AI 中文总结
该研究评估AI智能体从超市产品图像推断糖含量的能力，发现其准确率因场景和产品类型差异大，全球产品准确率高但本地产品与随机猜测无差异，指出AI营养工具应定位为辅助工具，需对齐本地生态的可审计基准。
AI 中文摘要
营养标签在法律上允许使用极小字体，降低了现实中的可读性，促使消费者依赖“AI营养透镜”和具备视觉能力的对话式智能体来获取饮食指导。我们通过一项有界且可验证的任务评估这种AI中介的建议能否在多大程度上替代受监管的标签：仅从包装正面图像推断两种包装食品中哪一种含糖量更低。我们采用二选一迫选游戏，在四个国家的超市场景（瑞典、美国、澳大利亚、哈萨克斯坦）中评估AI智能体系统。两项智能体共132次对比的结果显示，其性能存在显著的场景依赖差异：对于全球产品，智能体准确率达88.9%（与随机猜测相比p<0.0001）；对于瑞典的本地产品，准确率降至59.5%（p=0.29），此时AI的指导与随机猜测在统计上无差异。这些发现表明存在与训练数据覆盖不均一致的跨市场偏差，引发了对信任、公平性和问责制的担忧，尤其是当营养判断从可审计的公共标签转向专有推理管道时。我们得出结论，AI营养透镜应用更适合被定位为辅助性、教育性工具，而非受监管标签的替代品，我们强调需要与本地食品生态系统对齐的可审计数据集和评估基准。
英文摘要
Nutritional labels are legally permitted to appear in very small print, reducing real-world readability and encouraging consumers to rely on 'AI nutrition lens' and vision-capable conversational agents for dietary guidance. We evaluate whether such AI-mediated advice can meaningfully substitute for regulated labeling using a bounded, verifiable task: inferring which of two packaged foods contains less sugar from front-of-pack images alone. A Two-Alternative Forced Choice game was used to evaluate AI agent systems across four national supermarket contexts: Sweden, the USA, Australia, and Kazakhstan. The results (N=132 comparisons) across both agents reveal a significant performance divide contingent on context. For global products the agents achieved 88.9% accuracy (p < 0.0001 against chance). For local products (Sweden), accuracy dropped to 59.5% (p = 0.29), rendering the AI's guidance statistically indistinguishable from random guessing. These findings indicate a cross-market bias consistent with uneven training-data coverage, raising concerns about trust, equity, and accountability, particularly when nutritional judgment shifts from auditable public labels to proprietary inference pipelines. We conclude that AI nutrition lens applications are better framed as assistive, educational tools rather than as replacements for regulated labels, and we highlight the need for auditable datasets and evaluation benchmarks aligned with local food ecosystems.
CommentsAccepted on April 15th 2026 to the 2026 48th Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC)
