Does LLM reliance make nursing students anxious? A 473-person survey cannot establish cause
A Guangdong survey found a small association between self-reported LLM dependence and academic anxiety. Every measure came from one questionnaire, AI use was not recorded, and anxiety or burnout could also increase reliance.
The short version
- What happened
- Researchers invited 500 nursing students in Guangdong Province and analysed complete questionnaires from 473, including students undertaking clinical placements.
- Who it affects
- A nursing student needs clear rules for checking AI-generated clinical material and access to human supervision. A lecturer needs evidence of actual learning and wellbeing—not a questionnaire label treated as a diagnosis.
- What remains unproven
- Cross-sectional convenience sample
Keep in mind
Cross-sectional convenience sample
473 nursing students in one Chinese province
All constructs self-reported in one questionnaire
Selected scale items rather than complete behavioural assessment
No logged AI-use frequency, duration or purpose
No academic, clinical or patient outcome
Source analysis · 1 linked source. Primary analysed source date: 27 October 2026.
Read the full report · 4 min The evidence, methods, implications and limitations
The reported link was small—and direction remains unknown
Nursing students who described themselves as more dependent on large language models also reported slightly more academic anxiety in a peer-reviewed survey published on 27 October. The adjusted path coefficient was 0.138 on standardised self-report scales, with a 95% confidence interval from 0.006 to 0.278 and a p value of 0.048. That is a small, borderline association—not evidence that an LLM caused distress.
Researchers invited 500 students from higher-education institutions and clinical placements in Guangdong Province and retained 473 questionnaires. All participants were still enrolled students; 80.3% were women and almost 90% were aged 18 to 23. Recruitment used teaching staff, placement supervisors, student groups and online channels, so the 94.6% response figure does not make the convenience sample representative of nursing students in Guangdong, China or elsewhere.[1]
Four self-report scales were fitted into one statistical model
The questionnaire used selected items from published five-point scales for LLM dependence, 'AI addiction', academic burnout and academic anxiety. Internal-consistency estimates were high, but the labels are measurements of questionnaire responses, not clinical diagnoses. The study did not log which models students used, how often or for what tasks. It also did not measure grades, knowledge retention, clinical judgement, patient safety or independent performance.
Partial least squares structural equation modelling included age category and sex as predictors of anxiety and used 5,000 bootstrap resamples. The model explained 49.2% of the variation in anxiety scores. Fit evidence was mixed: SRMR was within the authors' reference value, while NFI was 0.812 against a 0.90 reference. Burnout had the strongest direct association with anxiety, while the direct AI-addiction-to-anxiety path was not significant. The authors also reported indirect paths through burnout. Because all four measures were collected once from the same people, those paths are statistical decompositions, not proof that dependence progressed into addiction, then burnout, then anxiety.[1]
The distinction that matters
What the survey can—and cannot—separate
The same one-time pattern can fit several different explanations.
Reported association
Higher LLM-dependence score
Small higher academic-anxiety score
Reverse direction
Anxiety or burnout already present
Students may seek more AI assistance
Shared causes
Workload, support or prior distress
Could raise both sets of scores
Support should target wellbeing and verifiable practice, not a presumed cause
The practical signal is to ask better questions. A student may rely on an LLM because they are already anxious, exhausted or short of supervision; workload or prior distress may raise every score. Conversely, uncritical dependence could weaken confidence when the tool is unavailable. This design cannot separate those explanations, and residual confounding remains substantial because AI-use patterns, workload and placement stage were not modelled in detail.
Nursing programmes can still teach source checking, disclosure, privacy and the limits of generated answers while preserving access to mentoring and mental-health support. They should not treat a high questionnaire score as addiction or restrict tools on the assumption that doing so will reduce anxiety. Stronger evidence would follow students over time, compare competing causal orderings, include usage logs and independent performance measures, and test support interventions. The authors reported no funding or commercial conflicts. Their full article became accessible after an earlier accepted-abstract release, allowing the methods and qualifications to be checked; the source date remains 27 October.[1]