武汉船舶职业技术学院学报

基于活动理论框架，探究生成式AI工具（ChatGPT、Grammarly）在信息工程和机械工程专业学术英语写作中的双重效应。通过对160名学生的写作日志、访谈录音及过程录像进行三级编码分析（Cohen's Kappa=0.82），发现：（1）AI使文献处理效率提升40%(p<0.01)，但28%样本因过度依赖出现逻辑连贯性下降，低能力组（IELTS≤6.0）的元认知监控失效显著；（2）行为模式呈现学科分化：筛选重构型在方法论创新性得分最高（Cohen's d=0.76），其核心特征为每千词实施3～5次AI建议批判评估；知识内化型能主动融合AI术语库与领域知识图谱，但高频使用AI(>5次/千词)导致学术元话语密度下降32%(Z=3.78,p<0.01);(3)构建学科导向的AI协同模型，确立人机分工阈值（如公式推导需≥70%人工参与），开发含工具层（语法修正）、策略层（结构优化）与元认知层（自我调节）的动态评估矩阵，并设计双循环伦理机制，通过语料过滤（偏差识别率92%）与过程溯源防控技术风险。研究为STEM学科AI辅助写作教育提供“效率-能力-伦理”协同发展的实证框架。

This study, grounded in the framework of activity theory, explores the dual effects of generative AI tools(ChatGPT, Grammarly) in academic English writing for the fields of information engineering and mechanical engineering. Through a three-level coding analysis(Cohen's Kappa = 0.82) of writing logs, interview recordings, and process videos from 160 students, we found that:(1) AI increased literature processing efficiency by 40%(p < 0.01), but 28% of the samples exhibited a decline in logical coherence due to overreliance, with significant failures in metacognitive monitoring among the low-ability group(IELTS ≤ 6.0);(2) Behavioral patterns revealed disciplinary differentiation: the selection-reconstruction type achieved the highest scores in methodological innovativeness(Cohen's d = 0.76), characterized by 3～5 critical evaluations of AI suggestions per thousand words; the knowledge-internalization type actively integrated the AI terminology database with domain knowledge graphs, but excessive use of AI(>5 times per thousand words) resulted in a 32% decrease in academic metadiscourse density(Z = 3.78, p < 0.01);(3)We constructed a discipline-oriented AI collaboration model, establishing thresholds for human-machine division of labor(e.g., at least 70% human involvement in formula derivation), developing a dynamic assessment matrix that includes a tool layer(grammar correction), strategy layer(structure optimization), and metacognitive layer(self-regulation), and designed a dual-loop ethical mechanism to mitigate technical risks through corpus filtering(with a bias identification rate of 92%) and process tracing. This research provides an empirical framework for the collaborative development of "efficiency-capability-ethics" in AI-assisted writing education within STEM disciplines.