Perceived AI knowledge linked to trust, anxiety, and reported reliance

Generative artificial intelligence has shifted from occasional experimentation to routine academic practice. The Student Generative AI Survey 2025 reported that 92 percent of university students had used AI tools in some form, up from 66 percent the previous year, while the Global AI Student Survey 2024 found that 86 percent of students used AI in their studies, with 54 percent using it at least weekly. A new study published in Frontiers in Psychology adds a psychological dimension to this picture. Researchers analyzed cross-sectional survey responses from 526 undergraduate students in design-related disciplines in China, using partial least squares structural equation modeling to trace how students’ perceptions of AI-generated content tools relate to their perceived knowledge of the technology, their trust and anxiety, their self-efficacy, their intention to keep using the tools, and their self-reported reliance on them in creative work.

The study focuses on design education, where generative AI intervenes directly in creative judgment rather than merely automating routine tasks. The disciplines covered included visual communication, graphic and brand design, industrial and product design, interaction and UI/UX design, and environmental, interior, and landscape design. AI-generated content, or AIGC, tools produce textual, visual, or multimodal output in response to prompts and can support students in generating alternatives, exploring styles, and accelerating early-stage design development. But the authors note that these tools may also shape idea generation, aesthetic judgment, and creative authorship, and that students may struggle to evaluate the reliability, originality, and appropriateness of what the systems produce. Prior review evidence cited in the paper suggests that over-reliance on AI dialog systems may weaken critical cognitive capabilities, including decision-making, critical thinking, and analytical reasoning, especially when users accept outputs without sufficient evaluation.

Unlike most technology acceptance research, which stops at whether people intend to use a tool, the study treats perceived technological dependence as a post-adoption outcome. The authors define it narrowly as students’ self-reported functional reliance on AIGC tools in current design work: incorporating the tools into routine tasks, preferring them in creative production, and expecting that working without them would reduce efficiency. They stress that the construct does not imply compulsive use, impaired functioning, or an objectively demonstrated inability to design without AI support. The measure includes an item on habitual reliance, giving it some conceptual proximity to habit and continuance intention, but its distinguishing emphasis is students’ anticipated difficulty in maintaining efficiency without the tools. Similarly, perceived AIGC knowledge refers to students’ self-assessed understanding of how the tools work, how to use them, and when they suit a task. The questionnaire recorded judgments about students’ own understanding; it did not test whether that understanding was accurate or whether students could actually evaluate generated outputs.

Data were collected between June and August 2025 through questionnaire links shared in course-related and student group chats. Of 537 submitted questionnaires, 11 were excluded because respondents reported no prior use of AIGC tools in design creation, leaving 526 for analysis. The sample included 270 male students (51.3 percent) and 256 female students (48.7 percent), spread across academic years, with 38.8 percent reporting occasional AIGC use, 34.6 percent frequent use, and 26.6 percent describing themselves as highly proficient users. All ten constructs — perceived reliability, ease of use, usefulness, intelligence, perceived AIGC knowledge, AIGC-related anxiety, self-efficacy, trust, usage intention, and perceived technological dependence — were measured with four items each on a five-point Likert scale. The analysis used partial least squares structural equation modeling in SmartPLS 4.0, with bootstrapping based on 5,000 resamples to test the hypothesized paths.

The measurement model performed well. Outer loadings ranged from 0.767 to 0.920, Cronbach’s alpha from 0.853 to 0.914, composite reliability from 0.901 to 0.939, and average variance extracted from 0.696 to 0.795, supporting reliability and convergent validity. Discriminant validity was assessed with the Fornell–Larcker criterion, cross-loadings, and the heterotrait–monotrait ratio of correlations; the HTMT between usage intention and perceived technological dependence was 0.743, and between perceived knowledge and trust 0.756, both below the prespecified 0.90 criterion. Because all constructs came from the same questionnaire, the authors checked for common method bias: Harman’s single-factor test extracted multiple factors, with the first accounting for 36.760 percent of item variance, and full-collinearity variance inflation factors ranged from 1.493 to 3.132, below the 3.3 reference value.

The structural results showed that three of four technological perceptions were significantly associated with perceived AIGC knowledge. Perceived reliability had the strongest association (β = 0.371, p < 0.001), followed by perceived usefulness (β = 0.272, p < 0.001) and perceived intelligence (β = 0.131, p = 0.005). Perceived ease of use, a cornerstone of the original Technology Acceptance Model, was not significantly associated with perceived knowledge (β = 0.061, p = 0.187), and its effect size was negligible (f² = 0.004). The authors suggest that an accessible interface may encourage use without necessarily changing students’ reported understanding, while noting that the ease-of-use items also touched on familiarity, which may reflect prior exposure rather than perceived effort, making the non-significant path harder to interpret. The four perceptions together explained 44.4 percent of the variance in perceived knowledge.

Perceived AIGC knowledge, in turn, was positively associated with trust (β = 0.666, p < 0.001), self-efficacy (β = 0.460, p < 0.001), and usage intention (β = 0.296, p < 0.001), and negatively associated with AIGC-related anxiety (β = −0.487, p < 0.001). The knowledge–trust link was the largest effect in the model (f² = 0.799), although the authors caution that students with a generally favorable view of AIGC may have rated both their understanding and their trust highly. Downstream, anxiety was negatively associated with usage intention (β = −0.247, p < 0.001), while self-efficacy (β = 0.124, p = 0.002) and trust (β = 0.291, p < 0.001) were positively associated with it. Usage intention was strongly associated with perceived technological dependence (β = 0.644, p < 0.001; f² = 0.708, a large effect). The model explained 59.1 percent of the variance in usage intention and 41.5 percent in perceived dependence.

Mediation analysis with 5,000 bootstrap resamples showed that trust, anxiety, and self-efficacy each carried a significant indirect effect in the relationship between perceived knowledge and usage intention: through trust (β = 0.194, p < 0.001), through anxiety (β = 0.120, p < 0.001), and through self-efficacy (β = 0.057, p = 0.005). Because the direct path from perceived knowledge to usage intention remained significant, the pattern is consistent with partial statistical mediation. The selection of indirect paths was based on the paths examined rather than on statistical significance, and it included the non-significant ease-of-use path. Perceived reliability, usefulness, and intelligence were also indirectly associated with usage intention through perceived knowledge, while the indirect path from perceived ease of use was not significant. Finally, perceived knowledge was indirectly associated with perceived technological dependence through usage intention (β = 0.191, p < 0.001).

The authors are explicit about what the cross-sectional design cannot show. All constructs were measured at a single time point, so the path directions represent theoretical expectations rather than an observed sequence: the data cannot establish whether favorable perceptions led to greater perceived knowledge, whether trust developed after perceived knowledge, or whether usage intention preceded reported dependence — the reverse ordering remains possible. The strong knowledge–trust association may also be inflated by shared response tendencies, which the diagnostic checks cannot rule out. Model fit warrants caution as well: the standardized root mean square residual of 0.037 met the approximate 0.08 guideline, but the normed fit index was 0.887, slightly below the 0.90 benchmark, and bootstrap-based exact overall fit was not established. The reported paths should therefore be interpreted within the specified model rather than as established causal structure.

The findings carry practical implications the authors frame carefully. Because the dependence measure does not distinguish useful reliance from over-reliance, higher scores alone should not be treated as evidence of harm. In design courses, they suggest, students could compare outputs for the same brief across prompt revisions, note inconsistent or unsuitable results, and explain why they accepted or rejected each output — though this exercise was not tested in the study. Course guidelines could specify acceptable AI-supported ideation, disclosure, originality, and responsibility for revising generated material, with sketches, prompt records, and design rationales helping instructors examine how students reached their decisions. For tool developers, side-by-side comparisons, revision histories, and interfaces showing how prompt changes affect outputs could support more reflective use. The sample was a convenience sample of Chinese design students with prior AIGC experience, actual use behavior was not measured, and the authors call for longitudinal and experimental research to test the proposed sequence over time.

Subject of Research: Psychological mechanisms linking perceived knowledge of generative AI tools to trust, anxiety, self-efficacy, and perceived technological dependence among design students

Article Title: Psychological mechanisms of perceived technological dependence on generative artificial intelligence tools: perceived knowledge, trust, anxiety, and self-efficacy

Article References: Wei, L., Zhang, S., & Heo, J. (2026). Psychological mechanisms of perceived technological dependence on generative artificial intelligence tools: perceived knowledge, trust, anxiety, and self-efficacy. Frontiers in Psychology, 17, Article 1928967. https://doi.org/10.3389/fpsyg.2026.1928967

Image Credits: AI Generated

DOI: 10.3389/fpsyg.2026.1928967

Keywords: generative AI, AIGC, design education, technology acceptance model, trust, self-efficacy, anxiety, perceived knowledge, technological dependence, usage intention, PLS-SEM, cross-sectional survey