Beyond Positive or Negative: Latent Profiles of University Students’ AI Attitude and Their Career-Development Correlates

Synthesis: Song, Hu and Li ran a three-wave, time-lagged survey of 379 first-year business-administration students at one Chinese university, measuring six AI-appraisal indicators at Wave 2 — perceived humanlikeness, adaptability and quality of AI, AI use anxiety, STARA awareness, and AI creative self-efficacy — and career outcomes at Wave 3. A latent profile analysis retained four profiles: Positive Empowerment (19.2%), Anxious Acceptance (37.5%), Low-Perception Detached (37.2%) and High-Perception High-Vigilance (6.1%). Learning agility, digital literacy, future work self-salience and perceived university digital support were prospectively associated with membership, chiefly separating Low-Perception Detached from the more engaged profiles, and Positive Empowerment and High-Perception High-Vigilance reported the highest career crafting and self-perceived employability despite opposite levels of threat. The central boundary condition is design: the three waves were only two weeks apart, no causal direction is established, and the smallest profile held 23 students.

Key Findings

1. Four profiles fit better than a single positive-negative axis. The four-profile solution had entropy 0.824 and average posterior probabilities from 0.883 to 0.949: Positive Empowerment (n = 73, 19.2%), Anxious Acceptance (n = 142, 37.5%), Low-Perception Detached (n = 141, 37.2%) and High-Perception High-Vigilance (n = 23, 6.1%). The adjusted Lo-Mendell-Rubin test stayed significant only through four profiles (p = 0.035) and turned nonsignificant at five (p = 0.376).
2. Opportunity and threat were co-activated in the smallest profile. High-Perception High-Vigilance scored above the sample mean on all six standardized indicators — 1.822 perceived humanlikeness, 1.674 adaptability, 1.753 quality, 1.521 AI use anxiety, 0.951 STARA awareness and 1.025 AI creative self-efficacy — while Positive Empowerment combined favorable capability perceptions (0.495, 0.704, 0.783) with the lowest threat of any profile (AI use anxiety −1.037, STARA awareness −0.653).
3. Low perceived threat was not the same as readiness. Low-Perception Detached sat below the sample mean on all six indicators (−0.778, −0.857, −0.831, −0.328, −0.398, −0.511), so its comparatively low anxiety and STARA awareness accompanied weak capability evaluations and the lowest AI creative self-efficacy in the solution.
4. Personal and contextual resources differentiated the detached profile. Higher learning agility raised the odds of Positive Empowerment over Low-Perception Detached (OR = 4.873) and of Anxious Acceptance over Low-Perception Detached (OR = 2.566); higher perceived university digital support gave OR = 3.728 for Positive Empowerment over Low-Perception Detached and OR = 2.052 for Anxious Acceptance over Low-Perception Detached.
5. Career outcomes tracked engagement, not threat level. Career crafting (Wald χ²(3) = 52.841) and self-perceived employability (Wald χ²(3) = 27.734) both differed across profiles. High-Perception High-Vigilance led at 4.222 and 3.964, Positive Empowerment followed at 3.912 and 3.624, and Low-Perception Detached was lowest at 3.407 and 3.140; the top two profiles did not differ significantly on either outcome.

How the study was run

Five hundred first-year students were invited, and 379 provided valid matched data across all three waves after quality screening and removal of eight multivariate outliers (Mahalanobis distance, D2 > 22.458). Participants were 283 women (74.7%) and 96 men (25.3%). Wave 1 captured learning agility, digital literacy, future work self-salience and perceived university digital support; Wave 2 captured the six AI attitude indicators; Wave 3 captured career crafting and self-perceived employability. AI was defined broadly as AI-enabled tools and systems rather than a single platform, and all measures were administered in Chinese via translation-back-translation. An item-level confirmatory factor analysis supported the six-factor structure, χ²(155) = 207.84, CFI = 0.982, TLI = 0.978, RMSEA = 0.030 and SRMR = 0.035. Profile membership was related to Wave 1 variables with R3STEP and career outcomes compared with BCH-adjusted means, both of which adjust for classification uncertainty.

What this means for practice

• Replace one-size-fits-all AI Literacy training with profile-based support. The four profiles spanned 19.2% to 37.5% of the sample, and advocates of a uniform program miss the Low-Perception Detached group entirely.
• For detached students, raise relevance before raising anxiety. Their low threat (−0.328 AI use anxiety, −0.398 STARA awareness) came with the lowest AI creative self-efficacy and the lowest career crafting (3.407), so discipline-specific tasks and mastery experiences matter more than warnings.
• Treat vigilance as engagement, not a problem to suppress. High-Perception High-Vigilance reported the highest career crafting (4.222) alongside the highest AI use anxiety (1.521), so support should target psychological Sustainability — bounded goals and information-overload management — rather than dampening concern.
• Build the contextual resource. Perceived university digital support carried the largest single odds ratio into the engaged profiles (OR = 3.728), pointing to accessible platforms, mentoring and technical assistance.
• Do not make the goal uniform positivity. Anxiety and career preparedness were not opposites, so Career Development and Readiness programs should assess appraisal accuracy, efficacy and proactive behavior together.

Limitations

• The sample was homogeneous — first-year business-administration students at one university in China, with women at 283 of 379 (74.7%) — so profile prevalence and correlates may be context-dependent.
• The three waves were two weeks apart and self-reported, so the Wave 1 and Wave 3 variables are prospective and time-lagged correlates, not causal evidence; the AI indicators were measured at one wave only, so profile stability was not examined.
• High-Perception High-Vigilance contained only 23 students (6.1% of the sample), so its characteristics and outcomes need replication in larger samples.
• The study used abbreviated, context-adapted indicators rather than the full AAAW scale, and neither STARA awareness nor AI creative self-efficacy is equivalent to the original framework's job-insecurity or personal-utility dimensions.

Citation

Song, H., Hu, A., & Li, X. (2026). Beyond Positive or Negative: Latent Profiles of University Students' AI Attitude and Their Career-Development Correlates. Behavioral Sciences, 16(9), 1460.