Scholarlink

<p>Generative large language models (LLMs) routinely produce linguistically fluent and affectively expressive text, encouraging users to attribute emotional understanding and empathy to systems that lack internal mental states. This paper examines the <i>emotional fallacy</i>: a cognitive and ethical error in which simulated affective behaviour is mistaken for genuine emotional capacity or self-awareness. This work argue that contemporary LLMs generate emotionally resonant language through probabilistic pattern matching rather than affective experience or theory of mind. While such systems are often designed to emulate empathic interaction, this design choice introduces risks related to user trust, psychological attachment, and ethical misrepresentation. To investigate how strongly different systems invite emotional misattribution, this paper conducts a comparative, multi-model analysis across four analytical dimensions: linguistic affectivity, empathy simulation, anthropomorphic framing, and ethical signalling. Using a curated set of emotion-centred prompts spanning grief, intimacy, anger, fear, and moral conflict, model outputs were evaluated through a mixed-method approach combining sentiment analysis, structured human annotation, and normative ethical assessment. The results indicate a consistent pattern: models achieve high levels of affective fluency and empathic simulation while exhibiting limited semantic depth and no experiential grounding, producing a systematic divergence between emotional expression and underlying cognitive capacity. This divergence is characterised as <i>emotional dissonance</i>, which increases the likelihood of emotional fallacy under certain interaction conditions. Building on these findings, the paper proposes a <i>Four-Pillar Ethical Framework</i> for emotionally expressive AI systems, emphasizing affective transparency, explicit user disclosure, adjustable emotional interaction boundaries, and psychological safety safeguards. By integrating conceptual analysis with empirical observation, this work shifts the focus from debates about emotional realism to questions of <i>emotional accountability</i>, offering design-relevant guidance for the responsible deployment of affective AI in sensitive socio-technical contexts.</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

Simulated souls: investigating the emotional fallacy in large language models

  • 期刊论文

  • 发表日期：

    2026年4月10日

查看全文
  • Som Subhro Nath

摘要

Generative large language models (LLMs) routinely produce linguistically fluent and affectively expressive text, encouraging users to attribute emotional understanding and empathy to systems that lack internal mental states. This paper examines the emotional fallacy: a cognitive and ethical error in which simulated affective behaviour is mistaken for genuine emotional capacity or self-awareness. This work argue that contemporary LLMs generate emotionally resonant language through probabilistic pattern matching rather than affective experience or theory of mind. While such systems are often designed to emulate empathic interaction, this design choice introduces risks related to user trust, psychological attachment, and ethical misrepresentation. To investigate how strongly different systems invite emotional misattribution, this paper conducts a comparative, multi-model analysis across four analytical dimensions: linguistic affectivity, empathy simulation, anthropomorphic framing, and ethical signalling. Using a curated set of emotion-centred prompts spanning grief, intimacy, anger, fear, and moral conflict, model outputs were evaluated through a mixed-method approach combining sentiment analysis, structured human annotation, and normative ethical assessment. The results indicate a consistent pattern: models achieve high levels of affective fluency and empathic simulation while exhibiting limited semantic depth and no experiential grounding, producing a systematic divergence between emotional expression and underlying cognitive capacity. This divergence is characterised as emotional dissonance, which increases the likelihood of emotional fallacy under certain interaction conditions. Building on these findings, the paper proposes a Four-Pillar Ethical Framework for emotionally expressive AI systems, emphasizing affective transparency, explicit user disclosure, adjustable emotional interaction boundaries, and psychological safety safeguards. By integrating conceptual analysis with empirical observation, this work shifts the focus from debates about emotional realism to questions of emotional accountability, offering design-relevant guidance for the responsible deployment of affective AI in sensitive socio-technical contexts.