Does AI technostress reduce teachers’ work effort? A 290-person Taiwan survey found an indirect link through loneliness—not a causal effect
Teachers who reported more difficulty adapting to ChatGPT also reported more workplace loneliness, which in turn was associated with lower self-reported effort. The peer-reviewed study measured all variables once in late 2023, used a convenience panel and did not observe actual work performance.
The short version
- What happened
- The study retained 290 ChatGPT-using public-school teachers recruited through a Taiwanese commercial survey panel in October and November 2023, with quotas across primary, junior-high and senior-high schools.
- Who it affects
- Teachers may gain useful drafting support while also taking on new checking, policy and adaptation work that is invisible in simple adoption statistics.
- What remains unproven
- Cross-sectional self-report data cannot establish temporal order or causation, and reverse causality or common-method bias remain plausible.
Keep in mind
Cross-sectional self-report data cannot establish temporal order or causation, and reverse causality or common-method bias remain plausible.
The convenience sample included 290 ChatGPT-using public-school teachers from a commercial panel, not a probability sample of Taiwanese educators.
Data were collected in October–November 2023 and may not reflect later products, policies or staff experience.
No school identifiers were retained, so within-school clustering could not be modelled; no objective work-performance or student outcome was measured.
The work-effort construct had average variance extracted of 0.420, below the conventional 0.50 threshold, although other reliability measures were acceptable.
Source analysis · 1 linked source. Primary analysed source date: 5 November 2026.
Read the full report · 8 min The evidence, methods, implications and limitations
The answer: the survey found an indirect association, not proof that AI use reduces effort
A peer-reviewed study of 290 Taiwanese public-school teachers found a specific pattern: teachers who reported more difficulty adapting to ChatGPT also tended to report more loneliness at work, and those who felt lonelier tended to report lower work effort. In the authors’ statistical model, technostress was associated with effort indirectly through loneliness. The direct relationship between technostress and effort was not statistically significant once the proposed pathway was considered. This supports a workplace-wellbeing question worth investigating; it does not show that ChatGPT caused anyone to become lonely or work less.
That distinction matters because every central variable came from the same questionnaire at one point in time. A teacher who was already isolated, overloaded or dissatisfied could rate both AI adaptation and work effort differently. School culture, staffing, digital support, self-efficacy and local policy may influence all three measures. The authors acknowledge that temporal order cannot be established, reverse causation remains plausible and unmeasured confounding could explain some of the associations. The useful finding is therefore a bounded signal about how technology strain and social experience may travel together—not a measured productivity loss from generative AI.[1]
Who took part, when the data were collected and what was measured
The researchers partnered with Metasurvey, a Taiwanese survey company, which recruited teachers from its registered member database and offered coupons as incentives. Data were collected in October and November 2023, during an early phase of school ChatGPT adoption. The sample covered 20 administrative regions and used quotas to include primary, junior-high and senior-high teachers: 96, 97 and 97 respondents respectively. Of the 290 retained questionnaires, 173 respondents were female and 117 male; 144 were aged 31–40. Teachers who said they had not used ChatGPT were screened out, and submissions were limited to one per internet address.
Respondents completed five-point scales covering six retained technostress items, three workplace-loneliness items, ten work-effort items and a transformational-leadership measure. The technostress scale was adapted to name ChatGPT, with two items removed after factor analysis. Reported internal-consistency values were 0.793 for technostress, 0.881 for loneliness, 0.877 for effort and 0.857 for leadership. Five bilingual experts reviewed the Traditional Chinese translation. These checks make the questionnaire more interpretable, but reliability within this sample is not the same as objective measurement of workload, teaching quality, attendance or pupil outcomes.[1]
The distinction that matters
What the survey can and cannot establish
The paper links four self-reported constructs, but its one-time design cannot establish the direction of cause.
Observed
Higher technostress ratings
Higher loneliness ratings
Observed
Higher loneliness ratings
Lower effort ratings
Not significant
Technostress directly
Lower effort after mediation
Still unproven
ChatGPT use
Caused loneliness or reduced performance
What the statistical model did—and what the leadership result actually says
The authors used exploratory factor analysis, regressions and the PROCESS macro in SPSS, controlling for gender, age, education, tenure and school level. Their proposed sequence treated GenAI technostress as the stressor, workplace loneliness as the strain and work effort as the outcome. The indirect path was statistically supported: technostress related positively to loneliness, and loneliness related negatively to effort. The non-significant direct effect means the data did not support a simple claim that higher technostress by itself corresponded to lower effort after the mediator was included.
Teachers who rated their leaders as more transformational showed a weaker negative relationship between loneliness and effort. But leadership did not significantly moderate the relationship between technostress and loneliness, and the complete moderated-mediation result was not significant. In plain language, supportive leadership may coincide with resilience once loneliness is present, but this survey does not show that a particular leadership programme prevents AI-related strain. Because leadership was also self-reported by the same respondents, a generally positive or negative view of the workplace could influence several ratings at once.[1]
Why schools and other employers should look beyond tool adoption counts
An organisation can record that staff have licences, attend training or use a chatbot without learning whether the surrounding work has become easier. Teachers may need to verify generated lesson material, interpret changing rules, protect pupil information, answer colleagues’ questions and maintain professional judgment while normal relationships and support channels shift. Those demands can create cognitive and social costs even when the tool produces useful drafts. The study does not quantify those tasks, but it makes a practical case for monitoring both adaptation burden and connection at work rather than treating uptake as success.
A responsible implementation would ask workers which tasks improved, which new checking or documentation tasks appeared, where help is available and whether staff still have time to discuss difficult cases with colleagues. Managers can define when AI is optional, what evidence must be checked, how errors are reported and who owns training. Peer support, protected learning time and access to human expertise are testable organisational responses. They should be evaluated against work quality, workload, absence, retention and service outcomes—not assumed effective because a leadership score correlated with effort in one survey.[1]
The main evidence limits are sampling, timing and common measurement
The convenience panel is not a probability sample of Taiwanese teachers. Participants were already registered with a survey provider, had used ChatGPT and accepted an incentive, so they may differ from non-users or teachers less willing to join online research. The study did not retain school identifiers, preventing adjustment for clustering: colleagues in the same school may share leadership, workload and AI policy. The authors also report an average-variance-extracted value of 0.420 for the work-effort construct, below the conventional 0.50 threshold, although composite reliability and item loadings were considered acceptable.
The 2023 collection period also limits freshness of the underlying behaviour even though the paper was published in 2026. Products, school rules and staff familiarity have changed since then. Taiwan’s policy and cultural context may not transfer to countries with different staffing, leadership norms or data rules. No pupil, parent or supervisor rating was collected, and the study did not measure actual ChatGPT frequency, output quality or administrative time saved. Stronger evidence would follow teachers over time, combine surveys with work records and compare implementation approaches across schools.[1]
What would change the assessment
A longitudinal study could establish whether increases in AI-related demands precede loneliness and whether either predicts later changes in effort. School-level sampling would allow researchers to separate individual experience from common leadership and policy. Randomised or carefully matched evaluations of training, peer-support time, workflow redesign or limits on mandatory AI use could test which interventions actually reduce strain without sacrificing teaching quality. Objective measures should include time spent checking outputs, error correction, sick leave, retention, observed work quality and student outcomes, with privacy protections.
Replication outside education is also necessary. Teaching involves intensive interpersonal work, professional autonomy and public responsibility; the pathway may operate differently in software development, health care, administration or manual work. For now, the defensible conclusion is modest: among these 290 teachers, AI adaptation strain, workplace loneliness and self-reported effort were linked in a pattern consistent with the authors’ theory. Employers should treat that as a reason to measure social and cognitive effects, not as proof that generative AI has reduced workforce effort.[1]