生成式AI輔助下敘事影像創作的人機互動研究：專業背景與工具熟悉度交互作用下的創作探討

Abstract

生成式人工智慧（generative artificial intelligence，GAI）技術迅速發展，特別是在生成對抗網絡（GAN）、變分自編碼器（VAE）與基於Transformer的模型（如GPT、DALL·E）等技術突破下，促使GAI廣泛應用於各種領域。此類工具在創作情境中展現出降低技術門檻、提升創作效率與促進個性化敘事的潛力。然而，GAI的應用亦引發創意同質化與創作者主體性受限等問題，其對不同創作者在人機互動層面的實質影響，仍有待深入探討。本研究以數位敘事影像創作為核心任務，讓創作者實際操作多種生成式AI工具，並採用混合方法設計，結合質性與量化資料，以影像創作專業背景與AI工具熟悉度為分類依據，將參與者劃分為四類創作者（專熟組、專生組、非熟組、非生組）。研究透過創作行為觀察、問卷調查與半結構式訪談等方式，分析各組在創作策略、操作模式與成果表現上的差異。研究結果顯示：專熟組展現出高度的創作自主性與流程調整能力，能有效結合AI工具進行複雜敘事建構；專生組雖操作經驗有限，但具備明確的專業素養與自我要求，能在過程中快速適應並取得良好成果；非熟組雖具操作直覺性與流暢性，卻因缺乏系統化創作理解，在敘事深度上受限；非生組則多依賴AI預設內容，作品易呈現樣式同質化。整體而言，創作專業背景與自我要求是影響創作成果品質的關鍵因素，而AI工具的易用性在一定程度上能彌補經驗不足，使各類創作者皆能完成基本創作任務。進一步分析也指出：專家組多將AI視為輔助性工具，重視生成結果與預期構想的一致性，並強調細節修正與流程掌控；非專家組則展現更高的開放性與彈性，傾向從AI意外生成中尋求靈感與創意激發。綜合上述實證結果，本研究根據Wallas的創作四階段模型，建構出一套AI輔助影像創作的心智模型，描繪創作者在創作不同階段中與AI互動的思維與決策機制，作為理解人機協作邏輯的參考依據。此外，亦進一步提出一個基於創作專業背景與AI工具熟悉度所劃分的四象限人機協作框架，以利未來教育者與創作者辨識不同創作主體之特性，並據以規劃適切的AI協作策略與教學設計，促進創作實踐與AI應用的有效整合。Generative artificial intelligence (GAI) technologies have rapidly evolved, particularly through advancements in GANs, VAEs, and Transformer-based models such as GPT and DALL·E. These tools are increasingly applied across domains and exhibit strong potential in creative fields by reducing technical barriers, improving efficiency, and enabling personalized narrative video expression. However, concerns have emerged regarding creative homogenization and the erosion of authorial agency. The influence of GAI on creator–AI interaction in narrative video and animation production remains insufficiently explored. This study investigates digital narrative video creation through hands-on engagement with multiple generative AI tools. Using a mixed-methods approach, participants were classified into four creator types based on their experience in digital animation creation and familiarity with AI tools: expert creators familiar with AI (E-AI+), expert creators unfamiliar with AI (E-AI-), non-expert creators familiar with AI (N-AI+), and non-expert creators unfamiliar with AI (N-AI-). Data were collected through behavioral observations, surveys, and semi-structured interviews to examine creative strategies, tool usage patterns, and characteristics of the generated content. Results indicate that expert-skilled creators exhibited high autonomy and effectively used AI to construct complex narratives, while expert-novices adapted using domain knowledge despite limited tool experience. Non-expert-skilled creators demonstrated fluent tool operation but lacked structural depth in their outputs, whereas non-expert-novices tended to rely on default settings, leading to more homogenized results. Professional background and intrinsic motivation emerged as key determinants of output quality, and user-friendly design helped bridge experience gaps. Further analysis revealed that expert creators utilized AI for precision and control, while non-experts valued the generative unpredictability of AI as a source of inspiration. Based on these findings, this study proposes a cognitive model of AI-assisted creation grounded in Wallas’ four-stage theory, describing creator engagement across the phases of ideation, incubation, illumination, and verification. Additionally, a four-quadrant collaboration framework is introduced to guide educators and practitioners in identifying creator types and developing appropriate AI integration strategies.