Algorithmic Advice and Professional Judgment in a High-Stakes Educational Vignette: A Contrastive Analysis of Teacher Responses and Generative AI Outputs

Artificial intelligence systems are increasingly used to predict, recommend, and guide educational interventions, yet less is known about how educators at different professional stages respond to high-stakes, AI-framed recommendations or how such responses differ from generative AI textual outputs produced under comparable scenario conditions. This vignette-based mixed-methods study examined responses to an algorithmic directive recommending that an artistically gifted student be redirected from an economically uncertain arts pathway to a statistically safer mathematics-related pathway. Participants included pre-service teachers (n = 119) and in-service teachers (n = 49). The study also constructed a contrastive textual corpus of 50 AI-generated outputs from five large language models under two prompt conditions. These outputs were not treated as equivalent to human responses, but as textual comparators for examining how different sources organized language around risk, agency, care, and alternative pathways. Open-ended responses and AI outputs were analyzed using qualitative coding, Epistemic Network Analysis, and Classification and Regression Tree modeling. In-service teachers showed the highest proportion of Accept responses, often citing economic risk and future security. Pre-service teachers more often rejected the directive, linking rejection to student-centered reasoning and meta-cognitive gating. In the AI corpus, no output selected Accept, although the balance between Reject and Other outputs varied by model; outputs most often proposed integrated alternatives while using language coded as humanistic concern and expressed tension. The findings indicate how risk, care, agency, and responsibility are textually and professionally organized in high-stakes AI-framed educational decisions.