Artificial Hivemind in K-12 Education: A Case Study

The growing use of large language models (LLMs) in K–12 education raises concerns beyond hallucinations, particularly regarding the potential homogenization of students' writing, reasoning, and independent perspectives. This study investigates the potential emergence of an artificial hivemind through a computational case study examining whether repeated LLM-assisted essay revisions increase assessed writing quality while reducing textual diversity. We analyze 1,818 student-written essays from the PERSUADE 2.0 dataset, grouped by original human scores into low-, medium-, and high-scoring groups. Five simulated rounds of revisions were conducted for each essay using Llama-3.1-8B-Instruct and Qwen3.5-9B models under three conditions: directive feedback (DirFeed), voice-preserving feedback (StyleFeed), and rewriting (RewriteFeed), yielding 54,540 revised texts. We evaluate changes in automated writing-quality scores, lexical and stylistic similarity, revision distances, and score distributions and rankings. Results indicate increases in mean judge scores across all three conditions, with greater improvements under directive and voice-preserving feedback than under rewriting when averaged across the two models. However, both feedback conditions increase within-model lexical similarity, even when the tutor is explicitly instructed to preserve student voice, while revised scores less closely preserve the original ordering of essays. Furthermore, feedback-based revisions continue to alter and expand essays, but later rounds yield small and inconsistent additional score gains. These findings suggest that improvements in automated writing-quality assessments may coincide with increasing textual similarity and weaker preservation of original score rankings. The study highlights a potential tension between LLM-assisted writing improvement and preservation of individual expression, motivating further investigation into artificial hivemind effects in educational settings.