大语言模型回应“AI取代教师”议题的语用身份建构与关系管理研究

Pragmatic Identity Construction and Rapport Management in LLM Responses to the Issue of “AI Replacing Teachers”

摘要: 本文围绕“AI是否会取代教师”这一具有职业敏感性和公共讨论度的教育议题，考察大语言模型生成话语中的语用身份建构与关系管理方式。研究选取ChatGPT、DeepSeek和文心一言在同一标准化提示语下生成的中文回应文本，将模型输出视为具有拟社会交际功能的算法话语实践，并综合语用身份理论、关系管理理论和评价资源分析方法，对自我定位、立场表达、情态缓和、条件化框定、互动对齐和焦点重组等现象进行质性细读与辅助性词项观察。结果显示，三种模型均回避“完全取代教师”的强断言，却形成了不同的身份完成路径：DeepSeek倾向于技术助手型身份，突出工具边界和人文价值；文心一言更接近教育观察者身份，强调制度条件、教育公平和社会风险；ChatGPT则以教育研究者身份展开概念区分和结构化论证。三者普遍将潜在威胁性的“取代”问题重构为“角色转型”或“人机协同”问题。由此可见，大语言模型在社会敏感议题中的生成文本具有规范化和调和化倾向。这一倾向可能与指令微调、偏好对齐及本研究的提示语设置有关；它虽有助于缓和职业焦虑，但也可能弱化对平台权力、教育不平等和责任归属的批判。

Abstract: This study examines how large language models construct pragmatic identities and manage rapport when responding to the educationally and professionally sensitive question of whether AI will replace teachers. The data consist of Chinese responses generated by ChatGPT, DeepSeek, and ERNIE Bot (Wenxin Yiyan) under the same standardized prompt. Treating the responses as algorithmically generated discourse with quasi-social interactional functions, the study integrates pragmatic identity theory, rapport management theory, and appraisal-based discourse analysis and employs qualitative close reading supplemented by lexical-item analysis to examine self-positioning, stance-taking, modal mitigation, conditional framing, alignment, and reframing. The findings show that all three models reject the strong claim that AI will completely replace teachers, but they organize this stance through different identities. DeepSeek tends to construct a technical-assistant identity by stressing AI’s instrumental role and its limits in humanistic education. ERNIE Bot adopts an educational-observer identity and foregrounds institutional conditions, educational equity, data security, and implementation risks. ChatGPT constructs an educational-researcher identity through conceptual distinctions and structured reasoning. In rapport management, the three models commonly mitigate categorical judgments and reframe the potentially threatening notion of “replacement” as “role transformation” or “human-AI collaboration”. These patterns indicate a tendency toward normalization and harmonization in LLM-generated discourse. They may be associated with instruction tuning, preference alignment, and the prompt conditions adopted in this study, rather than autonomous model intention. While mitigating occupational anxiety, they may also weaken critical attention to platform power, educational inequality, and responsibility allocation.