Do diabetes computing systems fade into continued use with cognitive disengagement?

A user state that looks like success in usage logs

The paper's central claim is that continued use and cognitive engagement can come apart. Participants still wore sensors, still opened apps, and still acted on glucose readings, so every external indicator of adoption looked healthy. What declined was self-initiated interpretation: comparing readings, contextualizing them against meals or activity, building explanations, and questioning whether the system's output was right. In the authors' description, users increasingly responded only when an alarm or recommendation prompted them.

This matters because the state is behaviorally invisible. The authors distinguish it explicitly from abandonment research, which assumes disengagement shows up in usage logs. Here the device remains in use and the data pipeline keeps operating, so retention metrics and usage frequency would not register the change. The paper therefore argues that long-term system value should not be judged by behavioral retention, usage frequency, or algorithmic accuracy alone, but by whether human interpretive agency is preserved over time.

Following the same patients across five interview waves

The evidence comes from a longitudinal interview study running from 2017 to 2025, with five waves of interviews with 37 patients with diabetes. The same interview protocol was applied at each wave, which allowed the researchers to trace within-subject change in usage context, interpretive strategy, and decision logic rather than comparing different people at a single moment. A single wave in 2025 added interviews with 15 practitioners: 5 clinicians, 5 designers, and 5 vendors involved in designing and deploying these systems.

The authors are explicit that the study aims at process-level mechanism, not population-level generalization. Participants came from a specific region and were diverse in urban/rural balance, education, and occupation but not statistically representative. The described stages and trajectories are mechanisms observed in a particular setting, and the paper states that generalizability to other healthcare systems and cultural contexts requires further study, especially in low- and middle-income settings where continuous glucose monitoring and automated insulin delivery adoption remain highly unequal.

Coding was checked rather than assumed. Of 246 sub-sample episodes, 31 (12.6%) reached a structured discussion step and 4 (1.6%) required adjudication by a fourth researcher; none ultimately required exclusion. Saturation was assessed separately for descriptive, trajectory, and mechanistic levels, and was re-evaluated each wave because new technology generations can reopen closed categories.

Three generations, each taking over more cognitive work

The study spans three broad periods of technology use. From 2017 to 2019, continuous glucose monitoring entered daily life and physiological data began appearing on screen, displacing intermittent fingerstick testing combined with apps that offered only general knowledge. From 2019 to 2021, real-time feedback designs matured, and part of the information-processing work shifted from physicians and patient self-querying toward system-mediated delegation. From 2021 to 2025, that data began feeding prediction, recommendation, and automated decision-making, including automated insulin delivery.

The authors describe these as macro-level distinctions without clear boundaries; individual adoption timing varied considerably. Some urban and highly educated participants adopted continuous monitoring earlier, and patients with particularly severe disease also tended to gain access earlier, while rural participants with lower education and worse economic conditions tended to adopt later. A few participants who started late never experienced long stretches of the earlier periods at all, because the technology they first met already bundled those capabilities.

The paper's framing is that each generation takes less effort than the previous one. That is the design goal inherited from the vision of technologies that disappear into everyday life. The study's question is what happens over years when that goal is pursued by substituting for users' judgment rather than scaffolding it.

Three stages, and why they are not a schedule

The analysis identifies three recurring configurations of system capability and user practice. The first is self-initiated exploration and sensemaking, where users compare, contextualize, and interpret data themselves; the authors link this to prior work treating reflection as a core component of self-tracking. The second is collaboration and delegation, where real-time feedback and recommendations redistribute part of the interpretive work to the system while participants still communicate with clinicians and interpret some data. The third is continued system use with reduced cognitive engagement.

Crucially, the authors state these stages are not a technologically determined or inevitable progression. Participants entered them at different times, some did not experience all three, and a small number retained or regained active interpretive practices. Some participants experienced only the first two stages, which the authors read as evidence that increasingly capable systems do not inevitably reduce engagement. Others stayed in a low-engagement configuration throughout, which the paper argues should not be labeled disengagement, since that requires a longitudinal reduction from earlier active practice.

The paper also reports that movement into the third stage need not be irreversible. Cases of resumed interpretation after a system failure or after reflection show users can return to more active practice. This is one reason the authors treat the stages as configurations to move through, remain within, or return to, rather than as a one-way ladder.

What pushes users toward disengagement

The paper names three drivers that bring the state about: functional expansion, delegation of interpretive authority, and information fatigue. Functional expansion means the system absorbs more of the work once done by the user. Delegation of interpretive authority means users increasingly treat the system's judgment as the judgment to act on. Information fatigue means the volume and repetitiveness of prompts wear down the appetite for examining data.

Three further conditions reinforce the state: clinical prioritization, information compression, and automation. Clinical prioritization places the clinician's judgment above the user's own reading of the data. Information compression reduces rich data into a single number, trend, or alert. Automation closes the loop between sensing and action with minimal user intervention. Together these make it rational for a user to stop interpreting: the system, the clinic, and the interface all signal that interpretation is no longer the user's job.

The consequence the authors report is less frequent independent health judgment and a shift from understanding to responding. They describe the risk less as a discrete lost skill than as reduced interpretive readiness: users compare, contextualize, verify, and question less often, and resume these practices more slowly when systems behave unexpectedly.

Why this is not just complacency or deskilling

The paper takes care to separate its state from adjacent concepts. It is not abandonment, because abandonment research assumes disengagement is observable in usage logs, whereas here users remain active on every external indicator. It is not automation complacency, which concerns insufficient monitoring of highly reliable automation in supervisory-control settings; the finding here is a broader longitudinal reduction in self-initiated interpretation, not limited to monitoring failures. It is also not simply deskilling, which the authors illustrate with a documented measurable decline in one professional skill after AI assistance is removed; what they observe is more diffuse, spanning observation, comparison, inference, modeling, and questioning.

The authors position the state in a specific gap: distinct from abandonment, broader than complacency, and more diffuse than deskilling. That positioning is the paper's main theoretical contribution, and it is also why the state is easy to miss. A system that keeps users compliant and responsive can look like a success on every metric the field usually collects.

Who moved faster and who held back

Four subgroup comparisons are reported, and the authors label them descriptive rather than tested moderation effects. Urban (n=15) and rural (n=14) completers showed broadly similar trajectory patterns, but rural participants generally gained access to new systems later and therefore entered each stage later.

Education showed a more visible split. Higher-educated participants (bachelor's degree or above; n=4 among completers) more often retained partial verification practices at later stages, and two of them had not entered the third stage within the observed window. Lower-educated participants (primary school or below; n=17 among completers) on average entered the third stage earlier, and some skipped the first two stages entirely. The authors caution that with n=4 and no statistical controls for confounds such as adoption depth, this should be read as hypothesis-generating.

Age did not sort the trajectories cleanly: both older participants (over 60 at study start; n=6 among completers) and younger participants (50 or under; n=13) displayed them. The difference was in how they described the process, with older participants more often framing it as the device watching for them and younger participants more often citing too little time or too much information. Participants who adopted technology earlier generally progressed through the trajectory more quickly, and sudden events, such as one participant's automated insulin delivery incident, prompted reflection.

Design for less work without losing agency

The paper's design argument is not that systems should demand more cognitive effort. It is that future diabetes computing systems should reduce unnecessary cognitive work while preserving opportunities to interpret, contextualize, and question system outputs. The distinction matters: the problem is not convenience itself but the removal of occasions on which users exercise judgment.

The authors are candid that these design implications are inferred from longitudinal interview analysis, not validated through a design intervention. They represent reasoned directions rather than verified solutions, and the second research question is answered at the level of design problems and directions rather than tested fixes. Future work would need to prototype, deploy, and longitudinally evaluate such directions, and the paper also calls for measurement instruments for depth of cognitive engagement that can be tracked in research and operational systems.

The paper extends its framework beyond health only as a generative hypothesis. It proposes continued use with cognitive disengagement as a sensitizing framework for automated driving and driver-assistance systems, where years of reliable assistance might similarly erode situational awareness, and for personal productivity and agentic AI systems that compress decision spaces into single-click approvals. The authors state plainly that these cross-domain connections are theoretical extrapolations, not empirically proven facts outside clinical contexts.

The metric that should change

The study's most consequential implication is a shift in what counts as long-term success. Retention, usage frequency, and algorithmic accuracy all remained compatible with the state the authors describe, because the device was still worn and the data still flowed. The variable that distinguished participants was whether they still compared, contextualized, verified, and questioned what the system produced, and whether they could resume those practices when the system failed.

That is why the paper's subgroup findings matter more than they first appear. The participants who retained partial verification practices, and the two who had not reached the third stage after eight years, show that the trajectory is not fixed by the technology alone. The authors' own framing points to the same priority: evaluate these systems by the longitudinal preservation of human interpretive agency, not by behavioral retention alone. The open question the study leaves is how to measure that preservation well enough to use it as a design target.

Sources used in this answer