人工智能生成的听力损失与痴呆患者问题回答的可读性

文献信息

中文摘要

目的： 本研究考察人工智能（AI）生成的针对听力损失（HL）、痴呆及二者关联的公众查询回答的可读性。
方法： 这项横断面研究向三个AI平台（ChatGPT、Claude和Perplexity AI）提交了20个标准化的面向患者的关于HL、痴呆及其关系的问题。数据收集时，默认模型分别为GPT-5.3 Instant、Claude Sonnet 4.6和Perplexity AI默认的“最佳”模型选择。所有60个回答均使用两种经过验证的指标评估可读性：Flesch-Kincaid年级水平（FKGL）和Simple Measure of Gobbledygook（SMOG）。美国医学会（AMA）/美国国立卫生研究院（NIH）建议患者教育材料的阅读水平为八年级。使用SPSS 31.0计算各平台和10个主题领域的可读性评分的描述性统计。
结果： 在所查询的三个AI平台中，听力相关问题的回答FKGL中位可读性处于高中高年级至大专年级水平（12.3-15.3），SMOG低1-2个年级水平（11.4-13.0）。对于痴呆回答，FKGL可读性为15.2-16.3年级水平，SMOG为12.9-14.1。总体及各领域均未发现AI平台间有统计学显著差异。Claude在HL和痴呆的症状及疾病管理领域大多获得高年级水平评分（FKGL和SMOG）。
结论： 公众使用AI聊天机器人寻求健康信息，但本研究发现ChatGPT、Claude和Perplexity AI对HL和痴呆问题生成的回答比AMA/NIH推荐的八年级基准高3-8个年级水平。这些发现与越来越多的文献一致，表明AI生成的健康内容在许多医学专科中常规超出推荐的可读性阈值。这对受HL和痴呆影响最大的老龄人群尤其令人担忧。作者敦促相关专业组织审查其面向公众材料的可读性合规性。

英文摘要

OBJECTIVE: This study examines the readability of artificial intelligence (AI)-generated responses to public queries regarding hearing loss (HL), dementia, and the link between them.
METHODS: This cross-sectional study submitted 20 standardized patient-facing questions about HL, dementia, and their relationship, to three AI platforms (ChatGPT, Claude, and Perplexity AI). At the time of data collection, the default models were, respectively, GPT-5.3 Instant, Claude Sonnet 4.6, and Perplexity AI default "best" model selection. All 60 responses were assessed for readability using two validated metrics, Flesch-Kincaid Grade Level (FKGL) and Simple Measure of Gobbledygook (SMOG). The American Medical Association (AMA)/National Institutes of Health (NIH)-recommends eighth-grade reading level for patient education materials. Descriptive statistics were calculated for readability scores across platforms and 10 topic domains (SPSS 31.0).
RESULTS: Across the three AI platforms queried, FKGL median readability for responses to hearing-related questions was at the upper high school to junior college grade level (12.3-15.3) and 1-2 grade levels lower with SMOG (11.4-13.0). For dementia responses, the readability was at 15.2-16.3 grade level with FKGL and 12.9-14.1 using SMOG. No statistically significant differences were found between AI platforms overall and by domain. Claude mostly received high grade level scores for domains regarding symptoms and condition management, for both HL and dementia with FKGL and SMOG.
CONCLUSION: AI chatbots are used by the public to seek health information, yet this study found that ChatGPT, Claude, and Perplexity AI produced responses to questions about HL and dementia at 3-8 grade levels above the AMA/NIH-recommended eighth-grade benchmark. These findings align with a growing literature showing that AI-generated health content routinely exceeds recommended readability thresholds across many medical specialties. This is particularly concerning for the aging populations most affected by HL and dementia. The authors urge relevant professional organizations to audit their public-facing materials for readability compliance.