Artificial intelligence in higher education and graduate employability: cognitive engagement, self-efficacy, and career adaptability in the AI era

Synthesis: Kong and Zhang synthesize 46 publications into an argument rather than a result. They argue that the link between AI-supported learning and graduate employability runs through three learning processes: cognitive engagement and self-regulation, Self-Efficacy and motivational regulation, and career adaptability. Three tensions then decide whether AI use builds capability or replaces it: augmentation against offloading, career agency against AI-related insecurity, and resource access against capability stratification. Organizing the literature by outcome level is the review's most useful move, and it produces a deflationary conclusion: nearly everything measured sits at the proximal end. The pathway to employment is a proposal awaiting test, not a demonstrated sequence.

Critical narrative review: the authors synthesize 46 publications interpretively and collect no data of their own. There is no pooled effect size and no formal risk-of-bias scoring, so every number below is either an attributed claim about an included study or a count of the corpus, never a finding of this paper.

What the review argues

The review's target is a habit of inference. AI adoption, access or frequency of use is routinely treated as a proxy for graduate capability, and the authors argue the inference fails in both directions: a student who uses AI constantly may learn less, and one who rarely uses it may learn more. Their alternative is to treat employability as developmental rather than as a labor-market status, following career-development accounts in which identity, psychological resources and the capacity to mobilize learning across contexts matter more than a checklist of skills.

On that account AI changes two things at once. It changes how capability forms, by widening access to feedback, explanation, rehearsal and task support. It also changes the environment in which capability is judged, by altering what employers expect a graduate to do unaided. The review treats the first as an empirical question about learning processes and the second as a structural condition that the included studies mostly cannot observe.

Three mechanisms carry the argument. Cognitive engagement and self-regulation determine whether AI-assisted work leaves anything behind after the task ends. Self-efficacy and motivational regulation determine whether students keep investing effort and form accurate beliefs about their own competence. Career adaptability — the concern, control, curiosity and confidence that allow a person to anticipate and respond to occupational change — is the bridge from learning to career exploration and adjustment. Crucially, the authors classify career adaptability as a career-development resource, explicitly not an employment outcome.

Three tensions cut across the mechanisms. AI can augment thinking or host offloading; it can support career agency or provoke insecurity about replacement; and it can democratize access or widen gaps between students who already have the disciplinary knowledge to judge its output and those who do not. The framework in the paper is a set of theory-informed propositions about how these resolve, drawn as directional arrows with one dashed link to employment outcomes that the authors themselves label untested.

The outcome ladder, and where the evidence stops

The review's most disciplined contribution is a four-level classification. Direct employment outcomes include employment status, time to employment, job quality, income, retention and employer evaluations. Indirect career indicators include perceived employability, career readiness, career exploration, career decision-making and career adaptability. Proximal outcomes include AI literacy, engagement, self-regulated learning, self-efficacy, motivation, anxiety and course performance. The fourth category is theory.

Claims about employment are then restricted to studies that measured employment. Very few qualify. The review reports a twelve-week AI module with 120 non-STEM students in the United Kingdom, in which AI capital rose and was associated with employment status six months after graduation — a single-course, non-randomized design that cannot show the course caused the difference. It cites a self-reported employment survey of 148 graduates in Peru, cross-sectional in design. It cites a 42-participant study of short-term AI mock-interview training that changed simulated interview performance, which the authors note is not a hiring outcome.

Almost everything else sits one or two rungs down. A Saudi business sample of 652 linked AI tool use with engagement, performance, perceived employability and empowerment. A multinational survey of roughly 9,384 students found ChatGPT use more strongly associated with employment confidence among undergraduates and social science students. A Chinese sample of 460 related AI awareness and career education to employability confidence. None of these follow graduates into work, and none can be read as an employment finding.

Why the three tensions matter more than any single association

The augmentation–offloading tension has the best evidence in the review, and it is still thin. The authors cite a meta-analysis reporting that generative AI in active-learning contexts was associated with greater cognitive engagement, with over-reliance linked to disengagement and shallow processing. A four-wave within-subject study in Mongolia (267 enrolled, 211 analyzed) found engagement and self-regulated learning rising during an AI-enhanced flipped phase, particularly agentic engagement and metacognitive regulation — though the fixed sequence means time and order effects cannot be excluded. A quasi-experiment with 68 students found some self-regulation and experience measures improving after metacognitive prompts were added, while the between-group difference in academic performance was not significant. On the other side, the review cites an association between AI dependence and weaker critical thinking, with cognitive fatigue proposed as one explanation, and a three-wave study linking ChatGPT use to procrastination and memory concerns.

The agency–insecurity tension is where the evidence is most contradictory. A 2 × 2 quasi-experiment with 200 STEM undergraduates found AI-assisted writing associated with higher intrinsic motivation and Self-Efficacy, with larger gains among lower-achieving students, though performance differences were not significant in every comparison. Against that, a survey of 315 students found AI anxiety negatively associated with career decision-making, with career adaptability mediating, and the hypothesized moderating role of self-efficacy unsupported. Null results appear too: no significant relationship between AI anxiety and career decidedness among English language teaching and literature students, a weak negative association among translation and interpreting students, and no significant prediction of lifelong learning tendency among pre-service teachers. The review's reading is that there is no single role for AI-related confidence.

The stratification tension is the least evidenced and the most consequential. Prior disciplinary knowledge, AI Literacy, language resources, tool access and institutional support plausibly shape what students extract from the same system, and students with weaker foundations may be least able to detect confident-sounding errors. Yet the same review reports structured AI support helping lower-achieving students, and a 2,622-student vocational sample in which the direct association between AI learning support and employability was not significant. Access and benefit are not the same variable, and almost no included study measured access as such.

What a narrative review can and cannot support

The design is a critical narrative review: an interpretive synthesis with a documented search of Web of Science, Scopus, ERIC and PubMed from inception to 1 July 2026, supplementary searches to 2 October 2026, two-author screening, and 46 included publications. What that design can support is a construct map, a defensible taxonomy of outcome levels, an identification of tensions and boundary conditions, and a set of testable propositions. What it cannot support is an effect size, a causal claim, a temporal order, a prevalence estimate, or a statement about how much AI use changes employment prospects.

The corpus limits interpretation further. Of the 46 publications, 24 were conducted in mainland China and 41 appeared in 2025 or 2026, so the evidence base is recent, geographically concentrated and still moving. Twenty-eight studies were cross-sectional surveys; twelve used intervention, quasi-experimental or longitudinal designs; four were qualitative; two analyzed documents or employment data. Only a small subset reported employment status at all. A synthesis of this material can tell you what the field believes and what it has measured; it cannot tell you what works.

What this means for practice

• Instructors. Design for retained capability, not task output. The review's calibration cycle is the concrete version: students attempt a task, use AI to extend or challenge their reasoning, verify against disciplinary evidence, then complete an unassisted explanation or transfer task. That sequence makes the difference between fluency and understanding visible to both parties.
• Curriculum and assessment designers. Embed AI literacy in disciplinary courses rather than in generic tool workshops. The review specifies what that means per field: validation and debugging in engineering and computing, safety, evidence appraisal and accountability in health professions, source verification and authorship in the humanities. Ask for artifacts of process — revision traces, verification decisions, justifications for what was delegated — which the review treats as evidence of learning rather than as integrity policing.
• Career services and administrators. Pair AI-supported occupational exploration with Reflection on which tasks are changing and which capabilities transfer; assisted and unassisted practice side by side gives students a realistic read on what they can do alone. Monitor who actually gets the benefit, since access to tools is not access to guidance, and the equity discussion turns on prior preparation and support rather than on licenses.
• Researchers. Measure employment. The gap the review keeps naming is longitudinal follow-up with career entry, job quality, employer assessment and performance on authentic tasks, with prior achievement, discipline and labor-market conditions as covariates.

Limitations

• The review design excludes formal risk-of-bias assessment and quantitative pooling by construction, so it cannot weigh studies by quality or estimate any magnitude.
• The search is English-language and limited to four databases, with no systematic gray-literature search and no formal assessment of publication bias; relevant work in other languages and outside these indexes may be missing.
• The corpus is concentrated: 24 of 46 studies from mainland China and 41 published in 2025 or 2026, which limits transferability across national and disciplinary contexts and makes the evidence base unstable.
• Twenty-eight of 46 studies are cross-sectional, so direction of association is unresolved throughout; the review is candid that stronger career adaptability may cause more AI exploration rather than the reverse.
• Measures are largely self-reported perceptions — perceived employability, confidence, readiness — which the review itself argues are not equivalent to employment outcomes.
• The framework is a proposal. Its arrows, mediation paths and single dashed link to employment are hypotheses the authors expect future longitudinal work to test.

Citation

Kong, L., & Zhang, H. (2026). Artificial intelligence in higher education and graduate employability: cognitive engagement, self-efficacy, and career adaptability in the AI era. Frontiers in Psychology, 17, 1987347.