映射授权边界：澳大利亚高等教育中生成式人工智能治理的比较政策小插曲研究

中文摘要 AI 辅助

澳大利亚大学通过重叠的政策、程序、指南和评估说明来规范学生对生成式人工智能（GenAI）的使用，但这些环境可能对相同的行为作出不同的分类。我们将15个标准化的学生使用小插曲应用于20所澳大利亚大学的公共政策环境，并产生了300个大学-案例分类。我们分别分析了约束性文书（A层）与完整的官方环境（B层），然后将每个组合分类为明确允许、有条件允许、潜在政策违规、明确禁止或不确定。我们使用归一化香农熵衡量跨大学的差异。我们将120个组合（40.0%）分类为明确禁止，97个（32.3%）为潜在政策违规，27个（9.0%）为有条件允许，56个（18.7%）为不确定；没有组合达到明确允许的严格阈值。披露的语言改写和披露的AI起草段落产生了最大的差异，而明确的评估禁止则产生了一致性。约束性文书在100个组合中对GenAI保持沉默，而指南澄清了88个组合。主要研究者主导编码并辅以AI协助，手动审查了所有201个排队行。一位独立的人类第二编码员，未使用任何AI协助，对盲选的75行样本进行了编码。总体一致性达到57.3%（未加权Cohen's kappa = 0.395；bootstrap 95% CI [0.244, 0.539]）。研究结果区分了许可门控架构与基于披露的架构，并表明书面政策对AI生成文本的保留的规范比仅过程辅助更清晰。比较政策小插曲测试评估书面治理是否支持可辩护的分类；它不预测不当行为或执法决策。

英文摘要

Australian universities regulate students' use of generative artificial intelligence (GenAI) through overlapping policies, procedures, guidance, and assessment instructions, but these environments may classify identical conduct differently. We applied 15 standardized student-use vignettes to the public policy environments of 20 Australian universities and produced 300 university-case classifications. We analyzed binding instruments (Layer A) separately from the full official environment (Layer B), then classified each combination as clearly permitted, permitted with conditions, potential policy breach, clearly prohibited, or indeterminate. We measured cross-university divergence with normalized Shannon entropy. We classified 120 combinations (40.0%) as clearly prohibited, 97 (32.3%) as potential policy breaches, 27 (9.0%) as permitted with conditions, and 56 (18.7%) as indeterminate; none met the strict threshold for clearly permitted. Disclosed language rewriting and a disclosed AI-drafted paragraph produced the highest divergence, whereas an explicit assessment prohibition produced unanimity. Binding instruments remained silent on GenAI in 100 combinations, and guidance clarified 88. The primary researcher led the coding with AI assistance and manually reviewed all 201 queued rows. An independent human second coder, who used no AI assistance, coded a blind 75-row sample. Overall agreement reached 57.3% (unweighted Cohen's kappa = 0.395; bootstrap 95% CI [0.244, 0.539]). The findings distinguish permission-gated from disclosure-based architectures and show that written policies regulate the retention of AI-generated text more clearly than process-only assistance does. Comparative policy-vignette testing evaluates whether written governance supports defensible classifications; it does not predict misconduct or enforcement decisions.

发表机构

• Charles Darwin University（查尔斯达尔文大学）

机构由 AI 辅助整理，请以论文原文为准。