AI辅助写作的元认知代价:监控准确性与自我调节的链式损耗
The Metacognitive Cost of AI Assistance: Cascading Impairments in Monitoring Accuracy and Self-Regulation
摘要
随着生成式人工智能广泛融入各行业信息搜集、整合与文书撰写等知识工作,个体在任务表现受益的同时,也可能为此付出元认知代价。基于元认知监控–控制框架,本研究选取论述写作与修改作为知识工作典型任务,考察AI辅助写作是否会削弱个体对自身作品质量的判断准确性,并进一步损害后续修改行为的自我调节效果。实验1在文本写作任务情境中探索AI使用对元认知监控准确性的影响,发现AI辅助写作在质量和创新两个维度上显著降低了元认知监控准确性。实验2进一步引入自主修改阶段,探索AI辅助是否影响后续自我调节行为。结果表明,AI辅助组更难准确选择初始得分更低的文章进入修改阶段,且修改后的得分提升幅度较小,初稿阶段的元认知监控准确性起到中介作用。研究将元认知框架拓展至人机协作情境,深化了对生成式AI辅助如何影响元认知的理论认识,提示人机协作评价不应仅关注一次性产出,也应警惕AI使用的元认知代价。
Abstract
As generative artificial intelligence becomes widely embedded in knowledge work across industriesincluding information gathering, synthesis, and document writingits benefits to task performance may come with an underexamined metacognitive cost: people may perform well with assistance while losing the ability to accurately monitor how well they themselves are doing. Drawing on the monitoringcontrol framework of metacognition, the present research took argumentative writing and revision as representative knowledge-work tasks and examined whether AI-assisted writing undermines individuals' accuracy in judging the quality of their own work and, in turn, impairs the self-regulatory efficiency of downstream revision behavior.Two experiments were conducted on the Credamo platform. Before data collection, sample sizes were determined through power analysis. In Study 1, participants were randomly assigned to an AI-use condition or a no-AI condition and completed 10 short writing tasks, each requiring a 200 to 300 Chinese characters response. After each text, participants rated the quality and originality of their own writing. Three trained raters independently evaluated each text, and standardized rater scores were aggregated into objective performance indices. Relative metacognitive accuracy was indexed primarily by GoodmanKruskal gamma, computed across each participant's 10 texts. Study 2 extended this design by introducing a delayed free-revision stage after a distractor task, allowing participants to choose which texts to revise.Across both studies, AI assistance did not reduce objective performance and improved some aspects of subjective or rater-based evaluations. However, AI users showed lower metacognitive monitoring accuracy: their self-evaluations tracked the relative quality and originality of their texts less effectively than those of participants who wrote without AI. Study 2 replicated this monitoring deficit and further showed that it carried over into regulation. Participants in the no-AI condition were more likely to select objectively weaker texts for revision, whereas AI users showed poorer selectivity and in some cases chose texts with slightly higher initial scores than the ones they left unchanged. In addition, the performance gains produced by revision were smaller in the AI condition. Monitoring accuracy at the drafting stage partially mediated the smaller revision gains in the AI condition.These findings suggest that the metacognitive cost of AI assistance is not limited to whether people can produce better outcomes with technological support. A more fundamental issue is whether people can judge the relative quality of their work and identify where additional effort is needed. Theoretically, by extending the monitoringcontrol framework to humanAI collaboration, the results deepen understanding of how generative AI assistance shapes metacognition. Practically, they suggest that AI systems should not only generate content but also preserve or scaffold users' diagnostic cues, so that improved output does not come at the expense of self-monitoring and self-regulationand that evaluations of humanAI collaboration should not focus solely on one-off output, but should also remain alert to the metacognitive costs of AI use.关键词
生成式人工智能/元认知监控/元认知准确性/自我调节Key words
Generative artificial intelligence/Metacognitive monitoring/Self-regulation/Monitoring accuracy引用本文复制引用
刘慕义,张志学.AI辅助写作的元认知代价:监控准确性与自我调节的链式损耗[EB/OL].(2026-11-15)[2026-11-17].https://chinaxiv.org/abs/202611.00137.学科分类
计算技术、计算机技术