Generative AI use and self-perceived employability among young adults in China: indirect associations through career decision-making self-efficacy and trust in generative AI

Synthesis: Students who used generative AI more also felt more employable, and the interesting part is why. The largest share of that association ran through students' confidence in their own career decisions rather than through their faith in the tool, so the main pathway is one a career course can deliberately build. Trust in the tool carried a smaller share, and a direct association remained, meaning use is not simply a proxy for confidence. The catch is that the outcome is a feeling about employability, not a job, a salary or a demonstrated skill, and one cross-sectional survey cannot tell whether use produces confidence or confident students seek out the tool. Read it as a case for teaching decision-making and calibrated trust, not for more screen time.

Key Findings

1. More use went with a stronger sense of employability. Students who used generative AI more often rated their employment prospects higher, and the association held with demographic and usage controls in the model.
2. Confidence in one's own decisions carried most of it. The larger indirect route ran through career decision-making self-efficacy — belief in appraising oneself, finding occupational information, choosing goals, planning and solving career problems — rather than through belief in the tool.
3. Trust in the tool carried a smaller share. Students who believed generative AI was accurate, private and benign also felt more employable, but this route was roughly half the size of the confidence route.
4. A direct association survived both routes. The two psychological variables did not account for everything, so use kept its own link to perceived employability. This is partial, not full, explanation.
5. The two explanations overlap. Career confidence and trust in the tool were more closely related to each other than either was to use, so the split between routes is a modeling choice as much as a psychological separation.
6. These were moderate scores, not enthusiasts. Average ratings sat just above the midpoint on every scale, and trust was the most dispersed — a sample of regular but not heavy users, many without work experience yet.
7. The design cannot order the story. Use, confidence, trust and perceived employability were all measured in one survey at one time, so reverse and reciprocal readings fit equally well.

How the study was built

The sample is 579 university students, half men and half women, nearly nine in ten aged 18–22, drawn from all 31 provincial-level administrative regions of China. The instrument contained 48 items from four established measures: a short GenAI usage scale, the 25-item Career Decision Self-Efficacy Scale–Short Form across its five domains, a three-item trust measure covering accuracy, privacy and nonmaleficence, and a 16-item self-perceived employability scale. All responses were five-point, with confidence ratings for self-efficacy and agreement ratings for the rest.

Internal consistency was high for every scale, the measurement model fit well, and a method-factor check moved the fit indices only marginally, which the author reads as no severe common-method bias. Effects came from 5,000 bias-corrected bootstrap resamples, and the model controlled for gender, age, educational background, work experience, and the frequency and duration of GenAI use — a longer list than the abstract's four controls.

Two routes of unequal weight

The theory separates two kinds of confidence. Career decision-making self-efficacy, from social cognitive career theory, is belief in one's capacity to appraise oneself, gather occupational information, select goals, plan and solve career problems. Trust in GenAI is belief in the competence and reliability of the system. Both could follow from using a tool that organizes career information quickly, and estimating them in parallel avoids claiming that one produces the other.

The larger pathway is the self-directed one, and that matters for higher education. Had it run mainly through trust, the message would be that students who believe the tool works feel better about their prospects — a promotional claim for the technology. Instead the heavier route runs through students' sense of their own decision-making, the kind of belief a career course can deliberately build. The trust route is real but modest, and the findings do not show whether use precedes trust or whether trusting students simply use the tool more, nor whether reported trust is calibrated to what the system can do.

Two features deserve more scrutiny than the paper gives them. First, the two mediators are strongly related, so their shares of the total association are not estimates of two cleanly separable mechanisms; some of the variance each explains is shared, and a model with a single confidence factor might fit these data nearly as well. Second, the usage background variables — work experience, frequency and duration — are almost interchangeable with each other and all enter as controls alongside the usage scale, with no collinearity diagnostic reported. The total association also differs between the two models presented, so the shares describe this specification rather than fixed proportions of a causal process.

What a perception score can and cannot show

Self-perceived employability is an appraisal of one's chances in the labor market, shaped by perceived capabilities, the reputation of one's university, the demand for one's field and general labor-market conditions. It is not a hiring outcome or a salary. The caution is not boilerplate: a student who uses GenAI heavily may feel more prepared while having developed fewer of the skills that the feeling refers to.

The surrounding literature makes that concrete. Another study of graduates found positive associations between perceived employability and confidence in GenAI, but the indirect effects through perceived usefulness and perceived ease of use were not statistically significant — favorable technology evaluations do not travel through the usual acceptance channels. Work on critical thinking distinguishes shallow dependence from deeper collaboration, so how a student uses the tool may matter more than whether they use it. Experimental evidence from high school mathematics shows that unguarded GenAI assistance can raise performance while the tool is present and depress it once the tool is removed. Together these mark the gap between assisted performance and independent capability that a self-report score cannot close.

The author also reports the counter-evidence directly: research among Chinese university students linking perceived GenAI impact to perceived skill devaluation and algorithmic anxiety, and the possibility that favorable employability perceptions coexist with dependence. The estimates describe one national sample at one point in time. They are consistent with a story in which GenAI-supported exploration builds real confidence, and equally consistent with a story in which confident students seek out more tools.

What this means for practice

• Instructors: Build the decision-making tasks, not just tool fluency. Self-appraisal, occupational information search, goal selection, planning and problem solving are the activities the larger pathway runs through, and they are teachable without any platform.
• Instructors: Pair assistance with calibrated trust exercises. Ask students to check an AI claim about a labor market or qualification requirement against a primary source, and to say what would change their mind — the trust measure captures belief in accuracy, privacy and nonmaleficence, not verified reliability.
• Advisers: Put the tool inside a structured search, not in place of one. Students who used GenAI more also explored more, but the study cannot tell purposeful exploration from anxious browsing, so require a written plan and a record of what was checked.
• Administrators: Treat self-perceived employability as a leading indicator, not an outcome, and track applications, interviews, offers and starting salaries alongside any confidence survey.
• Researchers: Measure GenAI use by purpose and reliance rather than frequency alone; the four-item scale cannot distinguish career exploration from application drafting.

Limitations

• The design is cross-sectional, so no temporal ordering or causal mediation can be established; reverse relationships are plausible, since students with greater decision-making confidence or trust may engage more actively with GenAI.
• The sample of 579 Chinese university students limits generalizability to other institutions, disciplines, regions and national contexts; comparison with employed young adults or across countries would require measurement invariance testing first.
• All constructs were self-reported in the same survey, which invites shared response tendencies and social desirability bias despite anonymous administration, partial reverse scoring and the method-factor check.
• The three-item trust scale provides limited differentiation across tools and tasks, and dimensional indicators and item parcels may obscure item-level measurement problems.
• The overall usage measure does not distinguish frequency, duration, task purpose or reliance patterns, so the model cannot assess dependence, skill devaluation or algorithmic anxiety as competing pathways.
• Self-perceived employability is subjective, and unmeasured factors including AI literacy, socioeconomic background and access to career support may explain the associations; following students into employment would clarify whether favorable perceptions correspond to stronger capabilities and better outcomes.

Citation

Yan, Z. (2026). Generative AI use and self-perceived employability among young adults in China: indirect associations through career decision-making self-efficacy and trust in generative AI. Frontiers in Psychology, 17, 1985156. https://doi.org/10.3389/fpsyg.2026.1985156