Ethics and Bias in AI for K-12 Education

The ethical deployment of large language models (LLMs) in K-12 education is among the most consequential — and least resolved — challenges facing the sector. Our analysis of 38 papers in this category reveals a field that has matured rapidly in identifying risks but remains far from establishing the evidence base needed for responsible deployment at scale. The research spans child-specific safety frameworks, bias detection across demographic groups, academic integrity, transparency, and the growing concern of cognitive offloading — where AI assistance may actively undermine the learning it purports to support.

A central finding is stark: adult-centric AI safety measures are insufficient for children. Younger learners face unique developmental vulnerabilities — susceptibility to misinformation, emotional dependency on AI personas, and exposure to age-inappropriate content — that generic safety guardrails do not adequately address. Researchers have responded with specialised benchmarks such as SproutBench (1,283 developmentally grounded prompts across 20 child-safety dimensions) and Safe-Child-LLM (200 adversarial prompts calibrated to two developmental stages), yet these tools remain largely untested in real classroom settings, particularly in low- and middle-income countries (LMICs).

Perhaps most troubling for education funders and policymakers is the evidence on cognitive offloading. A field experiment with approximately 1,000 Turkish high school students found that while GPT-4 assistance improved practice problem success by 48%, students subsequently scored 17% lower on unassisted exams. This suggests that without careful design, AI tools may create an illusion of learning while actively impeding the independent problem-solving skills that education systems exist to develop. The field urgently needs longitudinal studies, LMIC-specific evaluation frameworks, and clearer accountability structures before scaling deployment.