• Multi-rater protocol combining human judgment with three LLM architectures demonstrates robustness across model generations, stochastic variation, adversarial weighting schemes, and inference modes. • Disciplines with the highest female representation record the lowest AI Interaction Scores while already facing earnings disadvantages, revealing a dual disadvantage that unguided integration may compound. • Pedagogical recommendations are advanced that redirect AI usage toward interactions that accentuate human capacity over cognitive offloading. Despite near-universal AI adoption among university students, no established framework connects student AI interaction patterns to the career outcomes they may predict, leaving institutions without the evidence to guide integration or assess the cost of inaction. This study establishes a preliminary empirical bridge between observed student engagement and occupational outcomes through a novel multi-rater protocol combining human expertise with architecturally diverse large language models. We conduct a macro-structural analysis of discipline-level patterns across the National Center for Education Statistics field of study taxonomy by mapping aggregated data from 574, 740 student interactions onto a validated career-competency framework. The findings reveal a structural alignment between pedagogical practice and economic stratification. Disciplines with the highest female representation occupy the lowest range of AI Interaction Scores. These same disciplines already face substantial earnings disadvantages, with a 52, 090 gap separating the highest and lowest-earning fields. AI Interaction Scores correlate negatively with female representation (ρ = −0. 68) and positively with median earnings (ρ = 0. 72), suggesting unguided integration may entrench rather than ameliorate existing inequities.
Mullens et al. (Sun,) studied this question.