1. Background and Research Focus

In recent years, digital humanities (DH) research has shifted from large-scale data organization, text mining, and visualization analysis toward human-AI co-creation and the interpretation of meaning following the intervention of generative AI (GenAI). When GenAI is no longer merely a back-end tool but is capable of generating texts, images, and narrative frameworks, humanities research must reconsider who constructs cultural meaning, how interpretive responsibility is distributed, and how AI bias can be identified and corrected. This study originates from the creative project KNARPAY in Transmission: Beitou Society, Where Are You? by the Tree Tree Tree Person team(森人). It focuses on the belief culture of the Ketagalan people in Beitou, particularly the local historical memories embodied in Chi Wangye(池府王爺) and the Plains Indigenous(平 埔族) Earth God(土地公) worshipped at Baode Temple(保德宮). Since the Qing period( 清朝), these memories have gradually been obscured through processes of migration and settlement, colonial governance, and urbanization, surviving only fragmentarily in local historical writings, temple beliefs, oral memories, and research materials.

2.Theoretical Lens

Responding to Smith’s (2021) perspective on decolonization, this study regards GenAI as a cultural technology that must be examined and recontextualized, rather than as a neutral tool for knowledge production (Mohamed, Png, & Isaac, 2020). Through the use of GenAI to assist picture book creation, this study transforms the beliefs of Plains Indigenous peoples in Beitou and historically complex materials into an accessible narrative form for educational outreach. It also analyzes the usability of GenAI in local cultural translation, its cultural biases and symbolic misplacements, and how researchers may maintain interpretive agency and cultural appropriateness through a Human-in-the-Loop (HITL) iterative process.

This study adopts computational hermeneutics as its primary theoretical perspective. Kommers et al. argue that GenAI should not be understood merely as an automation tool, but also as a form of “cultural technology.” In processing texts, images, and narratives, GenAI participates in the organization and recontextualization of meaning. Therefore, GenAI-generated outputs are not neutral results; rather, they are produced through the interaction of data, models, prompts, user judgment, and cultural context (Kommers et al., 2026).

Computational hermeneutics emphasizes situatedness, plurality, and ambiguity. It argues that meaning must be understood within specific historical and social contexts, and that the same cultural material may allow for multiple valid interpretations. This perspective is particularly suitable for analyzing local beliefs and ethnic memories, as these are not closed, singular, or fully standardizable objects of knowledge. In addition, the concept of agency in human-AI co-creation helps explain the division of labor between humans and GenAI in this study. GenAI may function as a collaborator in ideation and material generation, yet cultural judgment, symbolic correction, narrative selection, and final responsibility remain with the researcher (Zhang, Wang, & Yi, 2025). Accordingly, “cultural appropriateness” in this paper does not refer to technical accuracy, but to whether GenAI-generated outputs correspond to specific cultural contexts, respect the historical complexity of local beliefs, and avoid reducing local culture to exoticized styles or stereotypical symbols.

3. Methodology and Workflow

The application of GenAI in this study is primarily framed through human-AI co-creation and incorporates the operational logic of Human-in-the-Loop (HITL). GenAI participates in text generation, visual ideation, and narrative translation, while the researcher continuously intervenes in the correction and interpretation of generated outputs through prompt design, cultural interpretation, content selection, and recontextualization. In other words, this paper regards human-AI co-creation as the main relational model for creative and knowledge production, while HITL is understood as a risk-control mechanism for maintaining historical accuracy, cultural sensitivity, and educational appropriateness.

The workflow consists of three layers: knowledge extraction, human-AI co-creation, and final production. First, the researcher establishes the narrative foundation through field observation, historical documentation, and contextual analysis of local beliefs. The textual materials are then input into ChatGPT and Gemini for script ideation, character development, and visual draft generation. Finally, the work is primarily completed through hand drawing in Procreate, followed by refinement, layout, and output using Photoshop, Illustrator, and InDesign. In practice, the researcher controls GenAI bias through prompt guidance and iterative revision. If the output corresponds to the cultural context, it proceeds to the production stage; if symbolic misplacement, cultural bias, or aesthetic inappropriateness occurs, the process returns to manual correction and prompt adjustment, as shown in Figure 1.

4. Findings

This study finds that when GenAI encounters highly localized materials such as the beliefs of Plains Indigenous peoples in Beitou, it can rapidly generate texts and images, yet it often reverts to generic cultural symbols, creating “ cultural friction” between globalized technology and local knowledge (Tsing, 2005). These biases include the generalization of deity images, confusion among ethnic symbols, and misplacement of religious atmosphere. The problem lies not only in the inaccuracy of images or texts, but also in the possibility that the model may replace local cultural contexts with dominant visual vocabularies.

5.Contribution and Implications

Therefore, GenAI errors should not merely be regarded as technical failures. They may also serve as analytical entry points, revealing how local cultures are ignored, replaced, or flattened in the generative process. This paper positions GenAI as a “controlled interpretive partner” rather than as the subject of cultural interpretation. GenAI can provide materials, associations, and clues to bias, but cultural judgment, appropriateness checks, and the responsibility for translation must remain with the researcher. The contribution of this study does not lie in proving that GenAI can replace the work of cultural interpretation. Rather, it proposes a GenAI-assisted model for cultural translation centered on researcher leadership, HITL control, and cultural appropriateness assessment. This model may serve as a reference for future applications in local cultural education and cultural heritage education.

Figure 1. An Iterative Workflow Model for GenAI-Assisted Cultural Translation

Based on the HITL Approach. (Source: Developed by the author.)