AI literacy and vocational students’ complex problem-solving skills: a moderated mediation model of systematic thinking, cognitive load, and perceived school ICT support

Synthesis: A three-wave survey of 784 students at six Chinese higher vocational colleges tested whether AI literacy is associated with self-reported complex problem-solving through two cognitive routes — systematic thinking and cognitive load — and whether perceived school ICT support conditions those routes. Total association was 0.480; the direct path was 0.248 and the indirect paths 0.151 (systematic thinking) and 0.081 (cognitive load). Both interactions pointed the same way: at higher perceived ICT support, AI literacy was more strongly tied to systematic thinking and more strongly tied to lower cognitive load, so the conditional indirect effects grew from 0.117 to 0.176 and from 0.038 to 0.127 across the moderator's range.

Key Findings

1. The analytic sample was 784 students from six higher vocational colleges in Guangdong Province (four public, two private), matched across three waves: 920 invited, 878 valid at T1 (95.4%), 827 at T2 (94.2% retention), 784 at T3 (94.8% retention, 85.2% cumulative). Waves ran in March, April and May 2025, about four weeks apart, with no intervention between them.
2. The sample was 409 male (52.20%) and 375 female (47.80%); 217 first-year (27.70%), 251 second-year (32.00%) and 316 third-year (40.30%) students, and 44.9% studied science, engineering, agriculture or medicine.
3. AI literacy correlated 0.421 with self-reported CPS, systematic thinking 0.448 with CPS, and AI literacy 0.471 with systematic thinking; cognitive load correlated −0.361 with AI literacy, −0.368 with CPS and −0.271 with systematic thinking. Means were 3.46 (SD 0.60) for AI literacy, 3.70 (0.67) for systematic thinking, 3.23 (0.82) for cognitive load, 3.78 (0.76) for perceived ICT support and 3.40 (0.69) for CPS, all on five-point scales.
4. The total effect of AI literacy on self-reported CPS was 0.480, SE = 0.036, 95% CI [0.409, 0.550]; the direct effect was 0.248, SE = 0.039, 95% CI [0.172, 0.324]; the total indirect effect was 0.232, BootSE = 0.028, 95% CI [0.181, 0.288].
5. Path coefficients were a = 0.499 (SE = 0.034) from AI literacy to systematic thinking, a = −0.487 (SE = 0.046) to cognitive load, b = 0.303 (SE = 0.035) from systematic thinking to CPS and b = −0.166 (SE = 0.026) from cognitive load to CPS, all p < 0.001. The specific indirect effect through systematic thinking was 0.151, BootSE = 0.028, 95% CI [0.097, 0.208]; through cognitive load it was 0.081, BootSE = 0.018, 95% CI [0.049, 0.119]. The model explained R² = 0.260 of systematic thinking, 0.140 of cognitive load and 0.352 of CPS.
6. Perceived school ICT support moderated the AI literacy → systematic thinking path, interaction b = 0.128, SE = 0.034, ΔR² = 0.012. Simple slopes were 0.387 (t = 9.549) at one SD below the mean and 0.582 (t = 13.913) at one SD above, both p < 0.001.
7. It moderated the AI literacy → cognitive load path in the same direction, interaction b = −0.352, SE = 0.046, ΔR² = 0.061. Simple slopes were −0.231 (t = −4.154) at low support and −0.769 (t = −13.394) at high support.
8. Conditional indirect effects through systematic thinking were 0.117 [0.053, 0.178] at low support (M − 1 SD = −0.764), 0.147 [0.096, 0.200] at the mean and 0.176 [0.122, 0.240] at high support (M + 1 SD = 0.764); the index of moderated mediation was 0.039, BootSE = 0.021, 95% CI [0.008, 0.089].
9. Conditional indirect effects through cognitive load were 0.038 [0.016, 0.064] at low support, 0.083 [0.052, 0.117] at the mean and 0.127 [0.082, 0.178] at high support; the index of moderated mediation was 0.058, BootSE = 0.013, 95% CI [0.035, 0.089].
10. A sensitivity analysis splitting cognitive load into its two dimensions found positive indirect effects through intrinsic load, 0.0487 [0.0232, 0.0785], and extraneous load, 0.0323 [0.0071, 0.0591]; the contrast between them, 0.0164 [−0.0233, 0.0576], was not significant. Both moderation indices stayed positive: 0.0357 [0.0170, 0.0586] and 0.0236 [0.0063, 0.0463].

What the three waves buy, and what they do not

The design is a three-wave survey, not a trial. AI literacy, perceived ICT support and demographics were measured at T1; systematic thinking and cognitive load at T2; self-reported CPS at T3. The four-week gaps separate the measures in time but do not establish change, because no construct was measured twice and no baseline levels were collected.

The outcome is the largest interpretive constraint. CPS was measured with an eight-item self-report scale adapted from Promma et al. (2025) covering problem identification, solution generation, iterative testing and stakeholder analysis (α = 0.928). Students rated their own capability, so "complex problem solving" here means perceived problem-solving capability, and every path above is an association among self-reports from the same people.

The instruments are established ones, translated and back-translated: AI literacy from Wang et al. (2023), 12 items across awareness, use, evaluation and ethics (α = 0.870); systematic thinking from Promma et al. (2025), eight items across structuring, reasoning, prioritizing and information integration (α = 0.949); cognitive load from Leppink et al. (2013), six items (α = 0.835); perceived school ICT support from Wang and Zhao (2021), three items (α = 0.840). The five-factor model fitted well (χ² = 542.029, χ²/df = 2.045, CFI = 0.977, RMSEA = 0.037) and beat every collapsed alternative, so the mediators separate.

Why the cognitive-load pathway is about effort, not difficulty

The negative a and negative b multiply into a positive indirect effect, which is easy to misread. Students with stronger AI literacy reported lower scores on the six-item composite of intrinsic and extraneous demands, and lower demands went with higher perceived CPS. Nothing here says AI literacy makes tasks objectively easier. It says familiarity with what a system can do, how prompts are framed and how outputs are checked removes peripheral effort — repeated searches, badly specified prompts, attention to irrelevant output — leaving working memory for problem representation.

The measure is a composite of intrinsic and extraneous items only. The four germane-load items were dropped because germane load indexes constructive processing rather than task or instructional demand, so the finding applies to those six items and not to cognitive load in all its forms. The item-level CFA nonetheless distinguished the two dimensions (one-factor RMSEA = 0.348; two-factor 0.051), and the sensitivity analysis found positive indirect effects through each separately, with a contrast CI spanning zero.

This is where the study sits closest to cognitive load theory, and furthest from any claim about critical thinking as a disposition. Systematic thinking here is a narrow, four-dimension construct — structuring, reasoning, prioritizing, integrating — and the authors explicitly decline to treat it as systems thinking in the broad sense. The 0.303 path from that construct to CPS is the strongest single coefficient in the outcome model.

The moderator is small in variance and large in interpretation

Perceived school ICT support adds only ΔR² = 0.012 to the systematic-thinking equation, and its own coefficient there is 0.251. What matters is the shape of the interaction, not its size. The AI literacy → systematic thinking slope nearly doubles from 0.387 to 0.582 between one SD below and one SD above the moderator. The AI literacy → cognitive load slope more than triples, from −0.231 to −0.769.

The mechanism the authors propose is opportunity plus relief from friction. Where students perceive accessible resources and timely help, they can test ideas, compare outputs and retry without spending working memory on unstable platforms or unfamiliar interfaces. Where support is thin, the environment itself becomes a demand.

One caution belongs with that story. The moderator is a three-item student perception, not an audit of institutional infrastructure, so the finding cannot underwrite a claim that buying software or staffing a help desk will move cognitive outcomes. Clustering, as noted below, is unmodelled, and the indices of moderated mediation, 0.039 and 0.058, are the results most exposed to that.

What this means for practice

• Vocational instructors. Teach the cognitive moves, not just the tool. The largest coefficient ran through structuring, reasoning, prioritizing and integrating information, so an assignment that has students state the problem, compare two AI outputs with explicit reasons and rank constraints does more than one that merely uses AI.
• Curriculum designers. Build AI literacy as a practical capability across programs rather than a standalone module. The scale spans awareness, use, evaluation and ethics, and the indirect effects assume all four are operating on an industry task.
• Administrators. Treat perceived access as part of the learning environment, and measure it before investing. The moderation is about what students believe they can reach and whom they can ask; a support service nobody knows about will not show up in these pathways.
• Researchers. Report conditional indirect effects at the moderator's actual ±1 SD values, not at arbitrary "high" and "low" points, and model the college level when sampling is institutional.

Limitations

• All five focal constructs are self-reported by the same respondents at all three waves, and CPS is perceived capability rather than observed performance; shared-method bias cannot be excluded, and the authors' Harman's test (first component 30.07%) and ULMC analysis are post hoc diagnostics only.
• Each construct was measured once, with no baseline for the mediators or the outcome, and the questionnaire does not establish that T2 and T3 responses referred to the same task; the paths describe between-person associations and cannot support within-person change or causal claims.
• Students were nested in six colleges, but class and college identifiers were unavailable, so clustering was not modeled and the standard errors may be understated.
• The cognitive-load composite excludes germane load by design, and the authors caution against generalizing to cognitive load as a whole.
• Attrition from 920 to 784 (136 cases) could not be tested against baseline characteristics, since nonretained cases had no individual records; selective attrition remains possible.
• The sample is confined to higher vocational colleges in one Chinese province, and Self-Efficacy, though discussed as theoretical background, was not measured, so it cannot be ruled out as an alternative mechanism.

Citation

Huang, M., Huang, M., He, T., Yi, X., & Deng, W. (2026). AI literacy and vocational students’ complex problem-solving skills: a moderated mediation model of systematic thinking, cognitive load, and perceived school ICT support. Frontiers in Psychology, 17, 1971672.