COMPARATIVE UX ANALYSIS OF GENERATIVE AI IMAGE USING UTAUT2 AND CONCURRENT THINK ALOUD

Abstract - Generative AI image platforms are gaining popularity for digital content generation but user adoption and actual usability is highly varied from system to system. Three platforms Imagen, DALL-E, and Leonardo. Ai are the focus of this study through a mixed-method research design model: the UTAUT2 model, SEM-PLS model and the Concurrent Think Aloud (CTA) method. 90 respondents were sampled for the quantitative phase and 15 professional designers for the qualitative CTA session. Results from SEM-PLS indicate that Habit (HT) is the best predictor of Behavioural Intention on all platforms. For Imagen and Leonardo. Ai, habit was the only important factor that showed dependence on habitual practice rather than perceived usefulness. DALL-E exhibited a more robust acceptance model, that is, it demonstrated significant impacts on Use Behaviour (UB) from Effort Expectancy (EE) and Behavioural Intention (BI). CTA analysis showed significant problems Imagen had difficulty understanding Indonesian language prompts, while DALL-E had slow-generating and unreliable tasks and Leonardo. Ai also showed basic prompts comprehension issues. Linking these two examples tells us that these two data sets is a huge gulf between perceived ease of use (quantitative), and the sort of real-time user experience described as perceived through qualitative eyes (the time-based user experience). This paper provides new insights on the adoption of generative AI and emphasizes the need for providing additional multilingual processing, system stability and increased interface feedback to improve user experience. Keywords — Concurrent Think Aloud, DALL-E, Generative AI, Imagen, Leonardo. Ai, SEM-PLS, UTAUT2摘要：生成式AI图像平台在数字内容生成领域的认可度与日俱增，但不同系统间的用户采纳度与实际可用性差异显著。本研究以Imagen、DALL-E与Leonardo. Ai三款平台为研究对象，采用混合方法研究框架，涵盖UTAUT2模型、SEM-PLS模型以及并发有声思维法（Concurrent Think Aloud, CTA）。定量研究阶段共招募90名受访者作为研究样本，定性CTA环节则邀请15名专业设计师参与数据采集。SEM-PLS分析结果显示，在三款平台中，习惯（Habit, HT）均为行为意向（Behavioural Intention, BI）的最优预测因子。针对Imagen与Leonardo. Ai而言，习惯是唯一显著的影响因子，其驱动逻辑依赖于用户的习惯性使用行为，而非感知有用性。而DALL-E的用户接受模型则更为稳健：努力期望（Effort Expectancy, EE）与行为意向（BI）均对其使用行为（Use Behaviour, UB）产生显著正向影响。CTA分析结果揭示了多款平台存在的显著问题：Imagen难以理解印尼语提示词，DALL-E存在生成速度缓慢与任务可靠性不足的问题，而Leonardo. Ai同样存在基础提示词理解障碍。将两类分析结果结合来看，定量维度下的感知易用性与通过定性视角捕捉的实时用户体验（即基于时间维度的用户体验）之间存在显著鸿沟。本研究为生成式AI的用户采纳研究提供了全新视角，并强调需通过强化多语言处理能力、提升系统稳定性以及优化界面反馈机制来改善用户体验。 关键词：并发有声思维法（Concurrent Think Aloud, CTA）、DALL-E、生成式AI（Generative AI）、Imagen、Leonardo. Ai、SEM-PLS、UTAUT2