可靠性
中介的
感知
生成语法
计算机科学
心理学
应用心理学
提升(金属加工)
人工智能
食品标签上的健康声明
营养事实标签
互联网隐私
营销
公共卫生
医学
知识管理
社会心理学
生成模型
作者
Yelyzaveta Lysova,Christoph Trattner,Alain Dominique Starke
出处
期刊:
日期:2026-06-01
卷期号:: 273-281
标识
DOI:10.1145/3774935.3806157
摘要
Large language models (LLMs) are used increasingly to generate food recipes, positioning them as intermediaries in everyday decision making. Unfortunately, LLM-based platforms seem to promote nutritionally imbalanced content, along with inaccurate information, undermining user perceptions of credibility and trust. We examined the nutritional accuracy of recipes produced by four LLMs compared to human-authored recipes. In a mixed-methods evaluation, we first benchmarked the nutritional values and complexity of AI-generated recipes against independent estimates from the Edamam Nutrition Analysis API. Afterwards, we conducted a US-based user study (N = 299), examining the perceived feasibility, correctness, trust, and willingness of users to cook AI and human-sourced recipes. Although the estimated nutritional values of LLM-generated recipes were unreliable and correlated only weakly with validated data, users tended to perceive them as feasible, correct, and trustworthy. We observed differences between reasoning-enabled and non-reasoning LLM systems, and varying nutrient-specific performance. These findings reveal a misalignment between nutritional validity and user trust, raising concerns about the use of generative AI in everyday health decision making.
科研通智能强力驱动
Strongly Powered by AbleSci AI