The purpose of this study was to investigate how the acceptability and adoption of AI-integrated dementia care are influenced by the stage of the care pathway, system centrism (AI versus human), and the level of involvement of the system in patients’ lives. Our findings demonstrate that public perceptions are highly context-dependent, with nuanced attitudes that shift as patients progress through the care pathway. We can infer from these findings that the acceptance of AI technology within dementia care is a multidimensional construct, shaped by a dynamic interplay of ethical, relational, and practical considerations rather than as a simple binary judgment of “yes” or “no.” This complexity highlights a central principle in digital health: technology integration is not merely a matter of capability but must align with human values, care contexts, and societal expectations.

A particularly salient result was the significant three-way interaction between Stage, Centrism, and Level of Involvement for acceptability. Attitudes toward AI versus human-centric systems were not fixed but evolved according to the patient’s progression through the dementia-care pathway. In the early Preventing-well and Diagnosing-well stages, higher acceptance ratings indicated a cautious openness to AI systems, even when these were designed for higher involvement. This finding suggests that AI technologies can be acceptable if framed appropriately, particularly as tools that enhance autonomy, support preventative measures, or facilitate early diagnosis. This observation is consistent with broader technology acceptance literature, which emphasises that perceived usefulness, transparency, and control are crucial determinants of adoption [33, 34]. For example, framing AI as an assistive partner rather than a replacement may help mitigate resistance and enhance user confidence.

As patients move into later stages (Supporting-well, Living-well, and Dying-well) the data indicate a pronounced preference for fully involved, human-centric systems. In these stages, acceptability shifts sharply away from AI-dominant scenarios, reflecting heightened concerns about trust, ethical responsibility, privacy, and dignity. The Dying-well stage underscores the ethical sensitivity of end-of-life care; participants rated highly involved AI systems as less acceptable, likely perceiving them as intrusive or diminishing human agency. Empathy on behalf of the participants may also play an important role in determining what is acceptable or unacceptable in end-of-life care [35], where scenarios that elicit more empathetic responses may cause participants to express more sensitivity and consideration when deciding what is acceptable. This aligns with prior research highlighting that acceptance of automation is contextually contingent and deeply entwined with perceived risk, moral responsibility, and the potential for loss of control [36]. Notably, the Supporting-well stage showed significant main effects for human centrism and full involvement, whereas in the Living-well and Dying-well stages, these preferences were revealed primarily through interaction effects, emphasising the compounded influence of multiple factors in high-stakes care contexts. This aligns with current technology acceptance theories such as the Unified Theory of Acceptance and Use of Technology (UTAUT2) model which notes that there are multiple factors affecting users’ intentions, which are then also moderated by other characteristics and contexts [33, 37]

Interestingly, moderately involved AI systems were consistently perceived as acceptable across all stages. This suggests that AI can occupy a supportive, collaborative role without triggering concerns over autonomy or ethical compromise. Such findings support the notion of adaptive autonomy in digital health design, where the degree of system involvement can be tailored to individual patient and caregiver needs. Moderately involved AI systems may function as decision-support tools, enhancing efficiency and care quality while maintaining a human-in-the-loop framework that is critical for sensitive and complex healthcare contexts. Examples of this may include risk assessment support derived from passive lifestyle monitoring by an AI system that can flag potential risk factors to clinicians who would then discuss modifications with the patient. In this way, the AI does not control any aspect of the patient’s life, nor does it remove any agency or intrude in any way.

When evaluating likelihood, the patterns were clearer and more consistent. Human-centric systems with moderate involvement were generally seen as the most probable scenario, particularly in later stages of the dementia-care pathway. This links closely to theories of risk perception such as the Risk as Feelings Theory [38] which suggests that scenarios that invoke more empathy during decision making, such as the later, more sensitive stages of dementia care, can override cognitive analysis and distort likelihood ratings. This divergence between perceived acceptability and likelihood highlights an important consideration: the public’s view on the adoption of AI may be malleable, influenced by factors such as perceived risk or personal innovativeness [39] as well as direct characteristics of the AI itself. Such findings have direct implications for digital health policy and implementation strategies, suggesting that any integration of AI must balance innovation with realistic expectations of technology readiness, professional norms, and societal trust.

From a translational perspective, these findings carry substantial implications for developers, clinicians, and policymakers. Technology design should prioritize adaptive, patient-centred approaches, incorporating flexibility in system involvement according to stage-specific needs. Human-in-the-loop models are particularly critical in later stages of care to safeguard autonomy, dignity, and relational trust. Policy frameworks should reinforce ethical safeguards, transparency, and human oversight while encouraging innovation in early diagnosis and preventative care. Moreover, the consistent acceptability of moderately involved AI indicates a practical pathway for implementation. AI may be integrated as a collaborative, augmented tool rather than a fully autonomous replacement, supporting decision-making and easing caregiver workload without compromising patient values. These insights align with broader principles in digital health, emphasising user-centred design, ethical deployment, and context-sensitive automation [40, 41].

Methodologically, the study leveraged vignettes to explore complex, hypothetical care scenarios, providing rich, nuanced data. However, this approach has inherent limitations. Participants responded to imagined scenarios rather than interacting with tangible systems, which may omit critical real-world factors such as usability, interface intuitiveness, system reliability, and affective experience. This limitation underscores the need for future studies to enhance ecological validity, for instance, by incorporating interactive prototypes or Wizard-of-Oz paradigms that simulate real-world interactions with AI systems [42]. Longitudinal designs would also be valuable, tracking changes in public attitudes as AI technologies evolve and as societal exposure to AI in healthcare increases.

Furthermore, the rapidly advancing nature of AI in digital health suggests that public perceptions are likely to shift over time. Cross-sectional snapshots capture contemporary attitudes but cannot account for temporal dynamics, changes in trust, or exposure-driven adaptation. Future research should therefore integrate longitudinal assessments to provide a more comprehensive understanding of technology acceptance trajectories, identifying factors that can positively or negatively influence adoption, trust, and perceived appropriateness over time.

In conclusion, this study demonstrates that public perceptions of AI-integrated dementia care are highly context-dependent and shaped by the interplay of care stage, system centrism, and level of involvement. Human-in-the-loop approaches are consistently preferred in later, more vulnerable stages, while moderately involved AI systems offer a broadly acceptable pathway across the care continuum. These findings carry important implications for technology design, policy development, and clinical practice. With our sample being made up of exclusively 55–64-year-olds, who will be the next generation to utilise technologies such as this, allows this study to highlight the necessity of adaptive, ethically grounded, and patient-centred approaches in digital health interventions, reinforcing that successful AI integration depends on trust, dignity, and the careful negotiation of human and technological roles. By acknowledging the nuanced perspectives of patients, carers, and the wider public, digital health stakeholders can ensure that AI-enhanced dementia care is not only technologically feasible but socially acceptable, ethically robust, and ultimately supportive of human agency across all stages of the care pathway.