Academic pressure and trust in artificial intelligence were both associated with greater reported reliance on generative AI in a 2026 survey of 229 journalism students at two Chinese universities.
The study also found that students with stronger AI literacy reported less dependence after other factors in the analysis were taken into account. However, the results do not show that stress causes students to rely on AI, or that training in AI literacy will reduce reliance.
What the survey found
Academic stress had the strongest link with reported AI dependence among the factors tested, with a statistical significance level of p < 0.001. Students who trusted AI more also reported greater dependence, even after an instructor had explained AI errors and risks.
Classmates’ and teachers’ use of AI was associated with greater trust in AI, but it did not have a significant direct link with dependence. Social influence showed a similar pattern: it was linked with greater trust, while its direct relationship with dependence was not significant.
The researchers used structural equation modeling to examine several connected relationships while accounting for the others in the model. The indirect pathway from social influence to dependence through trust was statistically significant.
What “dependence” meant in this study
The researchers distinguished dependence from addiction. Dependence referred to handing thinking and decisions to AI and finding it difficult to reduce reliance. The questionnaire included statements about feeling anxious when forbidden to use AI and being unable to persist in cutting back on AI for learning.
These scores were not a clinical diagnosis. The survey measured students’ reports about their behavior and attitudes rather than using AI-use logs or observing classroom work.
How the study was conducted
The participants were undergraduates taking a required research methods course for journalism and communication students at one public university in Fujian province and one in Guangdong province. Of 266 invited students, 229 usable responses remained after incomplete, rushed and straight-line responses were removed.
The class covered the use of Doubao, DeepSeek, Kimi and ChatGPT for literature searches, paper summaries and proofreading. Instruction also addressed fabricated data, citation rules and AI detection.
Two weeks later, students completed an in-class literature review for a midterm grade. AI use was optional but had to be marked, and the survey followed immediately afterward. Nearly half of surveyed students, 48.1%, said they used AI tools at least once a day. Another 39.7% used them every few days to weekly, while 57.2% typically spent 30 minutes to one hour per session.
Why the findings need caution
This was a one-time, self-report survey. Stress, trust and dependence were measured at the same time, so the study cannot establish which came first or demonstrate cause and effect.
The sample also limits how broadly the results can be applied. It came from one course at two universities, was 81.2% women and included too few men for meaningful group comparisons. All participants had recently received AI instruction for the assignment, which may have influenced their answers.
The researchers modified the model after examining initial fit results, and some coefficients differed between the study’s text and tables. The source therefore indicates that the direction of the relationships is more dependable than their exact size.
What students described
Interviews with eight students, together with open-ended survey responses, highlighted efficiency, peer influence, confidence and unclear rules. Students described using AI to manage heavy workloads and tight deadlines. Some said they combined and rewrote outputs from several tools to avoid detection.
Students also reported uncertainty about what level of AI use was non-compliant. They identified fabricated data and difficult-to-verify content as major problems, while fewer mentioned possible effects on their own thinking or independent learning.
Academic self-efficacy was not associated with stress overall. Among students who considered AI very useful, however, higher self-efficacy was associated with lower stress.
The wider evidence
Other studies described in the source point in a similar direction but have comparable limitations. A survey of 300 university students in South Korea linked academic stress and performance expectations with lower academic self-efficacy and greater AI dependency among ChatGPT users.
A two-wave survey of 452 Chinese university students found that the relationship between low need for cognition and problematic AI use was stronger under high academic stress and weaker among students with higher AI literacy.
In a separate two-wave study of 3,843 Chinese adolescents, the proportion showing AI dependence rose from 17.14% to 24.19%. Earlier mental health problems predicted later AI dependence in that study, but AI dependence did not predict later mental health problems.
Together, these findings suggest that emotional strain, academic demands and confidence in evaluating AI may be relevant to how students use these tools. They do not establish that stress or trust alone produces dependence, and whether AI-literacy training changes behavior would require a trial.
AI tools were used to assist with the preparation of this article.
Leave a Reply