华东师范大学学报(教育科学版) ›› 2026, Vol. 44 ›› Issue (11): 34-46.doi: 10.16382/j.cnki.1000-5560.2026.11.004

• 教育数智化转型 • 上一篇    下一篇

从薄概念到厚概念:教育大模型伦理评测的范式转向

郁锋, 刘豪   

  1. 华东师范大学哲学系,上海 200241
  • 接受日期:2026-09-01 出版日期:2026-11-01 发布日期:2026-10-24
  • 基金资助:
    国家社会科学基金项目“多感官交互的知觉知识论研究”(22BZX099)。

From Thin Concepts to Thick Concepts: A Paradigm Shift in the Ethical Evaluation of Educational Large Language Models

Feng YU, Hao LIU   

  1. Department of Philosophy, East China Normal University, Shanghai 200241, China
  • Accepted:2026-09-01 Online:2026-11-01 Published:2026-10-24

摘要:

教育大模型伦理评测面临宏观原则难以转化为微观算法决策的困境。现有实践在数据集标注、评测基准及价值冲突处理等方面多以薄概念为主导,评估逻辑与真实教育实践脱节。基于元伦理学厚概念的事实−价值一体化特征及近期AI对齐研究,本文提出转向一种厚概念导向的评测范式(Thick-Concept-Oriented Model,以下简称TCOM)。TCOM不以厚评测取代薄评测,而以分层互补为原则。厚实化所改变的是评价基本单位与推理方式——由抽象薄评测拓展为事实与价值不可分的厚概念,由固定规则拓展为情境整体构型。TCOM包括情境结构化提炼、厚概念基准匹配及厚度分层判定的伦理推理三大核心环节,并以低−中−高三级厚度分层实现评价标准的分级适配。通过两个层面的例示与探索性分析,初步显示该框架能够识别教育大模型在伦理回应中调用的价值资源及其维度分布差异,为教育大模型的合规发展与伦理治理提供具有情境敏感性的分析路径。

关键词: 教育大模型, 伦理评测, 厚概念, TCOM体系, 道德特殊主义

Abstract:

The ethical evaluation of educational large language models faces the dilemma of translating macroscopic principles into microscopic algorithmic decisions. Existing practices in dataset annotation, evaluation benchmarks, and the handling of value conflicts are predominantly guided by thin concepts, resulting in an evaluation logic disconnected from authentic educational practices. Drawing on the fact–value integration characteristic of thick concepts in metaethics and recent research on AI alignment, this paper proposes a shift toward a thick-concept-oriented evaluation paradigm (Thick-Concept-Oriented Model, hereinafter TCOM). TCOM does not replace thin evaluation with thick evaluation but adheres to the principle of hierarchical complementarity. What thickening alters is the basic unit of evaluation and the mode of reasoning—extending from abstract thin evaluation to thick concepts in which facts and values are inseparable, and from fixed rules to the holistic configuration of situations. TCOM comprises three core components of ethical reasoning: structured refinement of situations, matching against thick-concept benchmarks, and thickness-stratified judgment. It achieves graded adaptation of evaluation criteria through a three-tier hierarchy of thickness (low–medium–high). Through illustrative examples and exploratory analyses at two levels, the framework is preliminarily shown to identify the value resources invoked by educational large language models in their ethical responses and the differences in their dimensional distributions, thereby providing a context-sensitive analytical pathway for the compliant development and ethical governance of educational large language models.

Key words: educational large language models (Edu-LLMs), ethical evaluation, thick concepts, TCOM Framework, moral particularism