The study was conducted in mainland China between November 2025 and March 2026. In contemporary Chinese health care, the practice scope of registered clinical nurses is inherently multidimensional, encompassing direct bedside care, clinical administration/quality improvement, and practice-based continuing education and scholarly research. Consequently, frontline nurses interact with a heterogeneous spectrum of AI applications, including embedded clinical AI systems—institutionally integrated, algorithm-driven tools designed for direct patient care, risk monitoring, and operational management (eg, early warning scoring systems and clinical decision support systems)—and generative AI tools driven by large language models (LLMs; eg, DeepSeek [Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co, Ltd], Doubao [Beijing Chuntian Zhiyun Technology Co, Ltd], or ChatGPT [OpenAI]). Participants were recruited across 8 general hospitals (tertiary and secondary) and 1 affiliated primary community health center in 4 municipalities/provinces (Beijing, Shandong, Guangdong, and Sichuan). The community facility was included as an integrated primary-care arm of a regional tertiary health alliance to capture grassroots AI adoption. In this study, clinical nurses were defined broadly as registered nurses providing direct patient care, including those working in primary care settings, to reflect the full spectrum of AI deployment across different tiers of the health care system.

A purposive sampling strategy framed by a maximum variation matrix was adopted to ensure broad coverage across key demographic and clinical dimensions. Inclusion criteria were as follows: (1) registered nurses holding a valid license; (2) currently engaged in direct clinical care with at least 1 year of clinical experience; and (3) having active, hands-on experience with at least one AI-based system—such as embedded clinical AI systems or generative AI tools—in their daily workflow for ≥3 months. Exclusion criteria included nurses exclusively in administrative roles without direct AI clinical application. Participants were recruited via professional nursing association networks and hospital department announcements across participating institutions. A total of 37 eligible nurses were approached; 2 declined due to clinical shift conflicts and 1 withdrew before the interview for personal reasons, yielding a final sample of 34 (completion rate: 91.9%) participants. To construct the sampling matrix, we initially recruited nurses varying in age, professional titles, educational backgrounds, and hospital tiers, aiming to capture the widest possible spectrum of routine AI interaction experiences. Building upon this maximum-variation framework, we then deliberately applied intensified sampling to 2 subgroups of particular theoretical interest: tertiary hospital nurses (n=26, 76.5%) and male nurses (n=13, 38.2%). Tertiary teaching hospitals serve as the primary innovation hubs for advanced AI deployment in China, and oversampling this group allowed us to capture dense, firsthand accounts of complex AI system integration. Similarly, male nurses—a demographic minority within the profession—were intentionally oversampled to ensure that their perspectives on technostress and human-AI interaction were not marginalized in the final analysis. This layered sampling design prioritizes both the breadth of routine experiences and the depth of theoretically critical cases. Enrollment continued until data saturation was achieved, defined as the point where no new codes or themes emerged from the last 3 consecutive interviews.

Interviews were conducted by 2 qualitative researchers (XZ and JF) trained in advanced qualitative content analysis. Neither interviewer had prior personal, clinical, or administrative relationships with the participants. To ensure reflexivity, both interviewers maintained reflexive journals throughout the study to acknowledge, document, and bracket preexisting assumptions regarding the JD-R model (the detailed interview guide is provided in Multimedia Appendix 2). A total of 14 interviews were conducted face-to-face in private, quiet hospital conference rooms, while 20 were conducted via a secure online video conferencing platform (Tencent Meeting [Tencent]) to accommodate geographically distant participants across provinces. All interviews were conducted in Mandarin Chinese and lasted between 20 and 60 minutes. Field notes were recorded during and immediately after each interview to capture nonverbal cues, vocal hesitations, and emotional responses, which were subsequently integrated into transcript familiarization.

Audio recordings were transcribed verbatim in Mandarin within 48 hours. Selected verbatim quotes were translated into English by a bilingual nursing researcher and back-translated by an independent native English academic editor to maintain conceptual and cross-language equivalence. Data analysis was managed using NVivo 20 software (QSR International) and followed the DQCA systematic steps: (1) preparation and familiarization—2 researchers independently read the transcripts multiple times alongside field notes to achieve deep immersion in the data; (2) deductive coding—initial coding categories were generated deductively from predefined JD-R constructs, and transcripts were coded by assigning text passages to these predefined categories; (3) inductive coding—data that could not be categorized into the initial JD-R model were assigned new codes inductively using explicit rules; for instance, statements describing technology-driven peer hyper-competition and artificially escalated management performance baselines were inductively coded under digital involution. To ensure coding reliability, 2 coders independently coded an initial batch of 6 transcripts; (4) theme refinement, disagreement resolution, and rigor—the research team held regular consensus meetings to compare codes, resolve discrepancies, and abstract the codes into main themes and subthemes. Discrepancies were discussed until 100% agreement was reached. Theoretical saturation was audited through a cumulative codebook; code generation stabilized at Participant 31, and the final 3 consecutive interviews produced zero new codes or subthemes, confirming data saturation.

Through DQCA guided by the JD-R framework, we identified 3 overarching domains comprising 9 main themes and 24 subthemes that illustrate the complex double-edged sword effect of AI application in clinical nursing (Figure 1). Specifically, the gain path (job resources) demonstrates the empowering potential of AI by facilitating a workflow efficiency leap, providing clinical cognitive empowerment, and fostering professional capital appreciation. Conversely, the drain path (job demands) highlights the hidden and continuous costs of AI integration, conceptualized as perceived cognitive impediment, relational attrition that challenges humanistic empathy, and digital involution driven by performance inflation and competitive pressures. Furthermore, the interplay between these perceived gains and demands is contextualized by key perceived buffering conditions, specifically nurses’ proactive coping strategies, strict professional boundary demarcation, and the presence of a supportive organizational support architecture. The detailed thematic coding tree is presented in Multimedia Appendix 5.

AI emerged as a robust job resource for clinical nurses, triggering a motivational process. By interacting with various AI systems, nurses extracted perceived empowerment, which manifested across 3 distinct themes: workflow efficiency leap, clinical cognitive empowerment, and professional capital appreciation.

Nurses perceived generative AI tools and integrated clinical decision support systems as an on-demand knowledge repository that was particularly crucial during emergencies or unfamiliar cases. The systems provided rapid access to specialized clinical knowledge, complex pharmacological data (eg, drug interactions), and evidence-based guidelines. This instant access effectively masked experience deficits among junior nurses. As nurse N6 noted: “I use AI to organize and recall anatomical positions and functions; after sorting it out, I have a general idea of the clinical purpose and direction” (see Multimedia Appendix 5 for the full dataset).

When confronting cognitive bottlenecks, such as complex case analyses or practice-based academic reflection, generative AI tools served as a cognitive coach. They provided structural frameworks, established logical connections between disparate symptoms, and assisted in structuring care plans. Participants emphasized that AI stimulated secondary reflection and expanded thinking dimensions. As nurse N8 described: “When I'm stuck on how to approach something, I might need AI to point the way. For example, when writing a report for the first time, I needed AI to give me a logical pivot. It also told me how to structure the logic of the text” (see Multimedia Appendix 5 for the full dataset).

Embedded clinical AI systems (eg, early warning scoring models and closed-loop smart infusion pumps) acted as vigilant partners in direct patient care. By monitoring microphysiological trends, predictive models offered early warnings for clinical deterioration, effectively reducing continuous vigilance fatigue. Automated assistance in high-risk procedures was perceived to enhance clinical safety monitoring and foster technological trust. As nurse N23 noted: “The intelligent early warning system mainly monitors patients’ vital signs... If any data is abnormal, it immediately alerts us so we can respond quickly” (see Multimedia Appendix 5 for the full dataset).

Delegating routine tasks to both generative AI tools and embedded clinical AI systems generated significant positive emotional energy. Nurses reported a heightened sense of control over their work pace and a reduction in digital anxiety. The reliable, on-demand nature of AI assistance served as a psychological buffer during overwhelming workloads. As nurse N4 shared: “Even though it’s complicated, we still have AI, right ?... It boosts my confidence in my work. With that confidence, I’m less likely to be bothered by difficulties” (see Multimedia Appendix 5 for the full dataset).

Generative AI tools functioned as accessible tutors that lowered thresholds for acquiring new digital competencies. Nurses used AI to bridge knowledge gaps, learn nursing research methodologies (eg, SPSS Statistics [Version 20.0; IBM Corp] operations and statistical workflows), and rapidly acquire digital content creation skills for patient education (eg, short video scripts). As nurse N33 shared: “It forced me to learn data analysis. Before, I just followed orders; now, I interpret AI-generated reports and analyze chronic disease trends in my district... It transformed me from an executor into a manager” (see Multimedia Appendix 5 for the full dataset).

A value-driven expectation emerged from the data. By absorbing trivial administrative and physical tasks, AI implicitly redirected nurses’ finite attention back to the essence of nursing: direct patient care and complex emotional labor. Participants expressed future aspirations for advanced AI—such as intelligent ward-roaming robots or automated health education loops—to fully eradicate manual drudgery, thereby allowing them to reclaim their professional identity as empathetic caregivers. As nurse N1 stated: “It can take over more mechanical tasks, such as documentation or monitoring, allowing nurses to focus more on patient psychological care and complex clinical management” (see Multimedia Appendix 5 for the full dataset).

While AI promised cognitive relief, the unreliability and complex operational logic of current algorithms generated new dimensions of mental load, transforming nurses from care providers into algorithmic supervisors.

The democratization of consumer-accessible generative AI tools meant that patients and families frequently used these systems for self-diagnosis. Frontline nurses were forced to absorb substantial amounts of new interpretive labor—deconstructing AI-generated misinformation, managing patient anxiety, and rebuilding trust. As nurse N30 stated: “Patients and families create challenges by consulting unreliable online sources... This undermines our health education efforts” (see Multimedia Appendix 5 for the full dataset).

Participants harbored deep apprehensions regarding technology-mediated alienation caused by standardized algorithmic care pathways and automated communication scripts. Nurses feared a loss of humanistic care, perceiving algorithmic reliance as an emotional barrier that reduced holistic nursing to rigid, machine-dictated transactions. As nurse N12 noted, “Compared to humans, it’s deficient in humanistic care—a cornerstone of our profession” (see Multimedia Appendix 5 for the full dataset).

Embedded clinical AI systems (eg, performance-tracking algorithms and automated shift scheduling) and rapid technology iterations fostered professional insecurity. Senior staff facing the digital divide experienced anxiety regarding displacement and feeling managed by algorithms, which eroded their clinical agency. As nurse N23 described: “My greatest fear is AI wholly supplanting nursing roles, eliminating job opportunities” (see Multimedia Appendix 5 for the full dataset).

The accessibility and speed of generative AI tools lowered the threshold for complex administrative and quality-improvement deliverables (eg, report synthesis or presentation design), which paradoxically triggered performance inflation. Management implicitly raised baseline standards, demanding perfect deliverables and forcing nurses into a cycle of continuous overperformance. As nurse N33 illustrated: “Management increasingly judges quality by numbers rather than substance. Quantity gradually overshadows quality focus, and performance benchmarks keep escalating.” (see Multimedia Appendix 5 for the full dataset).

The accessibility of AI-generated content fostered an environment of intensified peer comparison. Nurses felt compelled to produce flawlessly polished materials to maintain their professional standing, leading to a state of competitive perfectionism. This subtheme highlights how AI-driven high standards became the new minimum requirement, forcing individuals to invest excessive cognitive effort into refining AI outputs to avoid being outperformed by colleagues. As nurse N23 noted: “I have to spend time after work every day learning related knowledge, and I feel a lot of pressure. Before, when we wrote health education materials, ‘good enough’ was fine. But now with AI, the requirements keep getting higher, the quality keeps improving — it’s very competitive” (see Multimedia Appendix 5 for the full dataset).

Contrary to labor-saving promises, findings reveal a paradoxical expansion of invisible digital labor across both technology classes. This includes troubleshooting embedded system errors, performing redundant dual-entry processes between legacy systems and AI platforms, and debugging generative AI outputs. As nurse N7 noted: “Nurses haven’t had responsibilities reduced by AI; instead, we’ve gained the added duty of monitoring AI for malfunctions” (see Multimedia Appendix 5 for the full dataset).

Nurses did not remain passive recipients of technological outputs; instead, they engaged in highly proactive behavioral crafting to harness AI’s use while neutralizing its risks.

Participants conceptualized both generative AI tools and embedded clinical AI systems strictly as high-capacity subordinate assistants or information aggregators, vehemently rejecting AI as an autonomous decision-maker. Recognizing AI’s lack of genuine empathy insulated nurses from replacement anxiety. As nurse N26 noted: “A textual assistant. For reference and support—nothing more” (see Multimedia Appendix 5 for the full dataset).

Nurses preserved clinical autonomy by insisting on the irreplaceable nature of physical touch and empathetic care. In clinical decision-making, they adopted a “think first, consult second” paradigm, ensuring human reasoning framed the clinical strategy. As nurse N6 stated: “I adhere to my own logic—AI won’t override my reasoning. Critical thinking is non-negotiable. With independent thought, AI becomes a tool, not a puppeteer ”(see Multimedia Appendix 5 for the full dataset).

Participants expressed a need for hospitals to transcend fragmented usage by deploying unified, institutionally sanctioned, vertically integrated medical AI platforms (eg, specialized medical LLMs trained on validated clinical guidelines). Centrally provided platforms lower cognitive thresholds and alleviate security concerns. As nurse N3 noted: “Hospitals could consolidate databases... Institution-led AI platforms would streamline retrieval” (see Multimedia Appendix 5 for the full dataset).

Nurses emphasized the need for structured organizational training focusing on underlying algorithmic logic (for embedded clinical AI systems) and prompt structuring and error avoidance (for generative AI tools), backed by clear institutional policies and ethical boundaries. As nurse N3 described: “If hospitals adopt AI, they should develop structured teaching programs” (see Multimedia Appendix 5 for the full dataset).

Consistent with the motivational pathway of the JD-R model, our findings validate that AI serves as a potent perceived job resource, initiating empowerment across 3 progressive dimensions: operational workflow, clinical cognition, and professional capital. First, regarding operational execution, our findings corroborate historical health information technology literature that emphasizes the physical and temporal benefits of automation [25]. We demonstrate a substantial workflow efficiency leap, where disaggregated technology classes yield distinct operational resources: generative AI tools streamline documentation automation and literature retrieval, whereas embedded clinical AI systems enhance shift scheduling and team synergy. These observations converge with recent qualitative scholarship, highlighting AI’s capacity to alleviate administrative burdens and optimize workflow management [6,21,26]. Beyond routine automation, AI provides critical clinical cognitive empowerment. Specifically, embedded clinical AI systems (eg, early warning scores) act as safety sentinels, whereas generative AI tools function as on-demand cognitive tutors that bridge knowledge gaps and scaffold clinical reasoning. This positions AI as an active adjunct to human cognition [27-29], directly boosting nurses’ self-efficacy [30,31] and rendering complex tasks manageable [32]. Ultimately, this operational and cognitive synergy fosters professional capital appreciation. By delegating mundane tasks to algorithms, nurses restore emotional resources, acquire cross-boundary digital skills, and experience a “core value regression”—redirecting finite attentional capacity back to the irreplaceable, humanistic core of patient care.

Despite the optimistic narrative surrounding AI empowerment, our study uncovers a perceived health impairment process driven by novel digital demands. Consistent with Chuang et al [33], who demonstrated the dual impact of AI on work and life well-being under the JD-R model, our finding reinforces the model’s suitability for unpacking the paradoxical effects of AI in clinical nursing contexts. A critical finding in this study is the manifestation of an AI productivity discrepancy rooted in cognitive impediment, which manifests through distinct technological mechanisms. While AI theoretically accelerates task completion, this temporal advantage is frequently neutralized by hidden cognitive loads. For generative AI tools, cognitive load stems from the threat of algorithmic hallucinations and omissions, forcing nurses into tedious prompt engineering and multisource verification; for embedded clinical AI systems, cognitive strain is driven by hypersensitive false alerts and alert fatigue [34]. As our data suggests, cognitive workload has not disappeared; it has merely shifted from primary content creation to secondary content verification. This verification pressure is further compounded by the psychological risk of automation bias (uncritical overreliance on AI outputs) versus the cognitive friction of overriding false alerts.

Furthermore, an important finding is that within the highly competitive environment of contemporary health care, the widespread accessibility of generative AI tools has inadvertently driven performance inflation. As AI lowers the technical barriers for producing polished administrative or quality-improvement work, nursing management implicitly raises baseline expectations. While prior literature often conceptualizes technological stress as individual operational difficulty, our findings uncover a novel sociotechnical phenomenon—digital involution—wherein frontline nurses are locked into a cycle of competitive perfectionism and invisible workload expansion (eg, dual-entry validation and prompt debugging). When combined with the interpretive labor required to bridge the gap between AI misinformation and patient understanding amid ambiguous legal responsibility frameworks, these demands constitute a substantial source of technostress. If left unaddressed, this invisible surge in task intensity, coupled with professional alienation and the fear of occupational displacement, threatens the long-term sustainability of the nursing workforce. These findings suggest that without organizational interventions to decouple AI efficiency from ever-increasing performance quotas, the double-edged sword of AI may inadvertently accelerate burnout rather than alleviate it.