Sociodemographic Biases in Educational Counselling by Large Language Models

This collection of recent CS/cyber arXiv papers highlights multiple systemic risks from large language models, AI governance, and alternative software ecosystems. Key findings: (1) LLMs exhibit measurable sociodemographic biases in educational counselling that are amplified by vague inputs and vary across models; (2) using LLMs as evaluators in healthcare (LaaJ) is growing but validation is weak, bias and oversight testing are largely absent, and the authors propose the MedJUDGE risk‑stratified evaluation framework; (3) a formal framework dubbed “LLM Psychosis” and a diagnostic scale (LCIS) is