Creativity and Awareness in Co-Creation of Art Using Artificial Intelligence-Based Systems in Heritage Education

Abstract

The article investigates a learning setting contextualising the use of artificial intelligence in heritage education, with a particular focus on AI systems utilising text-to-image processes. The setting is the one of a university interdisciplinary seminar in communication in South Tyrol, a border region in the north of Italy shaped by a strong cultural identity. The paper illustrates a didactic experience introducing a highly technical and, for most of the students in the chosen context, challenging topic, such as AI. The teaching addresses a critical approach to AI, such as dataset constraints, sustainability, and authorship, and focuses on text-to-image algorithms and artistic co-creation, namely, the shifting role of the artist from sole creator to initiator/collaborator shaping the AI system’s output. The aim of the paper is to contribute to the debate in heritage education on teaching and learning using AI-based systems. The latter are seen as a potential tool for the engagement of students in understanding heritage and its safeguarding and in the relationship between community, territory, and active participation, as emphasised by both the “UNESCO Convention on Intangible Cultural Heritage” and the “Council of Europe Framework Convention on the Value of Cultural Heritage for Society”. However, the current boundaries of AI, particularly in terms of bias and limitations of datasets, must be addressed and reflected on.

1. Introduction: The Recent Diffusion of Text-to-Image Artificial Intelligence-Based Tools

This paper aims to contribute to the ongoing discussion in heritage education by exploring the use of AI-based systems in teaching and learning. These technologies are viewed as promising tools for engaging students in understanding and preserving heritage, while also fostering connections between communities, their territories, and active participation. This perspective aligns with the principles outlined in both the UNESCO Convention on Intangible Cultural Heritage [1] and the Council of Europe Framework Convention on the Value of Cultural Heritage for Society [2].

At present, the use of artificial intelligence (AI) is revolutionising several fields, from engineering and communication to visual arts and education. Scientists have long worked with artificial intelligence, but for most people, AI was, until recently, an abstract concept. Although AI systems like smartphone voice assistants and search engine text predictions have been part of daily life for years, many were unaware of their presence. This changed with advancements in AI, especially in image generation, making the technology widely accessible. Referring to these tools, AI gained a lot of attention over the last decade but even more over the course of the last 3 years especially, starting from 2022. During this period, the release of test versions for several applications, including Stable Diffusion [3], DALL-E 2 [4,5], and Midjourney [6], occurred. Additionally, the autumn of that year marked the introduction of the now widely known and extensively utilised text-based AI assistant, ChatGPT-4 [7,8]. It was a time of considerable activity and constant talk surrounding AI, as evidenced by the large number of articles and papers that, by the spring of 2023, were written about it, many of which were critical in tone and feared the displacement of humans by an “AI invasion” [7].

The spread of AI-generated images raised awareness of AI’s rapid development and ease of use, even for amateurs. However, it also sparked debates about ethics, authorship, and increased concerns over fake news, particularly on social media.

The present article discusses the open issues around AI just highlighted, with a particular focus on text-to-image AI as artistic co-creation and its application within heritage education. AI-mediated artistic co-creation represents a fertile ground for innovation, offering new tools to explore cultural and natural heritages in innovative ways.

However, given the current limitations of datasets and the risk of creating generic, inaccurate, or even inappropriate images—that is, not adhering to elements of real heritage—it is necessary to take a critical approach to AI. With a touch of irony—within a university seminar on image production in the context of the promotion of local heritage—we posed the question “How can we express the idea that AI will always provide an answer, even if it is not correct?” to ChatGPT, which replied, “Artificial intelligence generally produces a result, even without accurate data, by processing the query as close as possible to the models it was trained on”. This prompt and response triggered a first-class discussion on AI with the students.

The paper presents a critical reflection on the use of AI, with particular reference to imaging tools, and reviews the state of research on its application in heritage education, contributing to existing discussions. It then provides a theoretical framework for the algorithms that underpin the tools used in the experience of the seminar, which is reported here as a case study. As AI-co-created images are part of broader reflections on artistic production, critical aspects of AI-generated art are presented.

Finally, a case study concerning the seminar just mentioned is illustrated. The teaching steps are described, followed by a qualitative analysis of two examples and related group discussions with students regarding the experience.

The aim of the research is to offer a critical analysis of artistic co-construction with the use of AI and to incorporate this reflection into the debate on the use of AI-produced images for heritage education purposes.

The research questions are as follows: How can awareness be raised about the use of AI tools for image production? Which aspects should be taken into account for the teaching practice in universities? What are the current limitations of these tools in the specific area of heritage education?

1.1. A Critical Approach to Text-to-Image AI: An Open Debate

The growing popularity of AI-generated art has raised significant debates regarding authorship, creativity, and copyright [7,8,9,10,11,12,13,14,15,16]. While non-AI-generated art is valued for the artist’s personal expression, AI-created works are often perceived as lacking emotional authenticity because they are produced by algorithms rather than human hands. This perception can affect the audience’s emotional connection to the work, leading to a more objective evaluation focused on the technical aspects of creation.

Public perceptions of AI-generated art are influenced by biases that can devalorise such works compared to those created by humans. A study has shown that people tend to evaluate art labelled as created by AI less positively, even when they cannot distinguish it from human art [9]. For this reason, previous studies have been concerned with developing benchmarks to evaluate how text-to-image models represent different cultures, improving the generation of culturally accurate images [10], and to assess cultural awareness and diversity in text-to-image models, highlighting gaps in the representation of global cultures [11].

Ethical issues also arise regarding the use of existing data for training AI models. Often, these systems rely on vast datasets that include copyrighted works, raising concerns about intellectual property and the consent of the original artists [12,13,14].

Concerning co-creation, it is necessary to consider whether only the end product is evaluated, i.e., the image resulting from the collaboration between humans and the AI program, or whether the entire process behind it is also included in the assessment. Just as a non-AI-generated work of art involves a lot of work and preparation, so too can AI-generated works of art. The phrase “man versus machine” is often used in discussions about AI, robots, and technological progress. However, in the context of artistic co-creation between humans and AI, the relationship becomes collaborative, with technology integrated into the artistic process. In the scientific community, there is a still open debate on exploring whether AI programs are merely tools or if they possess autonomy in decision making [7,17,18].

Starting from the emphasis on the indispensable role of human agency in the creation of art, particularly in the field of AI-generated art, it is appropriate to examine the importance of datasets in the fabrication of AI systems used in artistic endeavours because the system is fed with data during the training phase. Once the system has been trained to generate images based on a specific dataset, it internalises the patterns and characteristics of this dataset and takes them into account when generating new images. This allows the system to automatically generate new data starting from an external input of the user, but the result is still strongly affected by the characteristics of the dataset used. It is, therefore, not surprising that critical voices arise and ask whether the results can be described as art if they are inspired by and dependent on the used dataset. After experimenting with these tools, it was noted that generating creative and naturalistic images still remains one of the most difficult challenges for text-to-image models [19]. Furthermore, the effectiveness of the results depends on the quality and accuracy of the input text, with more detailed and context-rich inputs producing better results in terms of quality and accuracy than short and imprecise inputs. This emphasises the complex inter-relation between the quality of the input data and the resulting outcome and also highlights the underlying dynamics of text-to-image generation processes. The selected dataset plays a decisive role in the design of the results of the generated images. Optimal results in relation to consistency with the chosen prompt often depend on the use of large datasets that require a large storage capacity. Using large datasets not only requires adequate storage resources but also usually leads to better results.

Finally, regarding the issue of data and sustainability, several studies highlight the environmental impact of AI itself, for example, in terms of carbon emissions and energy consumption. The research goal is to develop AI that is compatible with environmental, economic, and social sustainability, balancing innovation, intergenerational justice, and resource equity [20]; in addition, tools to optimise energy consumption, the importance of regulations, and future prospects for more sustainable and accessible AI are analysed [21,22].

1.2. Context and Aim of the Research

The context of the research is the one of higher education, in a seminar on visual communication, with a focus on heritage education for students of a Bachelor in Communication Sciences and Culture program. It illustrates the steps needed to introduce students to AI, the activities proposed, and the reflection that emerged at the end of the seminar. Among the different AI-based solutions, the analysis in this paper is focused on the experiences of students related to the use of visual AI tools for the production of images starting from a textual prompt.

The aim of this work is to illustrate a practical experience in which AI is in dialogue with heritage education. At the same time, it illustrates, in a simple and non-technical way, aspects underlying AI, enabling teachers and students to gain skills in using AI tools and to develop critical thinking on the subject rather than just being passive users.

Within the seminar, the technical part concerning AI was explained in an experiential way: the students were first introduced to how a particular algorithm or piece of software worked and, later, how to use it, gaining direct experience. Such experiences made it possible to reflect on the potential, but above all, on the limits, of AI tools, as they raise a number of open questions. The reflection was focused on the following:

-

The general perception, what is written in the newspapers, what is on the social networks, and what is also in the videos on AI, such as the Ted Talks: for example, the videos found on the Internet often use dark colours, distressing background music, and help to convey the idea that artificial intelligence is stealing jobs;

-

The need to adopt a critical and not aprioristically negative perspective: this is necessary in order to understand that artificial intelligence is still far from putting into practice the concerns that lead to a totally negative view and that it can also be useful to create new ones, despite the open debate on the issue concerned different aspects of the use of AI;

-

AI’s limitations, in particular the fact that AI is not only hardware-based but also software-based, i.e., data-based;

-

The recognition that data are the most important thing and that this is what creates the biggest problems, because these data are selected or otherwise entered by people who are really a tiny percentage of humanity; this means that there are currently many limitations and biases in the datasets.

A major consideration is that it is not possible to know exactly how the algorithms work, but we can only approximately know what they are based on, since, at the end of the day, the jury of artificial intelligence is basically all based on the same algorithms. This is why it is necessary to have an introduction to these algorithms for students. They represent digital competences that can also be useful, for example, when creating applications to better understand how they work.

2. Theoretical Framework

A reflection on AI in the context of visual communication, in a university seminar with a strong vocation for heritage education, required an interdisciplinary impetus, bringing together themes, languages, and perspectives from fields that may be very distant from each other. It is a contaminating and generative exploration, from engineering to education, through visual communication and reflection on the artistic process. Nevertheless, the theoretical framework is presented in two sections that reflect the fields of research to which the two authors belong and, thus, their starting point.

2.1. Generative AI Platforms and Text-to-Image Models in Art Applications

In order to stimulate awareness and a critical approach to text-to-image AI-based tools, it is essential to give an overview of the algorithms underlying the tools that will be used in this research, especially those behind the text-to-image models. First of all, it is important to highlight that they belong to the branch of Machine Learning that is a technique that enhances system performance by allowing computers to learn from data and experience. Rather than being explicitly programmed for every task, the system is provided with large amounts of data, from which it can recognise patterns and develop models. These models enable the system to make predictions or decisions when presented with new inputs. Essentially, the system refines its ability to perform tasks based on past data, improving its accuracy and efficiency over time through repeated exposure to various datasets [8]. Among Machine Learning algorithms, the text-to-image models are based on generative methods. This means that it is necessary to give to the system one type of input so that the system can generate another type of output. As with text-to-image models, the user gives the system written input in the form of a text, and from this input, the system then works out an output in a sequence of certain processes, which can then correspond more or less to the ideas of the user.

The appearance of generative adversarial networks, so-called GANs [23], accelerated the use of AI in the process of creating visual art. It was highlighted that the crucial role of GANs has been responsible for a large number of recent AI artworks [24]. The method of generative adversarial networks is based on two components. These two components are so-called neural networks, one of which is referred to as a generator and the other as a discriminator, which compete with each other [25]. The fact that there are two components and that they compete against each other explains why these networks are labelled with the adjective adversarial. In concrete terms, this looks as follows: the generator observes the distribution of the data, while the other component, the discriminator, assesses the probability that its input originates from authentic data or is false data that was created by the generator. The two components are, therefore, in a kind of dynamic interaction between each other, and this leads to improvement of the results with the aim that, in the end, the discriminator is unable to differentiate between the real data and the fabricated data produced by the generator. In simple terms, the process involves two components: one generates random images based on the data it has learned, while the other evaluates these images by comparing them to the input it has been trained on. Through this interaction, the system gradually improves, with the generator learning to model the data more accurately and create new, realistic images. After training, the generator is able to produce novel samples that align closely with the patterns it has been taught.

Diffusion models [3] represent another method. Some of the most popular and most used models to create AI art are diffusion models, such as Midjourney and DALL-E 2. The operating procedure of diffusion models differs from that of generative adversarial networks. Among other things, the training phase is what distinguishes the two. Basically, for the diffusion model, there are two main things that occur in that phase, which are called the forward process and reverse process [25]. The model works with an input image to which, in the first part, the forward process or also the diffusion process, noise is added [21]. This noise is continuously added to the input image. This is continued until the input image is no longer recognisable and consists only of noise. Following the forward process, the second step—the reverse process—takes place. During this step, the model uses the learned insights from the forward process and is able to reconstruct the original input through the removal of the previous added noise. This second process of denoising diffusion probabilistic models refers to the fact that in the reverse process, the model estimates the probability of what the density of the image looked like at an earlier stage.

Software such as Midjourney or DALL-E are the most common between users in art application for their easy use and proper results. They allow users to create images by simply giving a written instruction. These instructions, which are used by the system to generate the result, are called prompts, and the activity of writing these requests is known as prompting. When composing a prompt, the user has the opportunity to direct the system according to their wishes and can give free rein to their imagination. This includes not only the description of the desired image content or the colour scheme but also makes it possible to demand a specific artistic style. While articulating the prompt, the user can have a decisive influence on the result, especially on the creative dimension. However, to gain a proper result, the user needs to prompt effectively and efficiently. The process of elaborating such a prompt is called “prompt engineering”. Instead of just writing normal sentences, it is way better to give precise instructions, focusing on meaningful keywords and a good structure [15,19]. After inserting the prompt, the algorithm then creates a result in no time at all and generates not just one image but four different versions, which sometimes differ more or less in style or design. This highlights the versatility of the AI system because based on a single prompt, the user obtains four different output results. Subsequently, the text-to-image user then has the practical option of further editing the images in order to better adapt the output with the desired result. It is possible, among other things, to select one of the four images and have a new version created from it. Since the text-to-image method is so simple—with simple meaning that anyone can do it without special qualifications—it is a very popular tool for creating realistic images. For this reason, it is only reasonable that this and other models have the potential to change a lot and have an impact on different fields of application and also on the art world.

2.2. AI Tools for Heritage Education: A Field Still to Be Explored

The field of heritage education is interdisciplinary. It crosses the line between history and archaeology, art history, and education. It addresses both formal and non-formal educational contexts. More recently, it also involves specific studies concerning the relationship between heritage digitisation and education [25]. The use of virtual and augmented reality dominates, at present, the research on heritage education, including often elements of gamification to better engage the user [26,27]. However, this area of digitally oriented research has not yet addressed the application of AI in heritage education, while this is already happening in heritage conservation, especially regarding material cultural heritage [28,29,30].

In general, we need to recognise that heritage education studies often remain fragmentary and do not always encompass all aspects of heritage: natural heritage and tangible and intangible cultural heritage [31]. When all three aspects are encompassed, heritage education offers actions, initiatives, and projects to enhance both citizenship education and Education for Sustainable Development, enhancing participation and active citizenship [32,33].

In particular, a perspective that stresses the centrality of intangible cultural heritage (ICH)—as promoted by the 2003 UNESCO Convention—bridges together the other two, overcoming the nature/culture separation and intercepting the three pillars of sustainability—ecological, social, and economic—thus responding well to the goals of Agenda 2030 [32]. In a real “change of paradigm”, ICH envisages a bottom–up approach of research and transmission of knowledge and practices that invites participation and active citizenship [34]. By extending the possibility of participation of active and committed children in the safeguarding of intangible heritage, for example, the community and the territory are explored with them, encountering, questioning, and giving meaning to the other elements of heritage, too. In line with Childhood Studies [35] and a World-Centred Education [36], children are, thus, recognised as cultural subjects and actors.

3. Research Methodology

The research starts from the need to fill a knowledge gap with respect to artificial intelligence tools and their use. Using AI tools without being aware of how they work allows for merely passive use. The effort required to understand a field as complex and technical as AI is, therefore, necessary and falls within the competence of scientific citizenship [37]. The research takes a qualitative approach and presents a case study, with a mixed method [38], in order to gain insights into the students’ learning experience. Sixteen individual papers, which included the description of the image co-creation process, as well as the prompts used and the final part with a self-evaluation reflection, were coded and analysed. These data were then compared with the recordings of conversations in two focus groups at the middle and end of the seminar.

The students’ work on local heritage campaigns on specific topics of their choice was analysed. The reflections that emerged in a group discussion at the end of the seminar were transcribed and analysed, bringing out common categories, recurring themes, and significant observations. The teacher also conducted non-systematic observations, noted on observation sheets that described the seminar steps, of the experiential moments in class on the use of AI software. These observations were useful for the analysis and interpretation of the data collected. The students were informed of the research, and the data were anonymised. The seminar consists of two parts, a theoretical one and an experiential one. These are illustrated below.

3.1. Reflecting the Creation Process in Non-AI- and AI-Generated Art

The first part of the seminar proposed a comparative analysis of the creation processes of non-AI-generated art and AI-generated art. The aim of such an analysis is to examine and contrast the processes involved in non-AI-generated art creation with those employed in AI-generated artworks. It also allowed us to explore the differences and similarities between these two approaches and to shed light on the work behind AI-generated images and artworks. The analysis addressed key questions posed in this article, such as the role of human artists, creativity, and authorship in art production with AI. While human skills, emotions, and interpretations play a role in art production, AI art introduces a new dimension by using computer algorithms to autonomously create images and artworks. This raised the question of the role of human agency, creativity, and authorship in art production. By comparing these two processes, the aim is to gain a deeper understanding of how AI complements or challenges non-AI-generated artistic practices. Furthermore, understanding the differences and similarities between non-AI- and AI-generated art making is essential to understanding the developing world of artistic expression. The results of this more general analysis were then used for experimentation and critical insights in the more specific field of art and heritage education, as highlighted in the next section of this article. In the subsequent sections, specific aspects of the artistic creation process are described, such as conceptualisation, execution, and iteration, and various aspects like creativity, authorship, and audience perception are examined, as

Table 1. Aspects of artistic creation process: a comparison between non-AI- and AI-generated art.

	Creation Process of Non-AI-Generated Art	Creation Process of AI-Generated Art
Conceptualisation	Artist’s inspiration and vision	Artist’s inspiration and vision
Training and learning	Self-taught, apprenticeship, courses, and schools	Training with datasets
Data acquisition	Personal experiences, historical or current events, etc.	Datasets containing text and/or images
Execution	Manual techniques and craftsmanship	Algorithmic generation
Iteration and refinement	Colour changes, etc.	Tinker with prompts
Technical constraints	Accessibility of material, financial resources, and correct preservation	Financial resources and technological difficulties
Creativity	Personal expression and skill, even in case of collaborative act	Collaborative act; dependent on dataset; artist has influence in pre- and post-curatorial actions
Authorship/ ethical considerations	Clear authorship and originality/possible theft issues	Shared authorship/concerns: intellectual property and copyright
Audience perception	Emotional connection and artist’s story	Focus on technical appreciation; perceived lack of authenticity and possible existence of negative bias

below illustrates. The aspects are analysed in detail in Chapter 4. Findings and Analysis.

3.2. Students Experiences with AI Tools Applied to Heritage Education

In the second part of the seminar, the students worked on experiential activities focused on the valorisation of heritage.

The latter was presented in its totality as natural, tangible, and intangible heritages. The reference framework is both UNESCO’s and the European Union’s [39,40,41,42].

The reflection involved the students starting from their personal perception of heritage, understood as a set of affective and valuable relationships, tangible and intangible, that they recognised as significant, in their daily lives, in the community, and in the territory in which they live. The profound relationship between community and territory, highlighted by the debate on intangible cultural heritage that led to the drafting of the UNESCO Convention for the Safeguarding of the Intangible Heritage in 2003, was the starting point for the preparation of the students’ papers. In fact, they reflected on how the 2003 Convention made it possible to go beyond the concept of cultural heritage to be preserved and conserved, introducing instead the concept of heritage to be safeguarded: no longer just handed down from generation to generation, but charged with a value starting from the perception that individuals give it, from their narrative representations. In this perspective, all heritages, even natural and material ones, are immaterial, and there is a need for awareness on the part of subjects with regard to their being cultural actors/actresses. In other words, even heritage valorisation actions through a visual communication campaign, such as those linked to tourist promotion agencies or companies in an area, can convey messages of care and the protection of heritage.

In the case of the seminar, the territorial context is the one of South Tyrol. This is an Alpine region in the north of Italy, on the border with Austria and Switzerland, characterised by a natural heritage recognised by UNESCO: the Dolomites.

The region also boasts an architecture that differs from the rest of Italy in that it is the expression of a local history that saw the region pass from Austria (the Tyrol) to Italy only after the end of the First World War. Elements of intangible heritage, too, are linked to this past, although it is precisely because it is a living heritage that is specific to the border regions. This is true of cuisine, which often reflects a unique Alpine culinary heritage.

The students were asked to create a narrative based on a heritage element or a regional good or a brand promoting the region. The exercise comprised the following tools:

• AI-generated image tools, such as automatic image generators based on deep learning models, to create visual representations of cultural heritage elements, regional products, or promotional concepts: these tools allowed the students to experiment with different styles, from naturalistic renderings to abstract reinterpretations;
• AI-powered text generators to assist in crafting compelling narratives, descriptions, and promotional texts: these tools provided students with creative support in structuring their storytelling while also raising questions about authorship, originality, and AI’s role in content creation;
• A virtual tour platform with augmented reality, enabling the realisation of immersive experiences where users could explore digitised heritage sites, interact with regional products in a virtual space, or navigate brand storytelling through interactive elements: the augmented reality application allowed the students to visualise cultural and territorial elements in an engaging way, enhancing digital storytelling techniques.

The works produced were then reflected upon in a group discussion plenary. The final reflection focused on the qualities and limitations of the tools used, fostering a critical approach to AI. The article presents only the analysis of the seminar activity concerning the production of images. The students’ works were very diverse and ranged from historical reconstructions of landmarks to fictionalised reinterpretations of cultural symbols, showcasing different approaches to representing cultural heritage through AI-generated content.

The heritage perspective, understood in the inter-relationship between community and territory and, thus, including both aspects of natural heritage and tangible and intangible cultural heritage, has a strong vocation to convey the theme of sustainability. The latter, however, becomes problematic when AI tools are employed. A critical approach to artificial intelligence compels the recognition that the use of AI could be problematic within a sustainability perspective due to energy consumption concerns. This aspect was dealt with in the seminar.

4. Findings and Analysis

This section first analyses two students’ works, describing the process of co-creation and the potential of AI-generated images for the promotion of heritage. The teacher’s request was that the students develop a visual and text narrative based on a cultural heritage element, a regional product, or a representative brand of the area. To do this, students had to use artificial intelligence tools to generate images. The goal was to explore the potential of these technologies for heritage promotion while analysing their creative possibilities and critical limitations.

The second part recalls the comparison between art created with and without AI. The analysis is drawn upon the students’ reflections, which emerged during the group discussions.

4.1. Same Exercise, Two Differing Outcomes

While the first example meets the challenge of combining an effective image from a communication point of view with the promotion of one or more elements of the local heritage (natural, cultural, tangible, and/or intangible), the second example is highly problematic due to the current limitations of AI, which does not always detect and consequently generate images pertaining to a local area.

In the first example, the student imagined a brand name for a new elderflower- and honey-flavoured drink. After trying out different names, the choice fell on “Monta”, which is close to the Italian word “montagna” and the English word “mountain”. The image processing steps using Firefly software are summarised in four steps (Figure 1).

In the discussion phase, the student reports how the various steps required to arrive at a satisfactory result confronted him, above all, with the need to compensate real natural elements with AI-generated images that have a strong visual impact and that also risk being somewhat generalistic. The result obtained shows one of the main chains of the Dolomites, the Geisler/Odle chain, in the vicinity of which both the “Geisler/Odle Nature Park Visitor Center” (Puez) and the “Infopoint Dolomites UNESCO World Heritage” (Zans/Zannes) are located. By this means, the new brand promotion for the drink is linked to the valorisation of the region, and this combination can be further conveyed, for example, with promotional material that makes the heritage topic explicit.

In this case, the AI tool recognised the prompt request referring to the mountain chain of the Dolomites. In the following example, however, the prompt concerning the generation of images related to the village of Egna/Neumarkt was not recognised. The students worked on 12 different prompts and composed the postcard shown in Figure 2.

The need to reach the widest possible audience justified the rich and varied composition of the images. The various images are intended to highlight different aspects of the landscape, history, wine, and food, as well as shopping, leisure, and cultural and sporting events.

The product generated by the AI system is appealing and communicatively effective. However, none of the images represent the village of Egna, and the student does not recognise this, relying on AI. In the case of image 1 in Figure 2, for example, the prompt was as follows:

1—A lively festival scene in Egna, Italy. A group of joyous people dressed in traditional Sud Tirol costumes is captured dancing energetically on a sunny day. The women wear colorful dresses with intricate patterns, while the men wear coordinating vests and shirts. The cobblestone street and rustic buildings in the background add to the festive, cultural ambiance as onlookers gather to enjoy the celebration.

Yet the resulting image represented neither a folk dance group from Egna nor from South Tyrol: the shape of the rooftops of the buildings and the dresses of the dancers do not look like local buildings and dresses. The student should have noticed this at first glance or should, in order to be sure, have taken a closer look at the picture or have checked it out on the Internet. The prompt for image 5 in Figure 2 also explicitly mentions the village of Egna:

5—A picturesque view of Egna, Italy, captured on a sunny summer day. The photo showcased the charming medieval town with its terracotta rooftops and historic buildings nestled on a hill. In the foreground, lush vineyards stretch out towards the town, leading the eye to the prominent church at the hilltop. The rolling hills and expansive landscape in the background add to the serene and timeless beauty of this idyllic Italian setting.

Again, the village in the AI-generated image was not Egna. This can be deduced from the brick buildings, the presence of two bell towers, and the hills in the background.

Finally, the other prompts are general and aimed at composing a memorable description, with expressions such as “Imagine a quiet summer morning in a picturesque vineyard, nestled in the rolling hills of the countryside...” (image 2, Figure 2), “Aerial view of a village with a church with two spires...” (image 3, Figure 2), and “A lively street scene with open-air cafés...”.

Discussions highlighted the need to avoid stereotypical or clichéd representations of local heritage, a risk that can arise when, for example, elements of intangible heritage, such as food, customs, or rituals, are overemphasised. Thus, it was acknowledged that the choice of having a variety of prompts to describe an area rich in natural and material heritages, with a view to the intangible (dance and music, above left) as “living heritage”, was the right one. Another positive aspect is the relationship between the past, the present, and the future, which is well represented by the two close-ups (bottom right) of a young girl and an elderly man.

However, the great constraint imposed by the limited amount of data and, thus, the small chance of the AI system being able to recognise the request to create a heritage picture related to a small village, as well as the fact that students generally admit to trusting AI tools, is the most problematic aspect that emerged during the discussion.

This raises concerns about the reliability and accuracy of AI in representing diverse and localised cultural contexts, highlighting the limitations of the datasets on which AI tools are trained. While AI systems can generate impressive results, they often rely on generalised datasets that may not fully capture the nuances of specific cultural or regional identities, especially in smaller, less-documented communities. This created the risk of perpetuating oversimplified or even inaccurate representations, as the AI tool might not have access to sufficient data to understand or correctly interpret the unique aspects of a local heritage.

The need for more tailored datasets is crucial to ensure that AI can generate representations that are culturally sensitive, contextually appropriate, and reflective of the diversity within different heritage groups. These datasets would need to include more specific examples of local traditions, history, and cultural practices and be designed to avoid reinforcing stereotypes or presenting an overly uniform view of natural and cultural heritages. It would also be important to involve cultural experts, community members, and local historians in the creation of these datasets, ensuring a more accurate and respectful representation of heritage.

Additionally, the critical engagement with AI technology is vital. Users, such as the students in this experiment, should be encouraged to not only trust the tool but also critically evaluate the images generated by it. This process involves questioning whether the AI tool’s interpretations are authentic and representative and understanding how the input prompts, datasets, and algorithmic biases shape the final output. By developing a more nuanced understanding of how AI works and its potential limitations, students can become more responsible users, capable of identifying and addressing the biases inherent in AI-generated art. This critical approach ensures that AI is used thoughtfully and in a way that respects the diversity and richness of cultural heritage.

4.2. Balancing Creativity and Awareness in Co-Creation of Art

The analysis highlights the evolving role of artists in AI-generated art, shifting from sole creators to collaborators who guide AI through prompts and refinements. While traditional art relies on intuition and manual skills, AI art depends on datasets and algorithms, raising questions about authorship, ownership, and ethical considerations.

Despite differences, both AI and non-AI art require creative input, technical problem-solving, and iterative refinement. Public perception of AI-generated art remains mixed, often seen as lacking emotional depth, although this may change as AI tools become more integrated into creative fields. Rather than replacing human artists, AI introduces new forms of artistic collaboration, expanding creative possibilities.

A key takeaway is that AI-generated art should be valued not just for its final product but also for its creation process, similar to traditional art. This approach acknowledges the role of human decision making in AI-assisted creativity.

This study was conducted in the context of a seminar on heritage communication and education. The first part of the seminar was dedicated to a detailed understanding of AI algorithms and the critical aspects of the co-creation of text-to-image products. Students then first explored text-to-image tools and later tested their own creativity in composing visual descriptions of natural and cultural (tangible and intangible) heritage elements of their choice.

• Conceptualisation

The methodology followed a structured approach, beginning with an introduction to AI-driven image generation, emphasising the role of datasets and the importance of input prompts in shaping visual outputs. The students explored how different descriptive approaches influenced image generation before selecting a visual storytelling theme.

Choosing the right AI tool was a crucial step, as different software offered distinct capabilities. DeepDream produces pattern-based images, DALL-E specialises in stylised artistic outputs, Midjourney generates striking visuals but is costly, and Adobe Firefly allows for targeted image refinement.

Artistic creation traditionally evolves through iterative refinements based on personal experiences and cultural context. AI-generated art follows a similar process but incorporates technical steps such as dataset selection, system training, and software evaluation. In the case of image generators, algorithms that have already been trained on standard datasets are used, but each time a specific prompt is used, image refinement is carried out, and the desired image is obtained after a few attempts, this constitutes an iteration of the training process. While both methods involve decision making and iteration, AI art requires computational resources and refinement, highlighting the intersection of human creativity and machine-driven processes.

• Execution

The iteration and refinement phase was time consuming, as the students worked to create textual descriptions reflecting their cultural or landscape heritage. These descriptions were used in AI text-to-image applications to generate visual representations, prompting the students to evaluate how well the AI outputs aligned with their original vision. This process fostered discussions on the interpretive nature of both human- and machine-generated content, highlighting AI’s strengths and limitations in heritage representation.

Technical constraints also emerged, with high-end AI tools requiring expertise and financial investment, while more accessible programs offered free trials or subscription-based models. This underscored the varying accessibility of AI for different users.

The creative process in AI-generated and traditional art follows similar iterative patterns but with key differences. Traditional artists refine their work through hands-on adjustments, while AI-generated art relies on refining prompts to guide outputs. Both require evaluation and revision, with human oversight playing a crucial role in shaping the final result. Constraints—whether material limitations in traditional art or computational and financial barriers in AI—affect both processes, influencing artistic outcomes and innovation.

• Creativity

The experiment demonstrated that AI significantly impacted the creative process by expanding students’ visual interpretations of their cultural and landscape heritage. While AI-generated images introduced unexpected and innovative perspectives, the students maintained creative control by carefully curating their prompts and selecting outputs.

The collaboration between human creativity and AI acted as a catalyst for artistic exploration, allowing the students to rethink their conceptualisation and discover new aesthetic possibilities. The findings highlighted that AI tools do not replace human imagination but rather extend it, reinforcing the idea that creativity is a dynamic process shaped by both human intent and technological influence. This aligns with broader debates on authorship and artistic agency, where AI challenges traditional notions of creativity while offering new opportunities for artistic expression.

• Authorship and ethical considerations

The project raised several ethical concerns, particularly regarding data usage, originality, and bias. AI-generated images were created using vast datasets including copyrighted materials, making it difficult to determine originality and proper attribution. Cultural misrepresentation was another issue, although mitigated by students working with their own heritage. The experiment highlighted the importance of transparency in AI systems and the potential for embedded biases.

Authorship emerged as a key debate, as AI-generated art complicates traditional notions of creativity. While students saw themselves as the primary authors due to their input in crafting prompts, the AI system played a significant role in shaping the final output. This suggests that AI-driven creativity is best understood as a shared process between human direction and machine execution, considering that the artists whose work was included in the training database, the programmers who wrote the code, and the workers who contributed metadata could also be considered participants in the creation of the images. The project underscored the need for ethical guidelines in AI’s application to heritage studies, fostering critical awareness of its impact on artistic creation.

• User perceptions

The experiment highlighted how AI-generated art is perceived differently from traditional, human-created artworks. While the students appreciated the novelty and technical sophistication of AI-generated images, they noted that these works often lacked the emotional depth and authenticity associated with human expression. Audience perception tended to focus more on the technological aspect rather than artistic intent, raising questions about the legitimacy of AI art. However, students acknowledged that as AI tools become more integrated into creative fields, these perceptions may shift. The project reflected broader debates in the art world about originality and artistic value, demonstrating that AI is reshaping creative expression and challenging traditional notions of art.

5. Discussion and Conclusions

This paper discussed some aspects of the use of AI tools in heritage education. Specifically, it referred to text-to-image tools. These are seen as potentially attractive, as they allow ample scope for creativity and experimentation in order to stimulate students’ engagement in the exploration and valorisation of heritage. Students could be personally involved in producing products. Yet such products—in our case, in the form of an image, but in the case of text-to-text AI tools, in the form of a text—are only partially created by the student. It was, therefore, necessary to explicitly address and reflect on the issue of awareness of co-creation when using AI tools. The limitations as well as the advantages need to be highlighted.

The fact that these AI tools are not open source—i.e., we do not know how they are trained or which algorithms they use—limited the results produced. It is possible to guess how they work from the available open models, but a significant limitation remains. It may be more or less adherent or correct with respect to the given prompt. In the specific case of image searches for local heritage elements, it is still unlikely that these will be recognised. Exceptions are heritage sites officially recognised by UNESCO, like the Dolomites in our example.

The potential of using AI text tools lies in the fact that they are accessible and inclusive, as images can be obtained automatically, overcoming the complexity of using certain digital programs (e.g., Photoshop) or allowing a user with a given disability to be involved in the activity, although some questions still remain open [43].

It is, therefore, only by being aware of these AI tools’ limits, possible to use them critically, considering areas where their use is not worthwhile and others where, for example, the ability to train the writing prompt can achieve increasingly better results.

Currently, one particular limitation is the representation of the intangible cultural heritage of images created using artificial intelligence. In fact, due to the extreme cultural diversity that is the expression of this heritage, there is a risk of the depiction of traditional aspects that do not correspond to reality. Even if expressions of intangible heritage, such as dance, may be similar, they are always linked to a community in a local area. They are, therefore, unique in their specificity.

In the two examples illustrated here, we found the following findings:

• A critical approach to AI—how it works, how datasets are constructed, and what the biases and current limitations are—is a crucial reflexive step to start experimenting in the classroom and to stimulate critical reflection in students.
• Yet such an initial step is not sufficient to avoid—as in the examples of text-to-image exercises—the user trust effect and the production of images that do not correspond to reality, e.g., as required in a heritage education context.
• Students need to be reminded to compare their product with photos and the literature of the chosen heritage elements in order to avoid the production of unintentional fake images.

The findings support the need to conduct further research towards the potential use of AI not only in the context of heritage digitisation [42] but also in heritage education.

From a technical standpoint, future research must focus on developing appropriate datasets for the study and promotion of cultural heritage. Unlike text, which can be expanded and refined relatively easily through continuous training, images require a fundamentally different approach. In text-based AI models, datasets can be augmented by generating variations of existing data, but for image-based AI, high-quality, well-contextualised visual datasets must be manually curated and structured. This presents two key challenges: first, someone must create or compile vast datasets of heritage-related images, ensuring that they are representative and diverse; second, training AI models on these datasets requires an enormous number of images, as well as metadata-rich contextual information to enable meaningful associations.

In the specific case of cultural heritage, copyright issues add another layer of complexity. Unlike general-purpose image datasets, heritage-related visual materials are often subject to ownership restrictions, making it difficult to access and integrate them into AI training processes. Additionally, the need for comprehensive and legally compliant datasets raises questions about data licensing and fair use, particularly when AI-generated images are used for public or commercial purposes. Moreover, the creation of heritage-focused AI models cannot rely solely on developers and engineers. It is crucial that dataset preparation is carried out in collaboration with cultural heritage experts, historians, and archivists. This interdisciplinary approach ensures that AI systems correctly interpret and categorise heritage materials, avoiding misrepresentations or the reinforcement of historical inaccuracies. Establishing an ontological framework—defining the relationships between objects, symbols, and historical contexts—is essential for AI to generate meaningful heritage-related images. Without this level of structured knowledge integration, AI-generated visuals risk becoming superficial or misleading representations of cultural history.

Ultimately, these challenges highlight that AI’s role in heritage studies and heritage education is still in its early stages. While the potential for AI in this field is evident, the current limitations are significant and multifaceted. Addressing them requires not only further research but also increased experimentation, testing, and technological innovation. AI systems need to be developed with greater cultural sensitivity, improved contextual understanding, and better legal frameworks to support their ethical and effective application in the heritage sector.