Perceived career impacts of artificial intelligence and career decision-making self-efficacy among Chinese college students

Synthesis: Students reading about AI's effect on the labor market are not neutral observers of their own futures, and this study asks how that reading reaches their confidence in making career decisions. Zhang surveyed 415 Chinese college students and modeled two pathways through Social Cognitive Career Theory: positive perceived AI career impacts were associated with stronger career decision-making self-efficacy through career planning (indirect β = 0.176), while negative perceived impacts had no direct association with self-efficacy at all but were associated with lower efficacy through career anxiety (indirect β = −0.184). Both valences therefore matter, but through different intermediates — a planning route for optimism and an affective route for pessimism. The model explained 54.6% of the variance in decision-making self-efficacy. The design is cross-sectional and self-reported, so the paths describe a plausible ordering of associations rather than a demonstrated sequence.

Key Findings

1. Positive and negative perceptions behaved differently. Positive perceived career impacts of AI related to decision-making self-efficacy directly (β = 0.186), while negative perceptions did not (β = 0.029, p = 0.406).
2. Planning carried the positive route. Positive perceptions predicted career planning (β = 0.488), which in turn predicted self-efficacy (β = 0.361), an indirect path of 0.176.
3. Anxiety carried the negative route. Negative perceptions predicted career anxiety (β = 0.588), which predicted lower self-efficacy (β = −0.314), an indirect path of −0.184.
4. The model accounted for a substantial share of variance. R2 was 0.546 for career decision-making self-efficacy, 0.346 for career anxiety and a smaller share for career planning.
5. Valence is not a single continuum. Treating "AI attitudes" as one score would average two opposite mechanisms, which is the paper's practical argument for measuring positive and negative perceptions separately.
6. An independent validation sample largely replicated the measurement model. A separate sample of 246 responses showed acceptable loadings (0.617 to 0.835) and composite reliabilities from 0.835 to 0.927.

How the model was estimated

The main sample was 415 college students recruited through a commercial online survey panel (Wenjuanxing) with purposive sampling; a second, independent sample of 246 responses was used to check the measurement model. Perceived career impacts of AI were split into positive and negative perceptions, and the model placed career planning and career anxiety as parallel mediators between those perceptions and career decision-making self-efficacy.

Estimation used partial least squares structural equation modeling, with a cross-path robustness model and a reverse model to probe alternative orderings — the reverse specification found self-efficacy predicting career planning at β = 0.534, which the authors report as a reason to treat the direction of that relationship as open. The career planning measure was adapted and shortened: four of its original ten items were dropped because their loadings fell below 0.60, and the retained items emphasize awareness, exploration and goal planning while career-related action, knowledge acquisition and adjustment are less represented.

What the two pathways imply for support

The practical distinction is that optimism and pessimism enter the system at different points. Positive perceptions work through planning behavior, which is something a course, an advising relationship or a career center can support directly. Negative perceptions work through anxiety, which responds to different interventions — accurate information about occupational change, and structured ways to act under uncertainty rather than reassurance alone.

Neither pathway was tested prospectively, and the reverse model's finding that self-efficacy predicts planning just as plausibly is a reminder that the causal story is not settled. What the study does establish is that the two valences should not be collapsed: a single "AI attitude" scale would hide the planning and anxiety routes that the data separate.

What this means for practice

• Instructors and advisors: Separate the two conversations. Students who are optimistic about AI need planning support that turns confidence into steps; students who are anxious need information and manageable action, not encouragement to feel better.
• Administrators: Treat career anxiety as a curriculum issue, not only a counselling one. The anxiety route predicts lower decision-making confidence, and planning support is the route the positive pathway runs through.
• Researchers: Measure positive and negative perceptions separately. Collapsing them averages two mechanisms with opposite signs, and the paper's own model depends on the split.

Limitations

• The sample came from a commercial online panel with purposive sampling rather than probability sampling, and over-represents science and engineering students.
• All constructs are self-reported and measured on one occasion, so the model establishes associations consistent with the proposed ordering, not temporal or causal sequence.
• The independent validation sample showed a career planning average variance extracted of 0.459, slightly below the conventional 0.50 threshold.
• The career planning measure lost four of its ten original items in validation, narrowing its content coverage relative to the original multidimensional scale.
• The reverse model fit the data plausibly, leaving the direction between self-efficacy and career planning open.

Citation

Zhang, H. (2026). Perceived career impacts of artificial intelligence and career decision-making self-efficacy among Chinese college students: the mediating roles of career planning and career anxiety. Frontiers in Psychology, 17, 1939461.