The research team, led by Kiran Abdul Sattar of the Institute of Family Medicine at Jinnah Sindh Medical University, surveyed 310 Bachelor of Medicine and Bachelor of Surgery students spanning every academic year. The scale of AI adoption they documented is striking. Ninety-four point one nine percent of participants reported having used ChatGPT, and 90.97 percent said they used artificial intelligence primarily for academic assistance. In a cohort of future physicians, nearly universal engagement with generative AI is no longer a prediction — it is the present reality, and understanding what drives that engagement has become a central question for medical educators worldwide.

To untangle the psychological drivers of AI dependency, the researchers measured four latent constructs using validated instruments: self-efficacy, assessed with six items; academic stress, also six items; performance expectation, captured with four items; and procrastination, measured with the Academic Situational Procrastination Scale. AI dependency itself was quantified with a five-item scale. Rather than treating these variables in isolation, the team employed Partial Least Squares Structural Equation Modeling, or PLS-SEM, a statistical framework that estimates relationships among unobserved latent variables through their measured indicators. The analysis was performed using SmartPLS 3.3.3, allowing the researchers to test how all four psychological factors simultaneously shaped AI dependency.

The structural model told a coherent story, though not the one many might expect. Together, the four constructs explained 24.4 percent of the variance in AI dependency — an R² of 0.244, respectable for a model of human behavior in a single cross-sectional sample. Within that model, procrastination emerged as the dominant force, showing the strongest positive association with AI dependency at a standardized path coefficient of β = 0.427 with a p-value of 0.001. In effect size terms, this corresponded to an f² of 0.239, a medium effect — meaning procrastination was not merely statistically detectable but substantively important in predicting which students came to depend on AI.

The interpretation is intuitively compelling. A student who habitually delays studying, defers assignments, and postpones exam preparation faces a moment when the work simply must be done, and generative AI offers an immediate, frictionless shortcut. The chatbot does not judge, does not require scheduling, and does not demand the sustained effort that independent learning does. In that light, AI dependency may function less as a technological phenomenon and more as a symptom of impaired self-regulation — a digital extension of the same executive-function weaknesses that produce procrastination in the first place. For medical educators, this reframing matters enormously, because interventions targeting time management and self-discipline may do more to promote healthy AI use than any policy restricting the technology itself.

Two other factors showed smaller but statistically significant positive associations with AI dependency. Performance expectation — a student’s belief that using AI will improve their academic results — carried a path coefficient of β = 0.143 (p = 0.013), a small effect. Self-efficacy, the confidence students have in their own capabilities, showed a small positive association as well, at β = 0.128 (p = 0.046). The self-efficacy finding is the most counterintuitive of the study. Classic self-regulation theory might predict that confident students would need AI less; instead, students who believed in their own competence appear to have felt freer to experiment with AI tools, integrating them into their workflow without anxiety about losing control. Confidence, in this reading, enables exploration rather than guarding against dependence.

The null result may prove the most talked-about finding of all. Academic stress showed essentially no relationship with AI dependency, with a path coefficient of β = -0.007 and a p-value of 0.935 — about as close to zero as statistics allows. The popular narrative that overwhelmed, anxious students turn to chatbots as a coping mechanism found no support in this sample. Whatever drives students toward AI, the authors conclude, it appears to be rooted in self-regulatory and motivational factors rather than in distress. That distinction carries practical weight: wellness programs and stress-reduction initiatives, however valuable for student mental health, are unlikely on their own to change how medical students use artificial intelligence.

The stakes of getting this right are unusually high in medical education. Overreliance on AI during the formative years of training threatens precisely the capacities that define a competent physician — critical thinking, clinical reasoning, and the ability to solve problems when no algorithm is available. A medical graduate who has outsourced the struggle of learning may struggle to tolerate the ambiguity of real patients, where no prompt produces a definitive answer. At the same time, AI used well can accelerate learning, personalize study, and free students to focus on higher-order reasoning. The difference between these outcomes lies not in the technology but in the psychological profile of the student using it, which is exactly what this study helps map.

The authors emphasize that their findings point toward a specific intervention strategy: programs focused on self-regulation and time management could support balanced AI integration in medical education while preserving independent learning skills. Teaching students to recognize procrastination triggers, structure their study time, and set boundaries around AI use may address the root cause of dependency rather than its symptoms. Such approaches align with a broader shift in educational psychology, which increasingly treats self-regulated learning as the master skill that determines whether new technologies amplify or erode genuine competence.

As with any cross-sectional study, the usual caveats apply. The data capture a single moment in time at a single institution, and associations — however precisely quantified — cannot establish causation. It remains possible that AI dependency itself fosters procrastination, or that both share an unmeasured underlying cause. Still, with 94 percent of medical students in this sample already using ChatGPT and the model accounting for a meaningful share of the variance in dependency, the study provides a rigorous empirical foundation for a conversation that medical schools everywhere can no longer postpone. The generation now filling anatomy labs will practice medicine alongside artificial intelligence for their entire careers; the urgent task, this research suggests, is not to keep them away from these tools but to ensure that discipline, not delay, governs how they use them.

Subject of Research: Psychological predictors of AI dependency among medical students in Pakistan

Article Title: Association of self-efficacy, stress, performance expectation, and procrastination with AI dependency among medical students at a public sector university in Pakistan

Article References: Abdul Sattar, K., Fatima, S., Ali, S. K., Zafar, H., & Abbas, S. M. (2026). Association of self-efficacy, stress, performance expectation, and procrastination with AI dependency among medical students at a public sector university in Pakistan. BMC Medical Education. https://doi.org/10.1186/s12909-026-10535-w

Image Credits: AI Generated

DOI: 10.1186/s12909-026-10535-w

Keywords: artificial intelligence, medical students, procrastination, self-efficacy, academic stress, performance expectation, AI dependency, medical education, PLS-SEM, ChatGPT, self-regulation, Pakistan

The study used a grounded theory approach, following the methodology of Strauss and Corbin, to build an explanatory model from the ground up rather than testing pre-existing hypotheses. Between August and October 2025, the researchers conducted semi-structured, in-depth interviews with 23 master’s nursing students drawn from multiple provinces in China. Participants were recruited through purposive and theoretical sampling, meaning the team deliberately selected interviewees who could illuminate the emerging theory and continued adding interviews until the model stopped generating new insights. The first 21 interviews fed directly into theory development, while two additional interviews were used to assess theoretical saturation. No new categories or shifts in the relationships between categories appeared in those final two interviews, giving the researchers confidence that the model had stabilized.

Analysis followed the classic three-stage coding sequence of grounded theory. The interview transcripts were first broken down through open coding, in which discrete concepts and events were labeled and compared. Axial coding then reassembled those fragments by identifying relationships among categories, linking conditions, actions, and consequences. Finally, selective coding distilled the material around a central phenomenon that organizes the entire model. The result is a process model describing how generative artificial intelligence becomes embedded in postgraduate learning and research routines, and the picture it paints is more nuanced than simple headlines about dependency might suggest.

At the heart of the model is what the researchers call a forward shift in the problem-solving pathway. Traditionally, a graduate student facing a research problem might first attempt to reason through it, consult textbooks or journals, seek advice from a supervisor, and only then, perhaps, try an AI tool. The study suggests this sequence is being inverted. Students now begin with the chatbot, asking it to summarize literature, draft text, interpret data, or generate ideas before they have engaged deeply with the problem themselves. Generative artificial intelligence, in other words, is migrating from the end of the pipeline to its front door, and that migration changes not just efficiency but the very structure of how students think.

What drives this shift? The researchers identified four contextual conditions that push students toward AI-first behavior. The first is impeded task progress: when a problem stalls, the temptation to break the deadlock with a machine-generated answer grows. The second is time pressure, a constant in clinical and academic life alike, which makes the slow work of independent inquiry feel like a luxury. The third is insufficient support, meaning that when supervisors, peers, or institutional resources fall short, the chatbot fills the vacuum. The fourth is ambiguous norms, the absence of clear institutional guidance about when and how AI use is acceptable. Together, these conditions create an environment in which turning to AI first is not a moral failing of individual students but a rational adaptation to structural circumstances.

Once initiated, the pattern becomes self-consolidating through what the authors describe as pathway reconstruction. This involves two intertwined behaviors: artificial intelligence-first use, in which the tool becomes the default entry point for new tasks, and cognitive task outsourcing, in which components of thinking, such as synthesizing evidence, structuring arguments, or refining language, are delegated to the machine. Each outsourcing decision reduces the friction of the next one, and over time the pathway itself is rebuilt around the technology. The student who once struggled through a literature review alone now cannot remember how they ever managed without one.

Crucially, the model identifies two reinforcement mechanisms that stabilize this pattern: Functional Benefits and Emotional Benefits. Functional benefits are the practical payoffs, including saved time, improved output quality, and smoother completion of demanding assignments. Emotional benefits are subtler but no less powerful, encompassing reduced anxiety, a sense of companionship with an always-available assistant, and relief from the isolation that often accompanies postgraduate research. Because both types of reward arrive immediately and reliably, while the potential costs, such as erosion of independent analytical skill, are delayed and diffuse, the behavior is reinforced on an uneven playing field. The brain’s learning machinery, responsive to immediate reward, tilts toward repetition.

The study does not portray students as oblivious to these risks. On the contrary, participants developed awareness that something might be lost when thinking is outsourced, worrying about accuracy, originality, academic integrity, and their own long-term competence. Yet the researchers found that this awareness rarely translated into stopping. Instead, students maintained the pattern through two mechanisms: delayed regulation, in which plans to cut back or verify AI output were perpetually postponed, and stage-based rationalization, in which use was justified according to the phase of a project, for example by telling oneself that AI assistance was acceptable for brainstorming but not for final writing, a boundary that proved porous in practice. The result is a form of negotiated coexistence, an uneasy but durable accommodation between the student’s critical judgment and the tool’s constant availability, rather than a clean decision to discontinue.

The implications for health professions education reach well beyond nursing. Graduate students in every discipline face the same four conditions: stalls, deadlines, thin support networks, and unclear rules. If the forward shift observed here is a general phenomenon, then the moment of educational intervention matters enormously. The authors argue that responses should move upstream, shaping when and how generative artificial intelligence is used rather than merely policing what students submit at the end. Practically, that means strengthening research competence so students have the skills to engage problems independently, reinforcing human support systems so that the chatbot does not become the only responsive mentor in a student’s life, and embedding artificial intelligence literacy alongside critical verification training so that students learn to interrogate machine output rather than absorb it.

The study also carries a warning embedded in its optimism. A negotiated coexistence is not a stable equilibrium; it is a truce that can tip either way depending on the environment. Institutions that provide clear norms, timely feedback, and robust mentorship could tip it toward reflective, bounded use, in which AI amplifies rather than replaces thinking. Institutions that remain silent could tip it toward habitual outsourcing, producing clinicians who can navigate a chatbot fluently but struggle to reason through an unfamiliar clinical problem without one. The grounded theory model built from these 23 interviews offers educators a map of the terrain: the conditions that trigger AI-first behavior, the pathway that reconstructs itself around the tool, and the reinforcements that lock the pattern in place. Reading that map early, the authors suggest, may be the best chance education has to shape a technology that is already reshaping its students. The research was funded by the Zhejiang Provincial Health Industry Science and Technology Plan and approved by the Ethics Committee of Sir Run Run Shaw Hospital, and it is published open access for the wider education and policy community to build upon.

Subject of Research: Generative AI-first problem solving behavior among master's nursing students

Article Title: When GenAI becomes the default: a grounded theory study of GenAI-first problem solving among master’s nursing students

Article References: When GenAI becomes the default: a grounded theory study of GenAI-first problem solving among master’s nursing students. (n.d.). https://doi.org/10.1186/s12909-026-10427-z

Image Credits: AI Generated

DOI: 10.1186/s12909-026-10427-z

Keywords: generative artificial intelligence, nursing education, grounded theory, postgraduate students, AI dependency, problem solving, medical education, qualitative research, large language models, cognitive outsourcing, AI literacy, China