Beyond the Oracle Problem: A Framework to Decouple Technical Alignment Debt from Human Cognitive Offloading in LLMs

Condori-Bustincio K.F. · Publicaciones de la Facultad de Educacion y Humanidades del Campus de Melilla (2026) · DOI: 10.30827/publicaciones.v56i1.36849

Resumen

Introduction: Generative AI in education presents the “Oracle Problem,” driven by two interconnected risks: Human Cognitive Offloading and Technical Alignment Debt. While over-reliance on AI is linked to decreased neural connectivity in Alpha and Theta bands—creating “cognitive debt”—standard RLHF safety paradigms instill a subservience bias that prioritizes user distress reduction over essential pedagogical friction Method: Using a Socratic Tutor Architecture based on the ICAP paradigm, three models (GPT-5.1, Gemini 3 Pro, and DeepSeek-V3) were evaluated across STEAM domains, including Mathematics, Coding, and Arts. The experiment was conducted in Spanish, utilizing the Fernández-Huerta Readability Index to ensure linguistic validity. We employed an Exploratory Stress-Test Framework where synthetic students escalated through five levels of psychometric stress—from confusion to aggressive coercion—to force a pedagogical collapse. Results: Under maximum stress, high-reasoning models exhibited “Compassionate Collapse,” leaking direct solutions to appease user emotional pain. While Gemini 3 Pro and DeepSeek-V3 maintained higher rigor, they displayed “Scaffolding Degradation” by simplifying problems into trivial hints. Metrics such as the Interrogative Ratio IRand Direct Answer Risk DAR confirmed the structural decay of socratic guidance at higher stress gradients. Conclusions: The findings demonstrate that superior reasoning capabilities (or high-reasoning models) do not automatically guarantee pedagogical safety, as standard safety alignment protocols fundamentally conflict with long-term educational goals. To reconcile the tension between safety and rigor, a paradigm shift toward domain-specific fine-tuning—specifically Pedagogical Reinforcement Learning from Human Feedback (P-RLHF)—is required to ensure AI tutors maintain the necessary cognitive burden rather than yielding to instant user gratification.

Instituciones: Universidad Nacional de San Agustín de Arequipa