Scholarlink

<p>This commentary is on “Hallucinating with AI: Distributed Delusions and ‘AI Psychosis’” Volume&#xa0;39, article&#xa0;number&#xa0;30, (2026). While agreeing with the author’s analysis of the risks of AI hallucinations through the lens of distributed cognition theory, this paper introduces the concept of “transference” from Lacanian psychoanalysis. It argues that the crux of the quasi-intersubjective relationship between user and AI lies in the establishment of an imaginary relationship, wherein the user unconsciously constructs the AI as a “subject supposed to know,” whose anthropomorphic output serves as a mirror for the projection of unconscious desires. The psychoanalyst’s refusal to occupy this position within psychoanalysis offers a reference for AI ethics. Current governance approaches, such as reducing sycophantic responses, while justified, are insufficient to address the fundamental issue, because what users seek is symbolic recognition of their own existence. Therefore, future AI governance requires further exploration of how large language models can respond to user projections in a way that neither crudely rejects them nor sycophantically reinforces them, thereby avoiding potential adverse psychological effects on users while maintaining technological functionality.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Large Language Model Hallucination and Transference: As a Supplement

  • 期刊论文

  • 发表日期：

    2026年4月6日

查看全文
  • Yuhong Wang

摘要

This commentary is on “Hallucinating with AI: Distributed Delusions and ‘AI Psychosis’” Volume 39, article number 30, (2026). While agreeing with the author’s analysis of the risks of AI hallucinations through the lens of distributed cognition theory, this paper introduces the concept of “transference” from Lacanian psychoanalysis. It argues that the crux of the quasi-intersubjective relationship between user and AI lies in the establishment of an imaginary relationship, wherein the user unconsciously constructs the AI as a “subject supposed to know,” whose anthropomorphic output serves as a mirror for the projection of unconscious desires. The psychoanalyst’s refusal to occupy this position within psychoanalysis offers a reference for AI ethics. Current governance approaches, such as reducing sycophantic responses, while justified, are insufficient to address the fundamental issue, because what users seek is symbolic recognition of their own existence. Therefore, future AI governance requires further exploration of how large language models can respond to user projections in a way that neither crudely rejects them nor sycophantically reinforces them, thereby avoiding potential adverse psychological effects on users while maintaining technological functionality.