Development and preliminary validation of a scale for evaluating the effectiveness of generative AI in museum cultural and creative product design

The technological revolution, particularly through the application of artificial intelligence (AI), has propelled a cultural renaissance that integrates traditional culture into a new phase characterized by digitalization, intelligence and automation (Liu and Wang, 2025). As indispensable repositories of human civilization, museums worldwide are transforming from static guardians of artifacts into dynamic hubs of cultural dissemination (Chunlan et al., 2025). Today, they function not only as exhibition halls showcasing national civilization but also as creators of aesthetic values and conveyors of the contemporary spirit. Among these developments is the creation of modern cultural and creative products through the innovative transformation of ancient cultural elements. This has become a vital strategy through which museums aim to bridge ancient civilizations and modern society, achieving mutual cultural and economic benefits. From pilot initiatives at the Palace Museum to parallel advancements in provincial and municipal museums, key cultural and creative products have not only generated substantial market revenue but also created new bridges between public engagement and history. As such, they bring artifacts to life beyond the museum walls (Wu et al., 2025).

However, developing the cultural and creative products of museums is no trivial endeavor (Cheng et al., 2024). It demands more than mere symbol collage or form imitation. It requires a profound interpretation of the historical, artistic and scientific value of artifacts, alongside creative transformation through modern elements (Qian et al., 2026). This process requires developers to possess cultural literacy, market insight and innovative design capabilities. Traditional development models frequently encounter obstacles. Many museum cultural products rely on layered subcontracting through creative agencies whose designers lack a deep understanding of the cultural significance of artifacts, resulting in superficial designs that become merely attention-grabbing gimmicks. Simultaneously, constrained by tight budgets and expensive innovation, the long cycle from initial design to prototyping and production also entails high trial-and-error costs. Furthermore, museum cultural and creative development risks homogenization, often being confined to traditional patterns and forms. This makes it difficult to consistently launch innovative products (Tkalich et al., 2025).

Given the rapid advancement of digital media, the cultural heritage sector is experiencing unprecedented transformation (Buragohain et al., 2024). The rise of generative AI technology not only enhances the efficiency of cultural product development but also accelerates the integration of culture and technology. As a cutting-edge field of AI, generative AI primarily learns from vast data sets to generate novel content in various domains, including text, images, audio, video and code. This is evidenced by its innovative application in such varied fields as healthcare, education, finance, agriculture and construction (Yazdani et al., 2025). Particularly with the widespread adoption of prominent Chinese domestic models, generative AI is swiftly permeating all aspects of product design and creation. Exemplifying the new productive forces, generative AI not only replicates existing styles but also ingeniously combines concepts and expands creative horizons, thus demonstrating human-like creativity (Chang and Tung, 2025).

By deeply learning from images of cultural relics and historical documents, generative AI rapidly generates creative proposals based on user prompts, substantially shortening design cycles. It can also dynamically optimize design directions using real-time social media data to enhance market relevance. Crucially, the widespread adoption of generative AI tools now enables small and medium-sized museums to initiate high-quality cultural and creative product development projects, promoting the equitable transformation of cultural resources. As digital technology and AI converge with the extensive and invaluable digital collections of museums, it is interesting to imagine which innovative methods for educational and cultural dissemination will arise. It remains unclear whether generative AI can serve as a proxy designer in creating cultural products for museums. Another key question is whether generative AI will accurately interpret the historical, cultural and national beliefs embedded in artifacts while designing popular products that preserve their essence and enhance the interactive experience, in line with modern aesthetics.

Theoretical possibilities have attained preliminary industry validation, stimulating widespread enthusiasm for their application. Some museums have begun implementing generative AI in support of curation and visual design (Heigl, 2025). These initiatives underscore the vast potential of this technology to ignite creativity and boost efficiency. Academic debates have summarized the principles, boundaries and ethical issues linked to generative AI applications, yet discussions often remain confined to idealized technical descriptions or isolated successes (Hagendorff, 2024). Despite extensive research into the capabilities of generative AI, a limited understanding still exists concerning its real-world performance, advantages and inherent limitations. The disconnect between application and evaluation involves the risk of turning this technology into a mere technical display, thereby hindering its translation into replicable, scalable methodologies.

Meanwhile, empowering museum cultural and creative design through generative AI is a multi-stage, complex process (Xu et al., 2025). It begins with understanding the artifacts themselves, transitioning into the extensive interpretation of their cultural significance, and ultimately culminating in the transformation of creative concepts. However, the currently fragmented studies of effectiveness frequently concentrate on isolated dimensions, evaluating separately the informational quality of generated content, precision of instructional descriptions or responsiveness of execution (Amini et al., 2024). These disparate approaches mean that the true capabilities of generative AI and its underlying mechanisms have not been fully revealed.

This study addresses these academic shortcomings by using the Delphi method. Integrating insights from experts across diverse fields, the aim was to construct a multidimensional, multi-structural evaluation system, as well as develop corresponding scales. Here, the term “museum cultural and creative products” refers to the commercial and educational merchandise − for instance, replicas, stationery, accessories, souvenirs and digital derivatives − that museums develop by reinterpreting the objects in their collections for the purposes of everyday use and public consumption. From the perspective of library and information science, each of these products is, essentially, a repackaging of cultural heritage information: the descriptive and contextual content carried by an artifact, including its provenance, materials, iconography, historical context and cultural meaning. Therefore, designing these products using generative AI depends on the accuracy with which the technology can extract, interpret, organize and re-express this cultural heritage information; this is precisely what the proposed scale is intended to evaluate. The aim is to address the practical challenges involved in applying generative AI to museum cultural and creative design. This would provide a quantifiable, comparable academic framework and action guide enabling technology empowerment. This shift will progress AI-assisted design from an experience-based to a science-based approach. From the perspective of library and information science, this study contributes to the understanding of how generative AI can interpret, organize and repurpose cultural heritage information. In doing so, it addresses several key concerns about information quality assessment and knowledge representation in digital cultural heritage contexts.

Early generative AI technologies were relatively primitive, and their application in museum cultural and creative development was often limited to simple imitation, making the capture of deeper cultural connotations somewhat difficult (Menotti, 2025). Research and practice focused primarily on the tool-like attributes of AI, emphasizing its capacity to improve efficiency (Madanchian and Taherdoost, 2025). Studies have documented processes whereby large language models were leveraged via prompt engineering to batch-generate numerous design sketches, from which viable options were selected (Suh et al., 2024). With technological iteration, generative AI gradually acquired cross-modal understanding and semantic association capabilities. This enabled auxiliary design support ranging from artifact feature extraction to creative extension. Some scholars even predict that AI will potentially catalyze a wave of technological revolution and industrial transformation, while also acting as a pivotal driver of future museum development (Hang and Li, 2025).

Recent practices have illustrated that AI tools can not only craft visual elements that closely align with historical and cultural contexts but also produce narrative-rich content specifically tailored to those contexts. This approach enhances the spiritual depth and user experience of cultural and creative products while substantially reducing time spent on creative ideation; this allows developers to explore an extensive range of cultural element combinations (He et al., 2025). Generative AI has been transformed from a simple tool into a technological intermediary, reshaping the principles and logic of cultural and creative design.

The subsequent phase of technological advancement introduced a new paradigm of human−machine collaborative development (Nampalli, 2025). No longer is generative AI merely a tool that executes commands; it now functions as a cooperative entity endowed with autonomous decision-making capabilities throughout the development process. Scholars have proposed that designers shift from being command issuers to becoming editors who combine the tasks of filtering, optimizing and integrating the extensive output of generative AI while activating their professional expertise and aesthetic judgment at pivotal decision points (Shah et al., 2025). This model balances the efficiency of generative AI with the expertise of human developers, creating a new paradigm of human−machine collaboration. Within this paradigm, developers establish value orientations to guide AI toward creative exploration, while AI leverages its robust data processing and generation capabilities to deliver a range of solutions (Nurfaizal et al., 2018). Studies have demonstrated that human−machine collaborative systems, particularly those incorporating continual learning and advanced algorithms, significantly enhance operational efficiency and quality in complex and dynamic environments (Hou et al., 2025). Especially in translating cultural symbols, AI technology effectively integrates historical contexts with contemporary aesthetic trends to create high-quality cultural and creative products (Yang et al., 2025). This connection broadens artistic expression, offering developers both innovation and convenience. It encourages a mutually supportive relationship between creativity and technology, revealing extraordinary possibilities.

Nevertheless, opportunities and risks coexist, with some scholars highlighting media regulation issues in generative AI applications. They have particularly cautioned against potential cultural misinterpretations and tendencies toward historical nihilism (Abbas et al., 2025). Concurrently, researchers have begun examining whether generative AI-driven cultural and creative development can accurately convey the core values of civilizations (Bu et al., 2025). Caution must be exercised in the face of algorithmic biases that can cause the desecration of cultural heritage, as evidenced by AI’s potential distortion of cultural icons and creation of inappropriate content. The vulgarization of traditional culture must be avoided (Alivizatou, 2011). Particularly when addressing sensitive subjects, AI-generated content requires meticulous human review. From this standpoint, value alignment is a pivotal issue across most fields.

Despite extensive research exploring the practical applications and potential of generative AI to enhance cultural and creative development in museums, there remains a lack of comprehensive assessments of how it can contribute to the preservation and innovation of cultural heritage. No systematic, quantifiable academic framework exists for defining what constitutes high-quality outcomes in generative AI development. Building upon a review of prior literature, this study argues that the key obstacle to the intelligent development of generative AI is not technological limitations but the lack of a model that accurately assesses its enabling effects. Fragmented, subjective evaluations fail to provide reliable guidance for practice, and they hinder the transformation of generative AI from an emerging technology into a reliable, positively productive force.

Human−computer interaction (HCI) theory emphasizes the bidirectional constructive relationship between people and technological systems. Technological development follows human intentions and cognitive patterns, while human behaviors continuously shape the evolutionary path of technology. The Data, Information, Knowledge, Wisdom and Purpose (DIKWP) model segments cognitive processes into five distinct layers, each playing a unique role in the progression from raw data to strategic decision-making. This has been interpreted as a bottom-up, unidirectional enhancement process forming a fully connected cognitive network. The model offers a preliminary framework for understanding how generative AI structurally achieves endogenous self-driving (Schembera, 2025). A higher-level purpose provides direction and evaluation criteria for cognition, ensuring that the perception-cognition-decision loop aligns with system objectives. Meter (2020) suggested that human engagement with AI that has a semantic personality requires the alignment of two DIKWP systems at the personality level. AI demonstrates distinct narrative styles and value orientations, manifesting its individuality through knowledge and information outputs (K/I layer). Such synergy could positively influence interaction outcomes.

Beaudouin-Lafon et al. (2021) introduced interaction generative theory (IGT), a novel framework for human−machine collaboration. It offers new insights into HCI research. IGT facilitates the understanding of current applications and envisioning of future systems through three progressive perspectives: analytical, critical and constructive. The former examines dynamic relationships among elements of HCI, revealing how technological behaviors align with user intentions. The critical outlook evaluates the limitations of the existing interaction models, encouraging reflection on generative AI and deeper content interpretation. Finally, the constructive perspective focuses on design principles for future systems, emphasizing co-evolution through feedback loop mechanisms. IGT provides a framework for assessing the role of generative AI in the cultural and creative development of museums. It reveals the fundamental technical logic behind AI-generated content and established the foundation for operational evaluations of intelligent communication effectiveness. Figure 1 presents further details.

In other words, the three perspectives of IGT can be translated into a systematic examination of generative AI spanning cultural understanding, value reflection and collaborative creation. This demands that AI not only recognizes the formal characteristics of cultural relics but also contextualizes them historically. This requires AI to develop progressive response mechanisms through human-like reasoning, which would enable users to guide AI in generating creative solutions that align with the inherent meanings of relics and the narrative logic of Chinese culture (Qian et al., 2026). More specifically, AI ought to extract surface-level symbols, interpret fundamental factual information and clarify the underlying cultural logic, thus generating creative output grounded in these foundational steps (Huang et al., 2026). This demonstrates that evaluating the intelligent performance of AI cannot rely on a single structural dimension. Instead, a comprehensive examination is required of its multidimensional manifestations in specific practices to conduct a systematic assessment.

The Delphi technique is a powerful forecasting and decision-making method. It empowers experts to voice their perspectives freely, ensuring that their insights remain uninfluenced by dissenting opinions. By fostering independent thought, the Delphi technique enhances the quality and reliability of collective decision-making. Through multiple rounds of opinion solicitation, statistical feedback and viewpoint refinement, divergent opinions converge to form a group consensus (Jorm, 2025). The research process must adhere to the following requirements:

• Complete anonymity must be maintained, with expert identities strictly confidential and participants unable to communicate with each other; this effectively mitigates group decision biases like the authority effect and conformity.

• Iterative controlled feedback is essential, whereby each round of survey results undergoes statistical analysis to generate responses; this guides experts to deliver targeted reflection based on the distribution of group opinions.

• Expert consensus must be the focus, which is achieved through multiple cycles that gradually converge opinions, resulting in a highly reliable collective judgment.

The effectiveness evaluation scale for museum cultural and creative products designed using generative AI involves multidimensional cross-indicators. As an emerging field, the innovative applications of generative AI in cultural heritage protection are now being explored in the literature and empirical studies, offering new strategies for cultural asset preservation and management. Yet, the data support remains insufficient. The Delphi method integrates perspectives from various fields by leveraging collective expert wisdom, thereby fully activating accumulated professional experience. The established consensus mechanism facilitated broad agreement on effectiveness evaluation standards at both academic and applied levels.

For this study, 22 experts were invited to participate. Their professional fields encompass the entire process of museum cultural and creative design, spanning disciplines like history, archaeology, management, economics, psychology, aesthetics, journalism and communication and computer science. During the invitation process, the researchers adhered to rigorous selection criteria to ensure that each expert possessed profound expertise and extensive professional and/or academic experience within their field; they needed to be holding senior professional titles or serving as principal leaders in relevant institutions. Specifically, each expert was required to have at least five years of professional experience in their respective field, with this expertise demonstrated through either published research or substantial practical engagement in museum cultural and creative projects. The panel composition was designed to ensure a disciplinary balance across the cultural heritage, design, technology and market domains. Details are provided in Table 1.

Before the study commenced, an ethical review and ethical approval were obtained from the academic committee of the principal research institution.

Having compiled the expert list, the first round of research was undertaken. Using semistructured in-depth interviews, 22 experts were interviewed (14 via online video conferencing and eight in person) from February to August 2025. Each interview lasted at least 60 min. The interviews covered five main aspects:

1. Preliminary concepts for the evaluation structure of museum cultural and creative product design, focusing on the integration of cultural elements and modern design principles, as well as the growth and market size of the industry.

2. Existing evaluation methodologies for museum cultural and creative products, their limitations, and suggested improvements, considering intellectual property protection, government policies and the balance between art and utility.

3. The technical application logic, practical experience and prospects of generative AI in museum cultural and creative design.

4. Potential opportunities, issues and challenges related to the use of generative AI to design museum cultural and innovative products.

5. Preliminary concepts and core indicator suggestions for the evaluation structure dimensions of generative AI-designed museum cultural and creative products.

The interview content was transcribed from recordings, which were processed using the three-tier coding framework of grounded theory (Corbin and Strauss, 2015). Two research assistants, both of whom were master’s students, independently completed the coding process. During the open-coding phase, the assistants reviewed the interview transcripts sentence by sentence, identified meaningful conceptual units and grouped similar initial concepts into preliminary categories. During the axial coding phase, the underlying connections between the preliminary categories were further explored and refined into subcategories and main categories. During selective coding, the core categories were distilled from the main categories. The independent research team reviewed and discussed the coding results of both coders, integrating the coding frameworks to ensure theoretical saturation. After the first round of analysis, the hierarchical relationships among the various categories were systematically organized based on the interview data. The position and function of each indicator within the evaluation framework were clarified (see Table 2).

The criterion for theoretical saturation is the inability to generate new categories from data. Upon reanalyzing the interview data, no new categories emerged from the textual materials, confirming that the proposed framework and categories had attained theoretical saturation.

Following the first survey round, four primary indicators, 11 secondary indicators and 39 tertiary indicators had been obtained. Based on this evidence, an initial evaluation framework was constructed. Thirty-nine measurement items were developed to align with the 39 tertiary indicators. These items were compiled by PhD students with specialized training, who have produced research outputs in their respective fields. Anonymous questionnaires were distributed to the experts; based on their professional experience, they were asked to rate each item according to importance, familiarity and judgment criteria. Details are listed in Table 3.

Each questionnaire item included a separate comments section where the experts could provide feedback. Two rounds of expert consultation were conducted, in September and October 2025. In the first round, 22 questionnaires were distributed, all of which were returned, so a 100% valid response rate was achieved. After the first round, the statistical findings were shared with the experts, along with a summary of the revision suggestions, and they were invited to participate in the second round. The second round also involved the distribution of 22 questionnaires and achieved a 100% response rate, with the return of all 22 indicating a remarkably high level of expert engagement.

Data from the two rounds of questionnaire surveys indicated that the expert authority coefficient, Cr, for all 39 indicators exceeded 0.7, demonstrating the high credibility of the expert opinion throughout this consultation process. The Kendall’s W coefficient of concordance was above 0.6, indicating strong agreement among the raters, and the Chi-squared test p-values were below the 0.05 threshold, suggesting statistical significance. Furthermore, the Kendall’s W in the second round exceeded that of the first, confirming the strong consistency within the expert evaluations and greater convergence of opinions.

The effectiveness evaluation indicator system was revised based on feedback obtained from both rounds of expert consultation. The process involved three basic operations: modification, addition and deletion. Changes or additions to an indicator were made only when proposed by three or more experts. For deletions, a threshold method was used to determine whether to remove an item based on three aspects: mean value, frequency of maximum scores and coefficient of variation (CV), as detailed in Table 4. These calculations led to the deletion of Q18 (narrative quality) from A-b-1 (narrative appeal), Q25 (cross-cultural relevance) from B-a-1 (cultural adaptability), Q35 (personalization) from B-b-1 (market value assessment) and Q39 (ideological leadership) from B-b-2 (educational value assessment). All the remaining indicators were retained and refined based on the expert feedback, ultimately forming the effect evaluation scale detailed in Table 5. The scale comprises four primary indicators, 11 secondary indicators and 35 tertiary indicators (specific measurement items).

Purposive sampling was used to select museum A, a national first-class museum in Luoyang, Henan Province, central China. In recent years, museum A has continuously developed its museum IP to integrate with cultural tourism, aiming to create related cultural and creative products. After obtaining informed consent, the researchers used data from the museum’s valuable artifact collection. Data collection was conducted using Doubao, a leading domestic generative AI platform. The research sample originated from the AI Cultural Relic Interpretation Report, and the AI Cultural and Creative Design Proposal (including the design interpretation and concept diagrams) was generated from the imagery and name information about the museum’s valuable artifacts that have been deemed suitable for public exhibition. A total of 516 artifacts were included, and these were categorized into ancient architecture, painted brick carvings, tomb murals, stone carvings, pottery, porcelain, bronzeware and wooden items. All the artifact images are high-resolution originals with excellent clarity, so they met the research requirements. Generated on November 3, 2025, the data set comprised 1,032 AI Artifact Interpretation Reports and AI Cultural and Creative Design Proposals, reflecting the integration of AI technology into cultural innovation. The specific generation process was as follows:

1. Step 1: Open the Doubao platform, URLLink to doubaoLink to the cited article.

2. Step 2: On the new dialogue page, input the following content sequentially.

Prompt 1: Generate an interpretation report on “Artifact Name” based on the artifact image provided. Requirements:

• When describing an artifact, it is essential to incorporate its historical background, cultural significance, artistic characteristics and value, because these elements collectively reveal the artifact’s identity and its role in multiculturalism.

• Use vivid language and clear logic while respecting historical facts and avoiding fictional details.

• To create thorough and detailed writing, adopt the methods of literary criticism by analyzing its distinct styles, recurring themes and linguistic nuances. Write in Chinese, using a minimum of 5,000 words. (Archaeological artifact images are sent as attachments).

Prompt 2: Generate a cultural and creative design proposal for “Artifact Name” based on the artifact image information. This must adhere to the following requirements:

• Integrate the generated interpretive report.

• Design culturally creative products with both social and market value, tailored to contemporary lifestyles.

• The output must include features such as the product name, design concept, functional purpose and target audience.

• It must include accompanying visual concept images; visual concept descriptions alone are unacceptable.

• Be as comprehensive and detailed as possible, and write in Chinese, using a minimum of 5,000 words. (Archaeological artifact images are sent as attachments).

1. Step 3: Wait for the model’s response and automatically receive the results. After generating the AI Cultural Relic Interpretation Report and the AI Cultural and Creative Design Proposal for one artifact, end the conversation and restart with a new one. Continue with Step 2 to generate the Report and Proposal for the next artifact, repeating these steps until all the data have been collected.

2. Step 4: Organize the AI Cultural Relic Interpretation Reports and AI Cultural Creative Design Proposals generated for each artifact into sequential categories to ensure the integrity of the data.

3. Step 5: Manually clean all the data by removing duplicates and invalid content; the data are then archived in PDF format using a standardized template.

The data were processed according to the following procedure: 516 AI Cultural Relic Interpretation Reports were submitted to Evaluation Panel 1 for collective review. The panel consisted of three experts, all holding senior professional titles in cultural relics and museum studies, in accordance with the provincial cultural relics bureau standards. A panel leader was appointed. The other 516 AI Cultural and Creative Design Proposals were collectively reviewed by Review Panel 2, also comprising three experts, one of whom was the panel leader. Two of the experts worked at national first-class museums, held senior professional titles in cultural relics and museum studies, and specialized in managing the cultural and creative projects of museums. The third expert was used by the city’s Culture and Tourism Group, being responsible for marketing the cultural and creative products of the city’s museums. Under the organization of their respective leaders, Evaluation Groups 1 and 2 conducted independent reviews of and collective discussions on the sample data. Each group appointed two research assistants to document the evaluation process. To ensure research rigor, none of the six experts participated in developing the evaluation scales.

During the collective discussion phase, the evaluation panel was required to score all the sample data individually, according to the researcher-defined rubric. Panel 1 was responsible for evaluating the foundational cognitive abilities and deep interpretive capabilities of the generative AI. Meanwhile, Panel 2 was responsible for measuring the value assessment modules concerning the creative output capabilities and cultural innovation of the generative AI. The evaluation variables comprised 11 metrics: factual accuracy interpretation, symbolic accuracy interpretation, hallucinatory degree, information interpretation difficulty, narrative appeal, semantic depth, cultural adaptability, aesthetic consistency, perceptual affinity, market value assessment and educational value assessment. These variables were rated using a five-point Likert scale, with 1 indicating completely disagree and 5 meaning completely agree. The scoring process was conducted independently by each evaluation panel under the guidance of the research assistants to ensure objective and impartial results. The scale validation process is detailed in Figure 2.

The 516-artifact data set was randomly split in half: Sample A (n = 258) for exploratory factor analysis and Sample B (n = 258) for confirmatory factor analysis, thus ensuring independent validation. Data analysis revealed that the Cronbach’s alpha coefficients for all the factors exceeded 0.7, indicating high internal consistency. The exploratory factor analysis yielded a KMO value exceeding the recommended threshold of 0.9, indicating that the data were adequate for factor analysis. Additionally, the p-value was very low, 0.000, well below the conventional alpha level of 0.05. Following factor rotation using the maximum variance method, 11 factors were extracted, explaining 74.129% of the variance. All the item loadings on their respective constructs exceeded the 0.5 threshold, indicating robust construct validity.

In confirmatory factor analysis, the model fit indices were χ²/DF = 1.431, RMSEA = 0.029, GFI = 0.926, CFI = 0.982 and NFI = 0.943, indicating a good model fit. The findings indicated that all 35 measurement items exhibited standardized factor loadings above 0.7 and were statistically significant, thus indicating robust measurement relationships. The AVE values for all 11 factors exceeded 0.5, the CR values surpassed 0.7 and the square roots of AVE. Based on the values exceeding the maximum absolute value of inter-factor correlations, the scale exhibits robust convergent and discriminant validity, as demonstrated in Tables 6 and 7. The scale design was rigorously evaluated for its scientific soundness, ensuring that it is a well-structured, effective measurement tool.

The foundational cognitive abilities that enable generative AI to understand and generate cultural content are underpinned by its capacity to learn from vast data sets and simulate human-like interactions (Epstein et al., 2023). The advanced creative behaviors of AI emerge from its computational understanding of the world, especially within specialized domains. This cognition originates from the extensive analysis that AI performs of multimodal data. This is optimized to develop a preliminary understanding of cultural symbols, artistic styles and other complex patterns in historical contexts. In museum cultural and creative product development, these foundational cognitive abilities are manifest as a precise grasp of the objective facts and visual symbols related to an artifact (Xia et al., 2026). When confronted with an artifact, the primary task of generative AI is to precisely locate and extract factual knowledge about it within its vast parameter space, which includes its name, era, place of discovery, material, function and other fundamental attributes. This process resembles the construction of a structured artifact information repository, whose accuracy directly determines the stability with which the foundation can be used for subsequent interpretation and creation.

Generative AI not only excels in analyzing textual attributes but also possesses strong computer vision capabilities that allow it to examine the visual symbolism of artifacts. This technology can recognize objective properties and interpret visual elements like ornamentation, form characteristics and color composition. With the assistance of AI, these elements in cultural artifacts can be assessed and restored. Additionally, AI technology facilitates the understanding of the cultural connotations of these artifacts, helping to animate them and render their stories more accessible. This level of recognition surpasses simple pixel processing; it relies on deep learning models to grasp the semantic layering of imagery.

Hallucinatory bias refers to the extent to which the outputs of generative AI deviate from factual accuracy by introducing fictional information. This necessitates strict controls during communication optimization. High hallucinatory levels may induce visual illusions or amplify model biases, producing factual misinterpretation and distorted cultural contexts in artifact analysis, which would compromise professional integrity. Consequently, at the foundational cognitive level, a key metric for evaluating the intelligent output of AI is its ability to maintain stable, low hallucination rates while preserving creative vitality. This ensures that the generated content exhibits high accuracy in its fundamental cognitive aspects.

The foundational cognitive capabilities of generative AI are constrained by the breadth and quality of the training data. They are also closely linked to the user prompt content. Should interpreting artifacts pose significant challenges, these foundational cognitive abilities may struggle to capture basic artifact information, thereby compromising the accuracy of subsequent creations. Although generative AI is promising in regard to big data analysis, it often misjudges information and fabricates details when confronted with incomplete data or complex scenarios.

Based on foundational cognition, the next stage of AI-driven logic seeks to surpass superficial descriptions of the physical features and surface symbolism of artifacts. It should explore their underlying contexts and values, requiring generative AI to master probabilistic reasoning and conceptual association. By analyzing the background of an artifact and connecting this to broader historical and cultural contexts, complex, rich semantic networks can be constructed. This integrates isolated artifacts into dynamic historical narratives, revealing their functional evolution and symbolic transformations in specific historical contexts. Via multimodal data correlation analysis, AI can identify the social structures, historical beliefs and aesthetic trends associated with artifacts, thereby reducing the cultural metaphors they convey. This process requires generative AI to not only comprehend “what” but also examine “why,” representing a transition from information integration to meaning construction.

Moreover, using its natural language generation capabilities, AI can form factual historical information into engaging narratives that imbue artifacts with a clear, accessible story. Such narratives present historical facts alongside their emotional and period context, helping audiences sense both continuity and change across civilizations. Built on semantic depth and narrative appeal, this deep interpretive layer moves AI-generated content beyond mere description and toward meaningful interpretation.

Intelligent creation logic and its most direct manifestation of value culminates in creative output capability. This involves integrating the cognition and interpretation developed over the previous two stages, leveraging large-scale model technology to transform cultural elements in innovative ways. This generates cultural and creative concepts that possess practical value and market potential.

The creative output of generative AI is not unrestrained or capricious; instead, it represents innovation firmly rooted in profound historical and factual underpinnings. It necessitates the integration of core cultural elements from artifacts with contemporary viewpoints, showcasing substantial cultural adaptability. While preserving cultural essence, it reconfigures symbols across eras to create derivative designs that harmonize aesthetic uniqueness with practical utility. Throughout this process, generative AI acts as a translator, identifying the essence and adaptable expressions of cultural elements. Its creations must demonstrate cultural adaptability while maintaining aesthetic consistency. Subsequent design solutions must harmonize with the aesthetic qualities of the original artifacts in terms of style, color and line. This avoids jarring stylistic discontinuity while connecting with contemporary preferences. Such holistic stylistic control is crucial if cultural and creative products are to embody both form and spirit.

In the context of cultural and creative development, museums increasingly focus on the perceived appeal of their offerings to meet market demands. This is a critical benchmark for evaluating their success and the quality of AI-generated content. In evaluating artistic and creative products, consideration must be given to not only craftsmanship and design but also emotional resonance, and the artist’s reputation, as well as the condition and scarcity of the artwork. Everyday items such as refrigerator magnets, bookmarks and accessories can be imbued with cultural memory and personal experience to become modern expressions of identity. By analyzing user behavior and cultural preferences, AI technology enhances emotional resonance in design. This transforms cultural artifacts from being mere replicas into elements that seamlessly integrate into contemporary life. Such tangible adaptation is pivotal in ensuring the timeless transmission of traditional culture.

Assessing the value of museum cultural and creative products should not be confined to market-level metrics like sales volume and popularity. Cultural and educational outcomes should also be considered, that is, whether these products raise public interest in the histories embodied in the artifacts and prompt the audience to connect present and past. Ideal cultural and creative development should act as a bridge, guiding audiences to understand the educational value of a museum. By sparking curiosity and imparting knowledge, museum artifacts can be integrated into daily life, ultimately fostering the regeneration and continuity of cultural memory.

The 11 dimensions measured across the four sections formed an evaluation framework for assessing the effectiveness of generative AI in museum cultural and creative design and development. The relationships among the structural dimensions are illustrated in Figure 3. In the diagram, the difficulty of artifact interpretation represents the situational dimension in practical applications, while the others represent the evaluation dimensions.

In this study, a three-round Delphi method and a structured group decision-making technique were used to gather and integrate expert opinions to construct a comprehensive evaluation framework. Widely applied in various fields like product development and policymaking, this method was chosen for its capacity to harness expert knowledge and achieve consensus through iterative rounds of anonymous feedback. Empirically validated, it offers scientific, objective and actionable standards and methodologies for evaluating the efficacy with which generative AI can enhance cultural and creative experiences in design.

The theoretical contribution of this research is to enrich the theoretical framework for generative AI applications in cultural creativity. The document delineates a comprehensive framework for evaluating multimodal content. It encompasses foundational cognition, in-depth interpretation, creative output and value assessment across discrete structural dimensions, thus providing future academic researchers with novel perspectives and directions. The detailed analysis and definition of each indicator clarify the operational mechanisms and influencing factors of generative AI in museum cultural and creative design, thus advancing intelligent communication theories.

In practical terms, this scale offers actionable guidance for museums, technology companies and practitioners. During AI-assisted design processes, these metrics can be used to evaluate the effectiveness of a generative AI application, enabling timely identification and adjustment of issues, thus enhancing the quality and competitiveness of AI-generated outputs. The scale also facilitates the selection of cultural and creative products that better align with market demands and possess cultural value, thereby promoting the effective transformation and dissemination of museum cultural resources. By addressing the human and financial resource constraints facing smaller museums, this evaluation framework might reduce reliance on specialist design teams.

The standardized assessment approach offered by the scale developed in this study could facilitate more systematic evaluations of AI-generated solutions. Nevertheless, further empirical testing across diverse museum contexts is needed to confirm these anticipated benefits. Data-driven feedback mechanisms enabled by this evaluation framework may help organizations better understand public interests, potentially supporting the transition toward more interactive and participatory models of cultural engagement. However, these prospective implications require further empirical validation through multi-site studies, user behavior analyses and longitudinal research.

The limitations of this study must also be acknowledged. Most importantly, the validation was conducted at one museum using a single generative AI platform and a fixed prompting workflow, so the external validity and generalizability are limited. Given the scarcity of empirical research on generative AI and its relatively short practical application period, the evaluation may not encompass all the effectiveness dimensions. Furthermore, as generative AI technology evolves and innovates, the evaluation scale would need to be refined if it is to meet new developmental demands.

Future researchers should continuously monitor trends in generative AI, promptly adjust and optimize the evaluation metrics and explore the differentiated needs of museums that deploy generative AI for cultural and creative design. By integrating regional cultural traits and distinctive collection resources, the evaluation criteria can be further refined. Simultaneously, extensive user behavior data analysis should be integrated to confirm the applicability and robustness of the evaluation criteria across various scenarios, which would enhance their universality and scalability. Interdisciplinary collaboration should emerge as a key direction, integrating perspectives from disparate areas like history, art studies, communication studies and computer science. This convergence would propel evaluation models toward intelligent, dynamic development, supporting museums in their digital transformation, enhancing the effectiveness of cultural dissemination and fostering the flourishing of museum cultural and creative industries. Ultimately, this would enable the charm and value of Chinese culture to be appreciated more widely.