Breaking career boundaries in the AI era: how AI literacy relates to university students' perceived employability through future work self salience and boundaryless career mindset

Synthesis: Qiu and colleagues surveyed 1,945 Chinese university students and modeled AI literacy as a second-order construct feeding two career-psychology mediators — future work self salience, meaning how clearly a student pictures a future working self, and boundaryless career mindset — on the way to perceived employability. AI literacy was associated with perceived employability at β = 0.307 directly and 0.164 indirectly, and the model explained 25.1% of the variance. The interesting result is the mediation pattern rather than its size: the future-self pathway (0.117) carried more than three times the boundaryless-mindset pathway (0.034). Every outcome is self-report and the reverse-order model fitted equally well, so what the study establishes is that these beliefs travel together — which matters for career preparation even though nothing here shows that teaching AI literacy makes anyone more employable.

Key Findings

1. AI literacy was associated with perceived employability directly. The standardized path was β = 0.307, with a direct unstandardized association of 0.354 (95% CI [0.293, 0.419]) after both mediators entered.
2. The future-self pathway dominated. The indirect association through future work self salience was 0.117 (95% CI [0.083, 0.150]) against 0.034 through boundaryless career mindset ([0.019, 0.052]) and 0.013 through both in sequence.
3. The total association was moderate and mostly direct. Total indirect association was 0.164 ([0.125, 0.200]) against a total association of 0.518 ([0.459, 0.576]), so roughly a third of the association ran through the career-psychology mediators.
4. AI literacy predicted both mediators unevenly. It predicted future work self salience at β = 0.503 and boundaryless career mindset at β = 0.246, and the model explained 25.3% of the variance in the former and 14.7% in the latter.
5. The model explained a quarter of perceived employability. R² was 25.1%, with latent correlations among the four constructs ranging from 0.288 to 0.503 — moderate associations rather than near-collinearity.
6. The reverse-order model fitted equally well. The authors report that the data cannot distinguish their specified ordering from the reversed one, which they treat as a limit on the mediation claim rather than a footnote.

What the career constructs add

Perceived employability is a belief, and the study's contribution is the decomposition of that belief. Students who score higher on AI literacy also see a clearer future working self — they can picture the work they will do — and that clarity, more than a general openness to moving between employers, is what travels with believing oneself employable. Boundaryless career mindset, the disposition to cross organizational boundaries rather than stay in one firm, adds little once the future-self term is in the model.

For curriculum designers this redirects attention. Adding another AI tools module raises a literacy score, but the pathway that carries the association is about vocational imagination: whether students can describe a specific future role that uses these capabilities. That is a career development task more than a technical one, and it is closer to advising and internship work than to a syllabus change.

Measurement and method

The measurement model was checked carefully by the standards of this literature: 2,230 questionnaires returned and 285 excluded for implausible completion times, failed attention checks or straightlining, leaving 1,945 (87.22%). The second-order AI literacy construct loaded four dimensions — awareness, usage, evaluation and ethics — at 0.782 to 0.981, composite reliability ran 0.895 to 0.952, and the Fornell–Larcker criterion was satisfied.

That is better measurement hygiene than most surveys in the area, and the paper is still candid about what it cannot do. Harman's single-factor test put the first unrotated factor at 36.824%, and the one-factor CFA fitted terribly (CFI = 0.498, RMSEA = 0.156), which the authors correctly describe as evidence against a single general factor rather than as evidence that method variance is absent. A competing model that combined the two mediators fitted far worse, which does support keeping them separate.

What the sample cannot support

Recruitment was convenience sampling with snowball distribution through class groups and student social networks, and 87.22% of returned questionnaires survived screening. Institution type was recorded but not institution identity, so within-institution clustering could not be modeled and the reported intervals are not cluster-adjusted — a limitation the authors state explicitly. With 25.30% of the sample in vocational colleges and 35.17% in Double First-Class universities, the range of institutional experience is wide but unmodeled.

The larger caution is substantive. All four constructs are self-reports collected at one moment, and the sensitivity analysis shows the boundaryless-mindset path nearly doubling in a hybrid specification (0.118 against 0.252). A study of beliefs, measured once, should not be read as a study of employment.

What this means for practice

• Teach the vocational picture, not just the tool. The pathway that carried the association was future work self salience, so an AI literacy module that never asks students to describe a specific future role leaves the strongest mediator untouched.
• Use AI literacy as an advising prompt. Asking a student to sketch how AI changes the work they want to do produces the same construct the study measures, and it is cheap to run in a tutorial or an internship briefing.
• Do not promise employability. Perceived employability is a belief, measured by self-report; employment outcomes were not collected, and the reverse model fits equally well.
• Administrators. Career services and AI literacy programs here address the same student belief from two directions, and the finding argues for connecting them rather than running them as separate initiatives.
• Researchers. Publication-quality measurement does not rescue a cross-sectional mediation; the next study needs performance-based AI assessment and actual employment data.

Limitations

• Cross-sectional, all self-report. Every construct came from the same respondents at one time point, and the reverse-order model fitted as well as the specified one.
• Convenience and snowball sampling. 1,945 analyzable responses from a non-representative recruitment chain, with no claim to national representativeness.
• No clustering adjustment. Institution type was recorded but not identity, so within-institution dependence could not be modeled.
• Path magnitudes moved across specifications. The boundaryless-mindset to employability coefficient was 0.118 in the full latent model and 0.252 in a hybrid model.
• Perceived, not actual, employability. No employment, salary or job-search data were collected.

Citation

Qiu, Y., Zhu, J., Liu, X., & Li, R. (2026). Breaking career boundaries in the AI era: how AI literacy relates to university students' perceived employability through future work self salience and boundaryless career mindset. Frontiers in Psychology, 17.