Completion Bias (The Ghostwriter)
Standard LLMs are mathematically compelled to resolve tasks as efficiently as possible. In education, this manifests as "Completion Bias"—where the AI provides the answer to resolve student friction, bypassing the neural encoding process entirely.
PII Exposure (The Training Trap)
Many vendors utilize real student interactions to "refine" their models. This creates a persistent data vulnerability, exposing sensitive student records and PII to the development cycle and potential model inversion attacks.
Ideological Drift (Impartiality)
Models trained on the open internet inherit the activist biases of their datasets. Without an architectural "Neutrality Layer," AI tutors present a significant legal and ideological liability for public institutions.
Subject Drift (Curriculum Delta)
Unconstrained AI often follows a student's tangent, drifting from GCSE Physics to English Literature or casual conversation. This dilutes the lesson's academic purpose and violates syllabus adherence mandates.
Pedagogical Hallucination
The most dangerous failure occurs when an AI provides a "hint" that accidentally confirms an incorrect student assumption. This creates a false sense of conceptual mastery while reinforcing fundamental errors.