Does using AI improve university students’ mental health?
A survey of 390 students in Bangladesh linked more frequent and more useful AI-supported learning with lower reported cognitive overload and better self-reported mental health. Because the study was cross-sectional and self-reported, it cannot show that AI caused either change.
The short version
- What happened
- The study surveyed 390 undergraduate and postgraduate students with prior academic AI experience at five purposively selected Bangladeshi universities, using non-probability recruitment and partial least-squares structural equation modelling.
- Who it affects
- Students may experience AI as helpful scaffolding or as another demanding system to learn and monitor.
- What remains unproven
- The cross-sectional design measured AI use, overload and mental health at one time and cannot establish temporal or causal direction.
Keep in mind
The cross-sectional design measured AI use, overload and mental health at one time and cannot establish temporal or causal direction.
All central variables were self-reported, increasing the risk of common-method, recall and social-desirability bias.
The sample comprised 390 students in Bangladesh and may not represent other institutions, countries or students without reliable digital access.
The study did not test a specific AI system, actual usage logs, objective learning outcomes or clinical mental-health outcomes.
PLS-SEM evaluates a specified association model but cannot eliminate unmeasured confounding or prove mediation as a real mechanism.
Source analysis · 1 linked source. Primary analysed source date: 8 October 2026.
Read the full report · 8 min The evidence, methods, implications and limitations
The direct answer: the survey found an association, not an improvement caused by AI
Students who reported using AI more often for academic work, and who viewed it as more useful, also reported lower cognitive overload and better mental health in a new Bangladesh survey. The researchers collected responses from 390 undergraduate and postgraduate students at public and private universities and analysed the proposed relationships with partial least-squares structural equation modelling, or PLS-SEM, using SmartPLS 4 and bootstrapping.
The result does not demonstrate that AI improved anyone's mental health. Exposure and outcome were measured through the same questionnaire at one point in time. Students who already feel less overwhelmed may be more confident experimenting with AI, while students with stronger institutional support may both access useful tools and report better wellbeing. Reverse causation, selection effects and common-method bias remain credible explanations. The defensible finding is a pattern of self-reported associations that merits longitudinal and experimental testing.[1]
Who was surveyed and what the questionnaire measured
The target population was undergraduate and postgraduate students in Bangladesh who had previously used AI tools for academic work. Researchers purposively selected five public and private universities to capture different institutions, locations and exposure to digital learning, then recruited respondents through non-probability sampling. The final structural model used 390 responses. That denominator supports the chosen statistical analysis but does not make the sample nationally representative; students who were reachable, willing and already experienced with AI may differ from those with limited connectivity, little confidence or no access.
Participants rated five-item scales covering perceived usefulness, frequency of academic AI use, cognitive overload, digital self-efficacy and student mental health. Experts reviewed the instrument, and a separate pilot with 120 students produced acceptable internal-consistency results and a five-factor structure. Even with that groundwork, the measures remain perceptions: frequency was not captured from usage logs, learning was not tested objectively and wellbeing was not assessed with a clinical instrument such as WHO-5 or DASS-21.[1]
What the model tested
The study examined two aspects of AI-supported learning: perceived usefulness of AI tools and frequency of using them for academic activities. It related those measures to student mental health and tested cognitive overload as a mediator—a possible pathway connecting AI use to wellbeing. It also tested digital self-efficacy, or students' confidence in their digital capabilities, as a moderator that might make the relationship stronger or weaker.
Both perceived usefulness and use frequency were positively associated with the study's self-reported mental-health measure. Both were also associated with lower cognitive overload, while greater overload was associated with poorer mental health. The structural model indicated partial mediation, meaning the estimated relationship with mental health operated partly through the overload measure and partly through other modelled paths. Digital self-efficacy strengthened the association between perceived usefulness and mental health, but its moderation of the frequency–mental-health relationship was not statistically significant.[1]
Why mediation language needs special care in a cross-sectional survey
A mediation model can describe whether the observed covariance is consistent with a proposed pathway. It cannot establish the temporal order required for a causal mechanism when all variables are measured together. For cognitive overload to mediate an actual effect, AI use would need to precede a change in overload, which would then precede a change in mental health. This survey did not observe those changes over time.
Self-report measures also share context. A respondent's mood, optimism about technology or desire to give socially acceptable answers can influence several scales at once. PLS-SEM and bootstrapping quantify model relationships and uncertainty under the chosen specification; they do not remove unmeasured confounding. Alternative models—such as mental health influencing perceived usefulness, or teaching quality influencing all three—could plausibly fit the same one-time data. The word 'mediated' should therefore be read as a statistical decomposition, not proof of a psychological process.[1]
What universities can use now without overstating the evidence
The study supports asking students not only whether they use AI but whether the tools reduce or add cognitive burden. A chatbot that clarifies an unfamiliar concept may free attention; a poorly integrated system that requires repeated prompting, checking and policy interpretation may create technostress. Universities can audit course design, access, training and support around those practical experiences without claiming a treatment effect on mental health.
Digital self-efficacy is especially relevant because the association between perceived usefulness and mental health was stronger among students reporting greater confidence. If that pattern transfers, an institution could unintentionally widen differences by introducing AI tools without basic literacy, accessible guidance and alternatives. The result does not prove that training will improve wellbeing, but it argues against treating software availability as equal access. Students need clear expectations, human academic help and the freedom to avoid a tool when it adds burden or conflicts with disability, privacy or assessment needs.[1]
Mental health is not a product metric
Perceived convenience should not be confused with clinical benefit. The paper did not diagnose mental illness, compare counselling outcomes, measure crisis risk or test a specific AI product. It does not show that generative AI is therapeutic, safe for disclosing sensitive information or an alternative to professional care. Universities should keep wellbeing services, safeguarding and disability support independent of vendors' engagement metrics and product claims.
There are also educational trade-offs the model cannot settle. Reducing unnecessary cognitive burden can help learning, but some effort is integral to practising recall, reasoning and writing. A tool that makes a task feel easier could support learning, bypass it or do both depending on the assignment. Future work should combine wellbeing measures with objective learning outcomes, actual usage logs and assessment design, rather than assuming that more frequent use or lower reported effort is inherently beneficial.[1]
The evidence that would change the assessment
A stronger study would follow students across a term, measure wellbeing and overload before and after adoption, and separate different uses such as explanation, feedback, drafting and answer generation. A randomised or carefully matched course-level evaluation could compare supported AI use, unsupported access and non-AI alternatives while reporting attrition, subgroup effects, academic performance and use of support services. Qualitative interviews could reveal whether lower load reflects helpful scaffolding or displacement of learning.
The authors report ethical approval from the Research Committee of Prince Mohammad Bin Fahd University's College of Business Administration and written informed consent. Funding came from the University of Debrecen Program for Scientific Publication and Princess Nourah bint Abdulrahman University's Researchers Supporting Project PNURSP2026R797; the authors declare no competing interests. Those disclosures aid interpretation, but they do not solve the study design's causal limits. For now, this is useful Bangladesh evidence about students' perceptions—not a reason to market AI as a mental-health intervention.[1]