International Open Access Journal Platform

大语言模型中针对老年群体的 “善意年龄歧视”研究

Benevolent Ageism in Large Language Models: A Comparative Analysis of Response Patterns to Young and Older Adult Users

张小玲

Psychology of China / 2026,8(4): 433-437 / 2026-04-16 31 23

View Download PDF

• Information:

  友邦人寿保险有限公司，上海

• Keywords:
  Large language model; Benevolent ageism; Elderspeak; Stereotype Content model; Human-computer interaction; Algorithmic bias
  大语言模型; 善意年龄歧视; 老龄语体; 刻板印象内容模型; 人机交互; 算法偏见
• Abstract: With the increasing integration of Generative Artificial Intelligence into daily life, whether their outputs harbor implicit social biases has become a focal point of research in psychology and human-computer interaction. This study, through a controlled experiment, aims to investigate whether Large Language Models exhibit systematic differences in their response patterns when interacting with users of different ages (25 years vs. 65 years). By analyzing the output texts across three scenarios—technology instruction, skill acquisition, and financial advice—the study found that LLMs display significant Benevolent Ageism towards older users. This bias is specifically manifested in the excessive use of Elderspeak, communication strategies based on the Deficit Hypothesis, and a tendency towards Excessive Risk Aversion in advice. The research suggests that although AI exhibits high Warmth in interactions, its default settings undermine the perceived Competence of older users. This “digital over-accommodation” may not only exacerbate Stereotype Threat for the elderly but also indirectly reinforce the Digital Divide. 随着生成式人工智能（GenerativeAI）在日常生活中的普及，其输出内容是否隐含社会偏见已成为心理学与人机交互领域的研究焦点。本研究通过一项控制变量实验，旨在探究大语言模型（LLM）在与不同年龄用户（25岁vs.65岁）互动时，其回应模式是否存在系统性差异。通过对科技教学、学习建议、理财建议三个场景的输出文本进行分析，研究发现，LLM对老年用户表现出显著的“善意年龄歧视”（BenevolentAgeism）。这种偏见具体表现为“老龄语体”（Elderspeak）的过度使用、基于能力折损假设（DeficitHypothesis）的沟通策略，以及过度风险规避（ExcessiveRiskAversion）的建议倾向。研究认为，AI在交互中虽然表现出高亲和性（Warmth），但其默认设定削弱了老年用户的胜任力（Competence）感知，这种“数字化过度照护”不仅可能加剧针对老年群体的刻板印象威胁（StereotypeThreat），还可能间接固化数字鸿沟（DigitalDivide）。
• DOI: 10.35534/pc.0804067 (registering DOI)
• Cite: 张小玲. (2026). 大语言模型中针对老年群体的“善意年龄歧视”研究. 中国心理学前沿, 8 (4), 433-437.