementation phases closely mirrored the well-known 'Gartner Hype Cycle' pattern 10 , with early enthusiasm, a peak of frustration as logistical challenges emerged, and later partial adaptation.”

• cites [10] Dedehayir O, Steinert M (2016) The hype cycle model: A review and future directions. Technol Forecast Soc Change 108:28–41. https://doi.org/10.1016/j.techfore.2016.04.005 sourceNot checked — no full text of the source

May combine two claims.

Decommissioning a chest radiograph AI triage tool in an NHS Trust was associated with both operational relief and a sense of clinical loss.

“Decommissioning a chest radiograph AI triage tool in an NHS Trust was associated with both operational relief and a sense of clinical loss.”

• AssumptionRests on an assumption: Differences in survey responses across timepoints reflect the effect of AI implementation or decommissioning rather than differences in respondent samples or other temporal factors.
• AssumptionRests on an assumption: The survey items validly measure staff attitudes and perceptions across the four timepoints.
• AssumptionRests on an assumption: Themes identified from free-text responses represent the views of the respondent population.

May combine two claims.

The sample size was small and predominantly composed of radiographers.

“Our sample size was small and predominantly composed of radiographers.”

May combine two claims.

Although radiographers valued the safety afforded by high sensitivity at a systems level, the workflow and emotional implications contributed to mixed feelings about the tool in practice.

“These experiences suggest that although radiographers valued the safety afforded by the high sensitivity at a systems level, the workflow and emotional implications contributed to mixed feelings about the tool in practice.”

• AssumptionRests on an assumption: Themes identified from free-text responses represent the views of the respondent population.

May combine two claims.

Staff perceived the AI-enabled pathway as logistically demanding but consistently believed that it had improved patient care, with some reporters regretting its removal.

“Staff perceived the AI-enabled pathway as logistically demanding but consistently believed that it had improved patient care, with some reporters regretting its removal.”

• AssumptionRests on an assumption: Differences in survey responses across timepoints reflect the effect of AI implementation or decommissioning rather than differences in respondent samples or other temporal factors.
• AssumptionRests on an assumption: The survey items validly measure staff attitudes and perceptions across the four timepoints.

May combine two claims.

Unlike diagnostic radiographers, reporters were not typically involved in direct patient communication at the point of triage and were therefore less exposed to the emotional weight of managing patient anxiety in response to AI-enabled positive findings.

“Unlike diagnostic radiographers, they were not typically involved in direct patient communication at the point of triage and were therefore less exposed to the emotional weight of managing patient anxiety in response to AI-enabled positive findings.”

Classified as probably not a substantial claim.

Free-text responses were analysed using simple inductive content analysis to identify recurring themes relating to workflow, emotional burden and perceptions of patient impact.

“Free-text responses were analysed using simple inductive content analysis 9 to identify recurring themes relating to workflow, emotional burden and perceptions of patient impact.”

• cites [9] Elo S, Kyngas H (2008) The qualitative content analysis process. JANm Leading globasl Nursing Research. https://doi.org/10.1111/j.1365-2648.2007.04569.xDigital Object Identifier (DOI) sourceNot checked yet

  Not checked in the free run.

May combine two claims.

Many staff reported their clinical workflow felt simpler without the need to act upon AI triage results, although belief that the AI tool had positively contributed to patient care remained high, with a subset of reporting staff expressing disappointment at its removal.

“Many reported that their clinical workflow felt simpler without the need to act upon AI triage results, although belief that the AI tool had positively contributed to patient care remained high, with a subset of reporting staff expressing disappointment at its removal.”

• AssumptionRests on an assumption: Differences in survey responses across timepoints reflect the effect of AI implementation or decommissioning rather than differences in respondent samples or other temporal factors.
• AssumptionRests on an assumption: The survey items validly measure staff attitudes and perceptions across the four timepoints.
• AssumptionRests on an assumption: Themes identified from free-text responses represent the views of the respondent population.

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The survey was anonymous without asking for personal data and respondents gave consent at the beginning of the survey for results to be used for research and education purposes.

“The survey was anonymous without asking for personal data and respondents gave consent at the beginning of the survey for the results to be used for research and education purposes.”

Classified as probably not a substantial claim.

The decommissioning survey retained 24 items unchanged from the original surveys to enable consistent comparison across all four timepoints.

“The 24 retained items included key Likert-scale statements regarding perceptions of patient care, departmental logistics and personal comfort with AI were repeated and unchanged to enable consistent comparison across all four timepoints.”

May combine two claims.

The study did not survey referring clinicians or patients directly, so inferences about their experiences are limited and based on second-hand staff perceptions.

“We also did not survey referring clinicians or patients directly, so our inferences about their experiences are limited and based on second hand staff perceptions.”

Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

The study was registered as a quality improvement project within the Trust (code RAD007), hence ethical approval was not required.

“This work was registered as a quality improvement project within our trust with the code RAD007, hence ethical approval was not required.”

Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

Withdrawing a functional AI triage tool due to external constraints may be experienced as a loss of capability rather than an improvement in care.

“Removing low-value care is typically viewed as a positive step however, withdrawing a functional AI triage tool due to external constraints (such as from funding constraints, contract expiry, changes in strategic priorities, staff changes) may be experienced as a loss of capability rather than an improvement in care.”

May combine two claims.

Staff experiences following AI tool decommissioning are complex, contextual based on staff roles, and reflect both operational relief and clinical benefit loss.

“This study has demonstrated that staff experiences following AI tool decommissioning are complex, contextual based on staff roles and are reflecting both operational relief and clinical benefit loss.”

• AssumptionRests on an assumption: Differences in survey responses across timepoints reflect the effect of AI implementation or decommissioning rather than differences in respondent samples or other temporal factors.
• AssumptionRests on an assumption: The survey items validly measure staff attitudes and perceptions across the four timepoints.
• AssumptionRests on an assumption: Themes identified from free-text responses represent the views of the respondent population.

May combine two claims.

The survey was conducted in a single, multi-site NHS Trust using one particular AI tool, and may not be generalisable to other settings, AI tools or use cases.

“Furthermore, the survey was conducted in a single, multi-site NHS Trust using one particular AI tool, and may not be generalisable to other settings, AI tools or use cases.”