一句话：用 LLM 给 84 个 AI 生成的 Web UI 原型打分，并与 39 位 HCI 专家评分对比，发现 RMI 最接近人类而 GPT-4o 表现不佳。
生成式 AI 已越来越多地用于 UI/UX 工程，但其自动化程度落后于其他软件工程活动。UI 原型的迭代流程包含创建与评估两个环节，而当前的 GUI Agent 缺乏对美学等主观质量维度的评判能力。作者让 ChatGPT 与 RateMyImage（RMI）对 84 个 AI 生成的 Web UI 原型进行评分，并以 39 位 HCI 专家提供的人类评分作为基准进行对标。结果显示，RMI 的评分与人类最接近（r = 0.546），而 GPT-4o 与人类参与者的一致性较差、内部一致性也较差。融合 LLM 之前的 UI 特征与模型评分后，预测人类主观评分的最优（AIC 最优）回归模型达到 R² = 0.423。论文最后给出了供 UI/UX 工程师在借助 AI 评估设计原型时参考的初步指南。
GenAI services are being increasingly employed in UI/UX engineering, whose automation remains insufficiently advanced compared to other Software Engineering activities. The iterative process involves creating and assessing UI prototypes, but up-to-date GUI Agents lack in evaluating subjective quality dimensions like aesthetics. We had ChatGPT and RateMyImage (RMI) rate 84 web AI-generated UI prototypes, and then benchmarked their ratings with human ones, provided by 39 HCI experts. RMI rated the most similarly to humans (r = 0.546), while GPT-4o showed poor agreement with human participants and poor internal consistency. The best (AIC-optimal) regression model for humans’ subjective ratings that incorporated both pre-LLM UI features and models’ ratings had R2=0.423. We also provide some preliminary guidelines for UI/UX engineers seeking to involve AI in assessing design prototypes.