Using artificial intelligence in craft education: crafting with text-to-image generative models

ABSTRACT Artificial intelligence (AI) and the automation of creative work have received little attention in craft education. This study aimed to address this gap by exploring Finnish pre-service craft teachers’ and teacher educators’ (N = 15) insights into the potential benefits and challenges of AI, particularly text-to-image generative AI. This study implemented a hands-on workshop on creative making with text-to-image generative AI in order to stimulate discourses and capture imaginaries concerning generative AI. The results revealed that making with AI inspired teachers to consider the unique nature of crafts as well as the tensions and tradeoffs of adopting generative AI in craft practices. The teachers identified concerns in data-driven design, including algorithmic bias, copyright violations and black-boxing creativity, as well as in power relationships, hybrid influencing and behaviour engineering. The article concludes with a discussion of the complicated relationships the results uncovered between creative making and generative AI.


Introduction
Artificial intelligence (AI) is increasingly embedded in everyday life with new products, services and systems (Stoimenova and Price 2020). The 2000s has seen AI-based technology adopted in an increasing number of use domains, user groups, devices and purposes (e.g. Russell and Norvig 2022). AI has also attracted visions of how it will influence the decisions of people, organizations and institutions (e.g. Lupton 2020). Much has been written about the potential of AI, but the growing extent of tracking, data mining, profiling and modelling has also fuelled anxiety over privacy and security threats, along with resistance to new types of power relationships, total surveillance, algorithmic bias, hybrid influencing, behaviour engineering and invisible processes of prioritization and marginalization (e.g. Bowker and Star 2000;Hendricks and Vestergaard 2018;Kramer, Guillory, and Hancock 2014;Noble 2018;Valtonen et al. 2019;Zuboff 2015).
Another defining feature of the 'Age of AI' is the ever-increasing automation of knowledge work (Brynjolfsson and McAfee 2014). While computer-based automation of tasks and jobs has affected labour markets since the 1950s , the latest wave of AI was made possible by the availability of massive amounts of data, increased computational power and more sophisticated statistical techniques (Darwiche 2018). As computers have taken over many routine-like and boring tasks that were previously performed by people, the computer revolution has also been said to liberate time and offer new opportunities for human imagination and creativity (Vartiainen et al. 2020). Since the mid-2010s, machine learning and neural network-based AI systems have slowly encroached on various kinds of creative domains that just two decades ago were widely thought to be relatively safe from AI-driven automation (Caramiaux and Fdili Alaoui 2022). Generative machine-learning models are already used, for example, for creating digital art (Audry 2021), composing music (Huang et al. 2020), free-hand sketching and drawing (Ghosh et al. 2019) and creative writing (Gero and Chilton 2019). While generative models present real threats to the status and role of the artist in contemporary societies, providing appropriate AI education and tools for learning may also make creative endeavours more accessible and democratic, and lead to novel outputs (Fiebrink 2019;Audry 2021).
Amidst these profound transformations, there is an evident need for educators to understand and acknowledge how artificial intelligence and machine learning are disrupting the labour markets and the prosperity of democratic societies as well as our everyday life (Hintz, Dencik, and Wahl-Jorgensen 2019). At this point, a considerable body of research has accrued on artificial intelligence and machine learning in the context of education. Holmes, Bialik, and Fadel (2019) wrote that most of that body of research was traditionally focused on (a) AI tools to support learners, (b) AI tools that are used to study learning processes, and (c) AI tools for supporting administrative functions in schools. In addition, the machine learning revolution of the 2010s did not go unnoticed by education researchers, and literature reviews have identified a rising number of initiatives on teaching AI to novice learners. In the past few years, those literature surveys have revealed a range of AI tools and instructional units (Marques et al. 2020), research reports on AI education for all (Long and Magerko 2020) and AI initiatives in K-12 (Zhou, Van Brummelen, and Lin 2020;Tedre et al. 2021). Researchers in the field of Human-Computer Interaction and Design Research have begun to investigate the way artists are using AI in their creative practices (Fiebrink 2019;Caramiaux and Fdili Alaoui 2022;Audry 2021). However, generative AI-in which the breakthrough applications are much more recent (Bommasani et al. 2021)-has received little, if any, attention in craft education.
As an early work on generative AI in craft education, the aim of this article is twofold. First, it introduces the reader to some aspects of state-of-the-art algorithmic art-more specifically, text-to-image generative AI for the creation of visual images from text prompts. This introduction is intended to contribute new perspectives to discussions in the field of craft, design and technology education on the role that AI-powered technology can play, especially in the early phases of the ideation and design process. Second, the article aims to shed light on current and future Finnish craft educators' (N = 15) insights into the potential and challenges of this still nascent technology. Following the research-creation approach (Lupton and Watson 2022), we organized a hands-on workshop in teacher education, in which the participants collaborated to create visual images using generative AI, and then reflected on the role of AI and automation in the process of ideation and product design. Instead of promoting hype-surrounded imaginaries 1 of possible AI futures, the aim of this exploratory case study was to use creative making to stimulate discourses and capture insights into the everyday realities, hopes, concerns and future imaginaries that characterize teachers' encounters with generative AI. The paper addresses the following research question: What kinds of benefits or shortcomings do craft teachers associate with ideating and designing using text-to-image generative AI?
This article begins by introducing the specific role that visual materials and techniques play in the craft design process. It then presents insights into how text-to-image generative AI has very recently become an increasingly popular means of visualizing ideas and creating digital content. Next, it presents the research methodology, research setting and analysis that illustrate the potential benefits, shortcomings and concerns that craft teachers associate with AI and the automation of creative learning. The article concludes with a discussion of the tensions and complicated relationships that were uncovered through creative making with generative AI.

The role of visual materials and techniques in the craft design process
Craft education at its core emphasizes learning by designing and making digitally or materially embodied artefacts (Seitamaa-Hakkarainen and Kangas 2013). The goal of craft education has long been to mirror the expert practices and to cultivate designerly ways of thinking and practicing (Pöllänen 2009;Seitamaa-Hakkarainen and Kangas 2013). Using the notion of a 'holistic craft process,' researchers have emphasized the creative nature of craft education that generally involves four phases: ideation, design, making and reflective evaluation of the output and the process (Pöllänen 2009).
Seitamaa-Hakkarainen and Kangas (2013) characterized craft and design as a form of inventive learning that requires going beyond what is already known and seeing beyond the obvious. Designing can be conceptualized as a process of creating something new, something that does not yet exist but 'ought to be,' like Herbert Simon (1969) has argued. Designing has also been characterized as a complex and iterative problem solving process, where the designers work with open-ended and ill-defined design tasks (Goel 1995). With this as the starting point, design-oriented learning is typically organized around complex real-life challenges rather than build-a-thing tasks or step-by-step product making that was once common in craft education (Pöllänen 2009;Seitamaa-Hakkarainen and Kangas 2013;Vartiainen et al. 2020).
The craft design process begins by exploring the design problem, along with seeking inspiration and visual reference material. When generating novel ideas, designers typically seek various sources of inspiration, such as photographs, artwork, swatches, previous garments, artefacts or natural objects (Keller, Pasman, and Stappers 2006;Petre, Sharp, and Johnson 2006). These external sources of ideas may provide, for example, forms, materials, patterns, colours or textures that trigger designers' imagination (Eckert and Stacey 2000;Omwami, Lahti, and Seitamaa-Hakkarainen 2020). While designers may use real physically present artefacts and objects as sources of inspiration (Eckert and Stacey 2000), the Internet has become the dominant source for finding visual reference material (Keller, Pasman, and Stappers 2006). Eckert and Stacey (2000) argued that almost all design proceeds by transforming, combining and adapting elements of previous ideas and designs. Designers are not just gathering a collection of various kinds of reference material, but they use combinations and modifications of them to generate new ideas and thoughts (Eckert and Stacey 2000). Complexity in the overall design arises from such combinations and configurations among simple design elements that stand in a complex network of relationships from which novel ideas and design properties may emerge (Petre, Sharp, and Johnson 2006). Furthermore, these configuration elements in a design process are also affected by a number of external and internal constraints (Seitamaa-Hakkarainen and Hakkarainen 2000), whether they are cultural, ethical, legal or economic questions, or more pragmatic conditions (e.g. available time, material, resources and tools).
What is more, research has shown that sketching and visualization play a key role in externalizing and further developing design ideas (Eisentraut and Günther 1997;Omwami, Lahti, and Seitamaa-Hakkarainen 2020), especially in craft education (Pöllänen 2009;Seitamaa-Hakkarainen and Kangas 2013). The creation of various kinds of representations, such as sketches or prototypes, allows the designer to explore, develop and evaluate the form and function directly (Hennessy and Murphy 1999). Moreover, by creating many sketches or visual representations, a craft designer may explore alternative solutions and test them before bringing the designed product to production (Seitamaa-Hakkarainen and Hakkarainen 2000). As the idea advances, more detailed drawings or visualizations are created, including technical drawings and technical illustrations (Pei, Campbell, and Evans 2011). A further benefit of visualization is that through externalization, the mental ideas of the designer become visible and can be communicated to others for feedback (Seitamaa-Hakkarainen and Kangas 2013).
In the making phase, design ideas are further transformed into material embodied forms and functional prototypes (Härkki 2018). In the field of craft education, handson manipulation of materials is typically an intrinsic part of the design process as it affects both the process and the outcomes of activity. According to Härkki (2018), material exploration facilitates the testing, evaluation and further development of design ideas and in an otherwise impossible manner. Accordingly, craft making relies heavily on material knowledge, such as knowledge of available materials, their technical properties, their sustainability and ways of manipulating and processing the required materials and tools (Härkki, Seitamaa-Hakkarainen, and Hakkarainen 2016). The importance of skills in handling materials and tools has also been analyzed through concepts such as tacit knowledge (Polanyi 1966) and embodied thinking (Patel 2008) that characterize the practices and knowledge of expert designers and craftspersons (Groth 2017).
Although people can design and make material artefacts without any technologically advanced tools, various kinds of environments, such as computer-aided design (CAD) and 3D modelling, have been developed specifically to support modelling and form-based prototyping in the early stages of the design process (Mathias et al. 2019). Such technologies may stimulate, facilitate, guide, externalize and otherwise mediate the process of design, thereby affecting the process of design as well as the kind of products designed. Moreover, material making and form giving can be enhanced by digital fabrication and computer-aided manufacturing technologies, such as 3D printers, laser cutters or CNC routers, that have been increasingly integrated as part of craft education (Vartiainen et al. 2020). While there is steady growth in the use of 3D modelling and fabrication technologies in school education, there is, to the best of our knowledge, no research on applications of text-to-image generative models in school education or in teacher education.

Generative models for producing photorealistic images
Recent developments in generative AI techniques have made text-to-image generative models increasingly popular for creating visual content and artwork based on text prompts in natural language (Dhariwal and Nichol 2021;Liu and Chilton 2022;Oppenlaender 2022). While the use of ML in creative work has historically required deep knowledge of ML and programming skills (Fiebrink 2019), what the recent models do is generate highly aesthetic images from the user's text prompts, requiring little to no understanding of the underlying technology (Oppenlaender 2022).
For example, the user can write a prompt (a text description of what the user wants) that tells the AI to imagine and visualize evening gowns with a description of the preferred style, colour, materials, scenery, lightning, and so forth ( Figure 1). A prompt for that could be, for instance, 'pink evening gown, clown gathering style, made of silk and uranium, with Mount Rushmore background, in dusk,-ar 16:9' (Figure 2). Based on the user's prompt, the AI then outputs an original image that has never existed before (McGeehan 2022). Figures 1 and 2 present evening gowns generated from the authors' text prompts (see Appendix A).  Since the mid-2010s, the AI community has seen a string of new types of text-to-image generative models, each better than the previous one (Dhariwal and Nichol 2021;Saharia et al. 2022). They all employ neural networks, trained using massive numbers of pictures and their textual descriptions publicly available on the Internet and in public repositories (Saharia et al. 2022, 9;Bommasani et al. 2021). The two most common image databases used to train text/image AI are LAION-400M and LAION-5B, which provide 400 million and 5.85 billion image-text pairs, respectively. 2 Gradually, by passing millions and millions of images and textual labels that describe them through neural networks, those neural networks learn to create similar-but never the same-images from patterns they have learned, guided by the user's prompts.
The current state of the art in this field relies on a class of techniques called diffusion models, which are able to achieve photorealistic quality and generalize to images that have never previously appeared in the set of pictures used to train the model (Dhariwal and Nichol 2021). That is, even if there were not a single image in the world of 'Yoda wearing a knitted heavy metal band fan shirt on Times Square,' the latest generative models would be able to create a photorealistic image of exactly that, as the models have been trained with plenty enough examples of each of the individual elements in that prompt (Yoda, wearing, knitted, heavy metal band fan shirt, Times Square).
Some of the best-known tools for text-toimage generation have become well known outside the AI community-take Dall-E and Midjourney, for example. Both implement a low-floor, high-ceiling and wide-walls principle (Resnick and Silverman 2005): They have a very low barrier of entry, as users can start to get the results they desire with next to no training (low floor), and with training, the results scale up to commercial-level and award-winning results 3 (high ceiling), and the results extend to a broad range of applications (wide walls). The images in this article were made using Midjourney, a popular community-based text-to-image generation tool.
Midjourney and other similar tools have quickly given birth to interest-driven contentcreation communities (Oppenlaender 2022). Similar to existing maker communities, these communities connect people from various backgrounds so that they can share their ideas, designs, instructions and works of art (Martinez and Stager 2016). In those communities, designers, programmers, artists, hobbyists and anyone interested actively share their work, advice, prompts and prompt-crafting hints. In this way, the community creates learning resources for others to build on when creating or customizing their own ideas. This growing ecosystem of services, easy-to-use tools and enthusiastic communities also contributes to making text-to-image generative content creation more accessible to newcomers and non-technical users (Oppenlaender 2022). Oppenlaender (2022) pointed out that with text-to-image generative systems, basically anyone can create impressive and novel content. This also raises new questions about whether text-based generative content is creative and what the level and nature of human creativity involved in the creative process actually is (Oppenlaender 2022). Audry (2021) argues that generative AI nurtures new human-machine relationships by displacing and reconfiguring the agencies involved in the creative endeavours. The creativity involved in text-based generative content is said to be connected to human-computer cocreativity (Rezwana and Maher 2022) and the iterative practice of prompt engineering (Gwern 2020) or prompt-crafting. The creation of text-based generative content is typically driven by users' creative explorations and interactions with generative models that are trained using Internet-scale quantities of data (Gwern 2020;Oppenlaender 2022). Oppenlaender (2022) argues that human creativity in text-based generative art is not located in the end product (the digital image), but emerges from the evolving interaction between humans and the AI (e.g. prompt engineering). In other words, the creation process is crafted in response to human initiative, but the system responds with its own decisions and thus, people need to engage in an adaptive process as they try to shape the system to respond to their desires and needs (Audry 2021). In practice, this may involve the iteration of textual prompts, along with image iteration (Figure 3).
In addition, text-to-image generative models have prompted discussions about whether these tools are creating new opportunities for artists and designers, or if they are disrupting the job markets for the creative fields instead (Audry 2021). There has been a backlash against the common practice of training models using publicly available images on the Internet with no curation, no consent from or knowledge of the image owners and no ability to exclude one's pictures from the training data, other than making pictures private altogether (Paullada et al. 2021). The ability to near-faithfully reproduce styles from art history, pop culture and individual artists has also raised critical voices and concerns surrounding copyright violation (Liu and Chilton 2022). Moreover, as an uncurated data-driven system, the digital content created by generative AI can incorporate the biases and stereotypes that have long existed in society, such as favouring or marginalizing individuals or groups based on their ethnicity, gender, sexual orientation and other attributes. In addition, democratic society has been challenged by fake content (Valtonen et al. 2019).
To sum up, the rapid development of textto-image generative models may open up novel opportunities and challenges for fostering innovation, creativity and participation in school education. However, as the technology involved is very recent, little is known about how it will change the dynamics of learning communities and classrooms. There is an evident lack of research on teachers' insights into AI, experiences with it, and attitudes toward it, as well as on how AI can liberate or impede the values and skills emphasized in craft, design and technology education. There is a need to explore the AI experiences of teacher educators and pre-service teachers, as they play a crucial role in supporting students in learning the creative competencies needed today and in the future.

Methodology
This study builds on our previous work and design-based research on integrating AI topics into school education (Vartiainen et al. , 2022. In this exploratory case study, we used a research-creation approach developed for understanding imaginaries concerning emerging technologies, such as AI and AI-related automation (Lupton and Watson 2022). The research-creation approach engages participants in the creative making of artefacts with the intention of inspiring discussions and helping to make tangible the complicated relationships with emerging digital technology (Lupton and Watson 2022).

Context of the study
The empirical data for this study were collected during a workshop that was held in a Finnish university programme on craft teacher education in the fall of 2022. Unlike many other countries, Finland has included craft into general education as a compulsory school subject for all students (Pöllänen 2009). To qualify as a craft subject teacher, teachers are required to take a 5-year academic programme that contains the Bachelor of Arts (Education) degree (180 ECTS) and the Master of Arts (Education) degree (120 ECTS). Craft teacher training follows the Finnish National Core Curriculum for Basic Education (FNBoE 2014), which emphasizes the holistic nature of the craft process (Pöllänen 2009). In addition, Finnish craft education is oriented toward a critical understanding of everyday technological phenomena, as well as the use of technology for enhancing inventive, creative and inquirybased learning (FNBoE 2014).
The workshop was designed and run by a multidisciplinary team of researchers in computer science and education research. The design and implementation processes in the workshop were co-configured with teacher educator experts to serve their specific needs and requests. Before the workshop, we held discussions with the teacher educators and set the goal of the 3-hour workshop as a way for the participants to learn and experiment with text-to-image generative models as well as to reflect on the role that new AI-powered algorithmic art technology may have on education.

Participants
The call for the workshop was distributed openly to all pre-service craft teachers and teacher educators in craft science involved in the educational programme. The number of participants was limited to 15 due to limited technical resources. The places were filled in order of registration.
The research was conducted in accordance with the instructions of the Finnish National Board on Research Integrity (TENK 2019). All data collection procedures received institutional approval by the administration of the Department of Teacher Education. In the fall of 2022, an introductory email was sent to the teacher educators and student teachers, including information about the workshop, as well as the data collection and research related to it. As per the institutional policy required by the university, a participant information sheet was sent to those who had registered for the workshop (including a description of the purpose of the study, a statement that participation in the study was entirely voluntary and that participants had the right to withdraw, information about the types of personal data that would be collected, how the data would be processed, and how the research results would be reported). In addition, participants were asked if the digital artefacts produced in the workshop could be published in research publications and used in presentations regarding the research results.
The group of 15 workshop participants included 5 teacher educators and 10 pre-service teachers. To ensure the anonymity of the participants, no demographic details were collected. The participants were informed that no previous experience with AI was required and that a diversity of perspectives was welcome.

Workshop
Before the workshop, all participants were asked to write down their preconceptions about AI and its relationship to the design process and the teaching of design. In addition, the participants were asked to elaborate on what they wanted to know about AI. This information was used to tailor the content of the workshop to meet the interests of the participants. In addition, these preconceptions and learning needs about AI were anonymously shared at the beginning of the workshop.
The workshop's pre-task revealed that the topic was new for almost all of the participants. Accordingly, the workshop began with an introduction to AI and the mechanisms of text-to-image generative models in particular. This introductory lecture also provided insights into the ethical concerns related to AI, such as algorithmic bias and copyright debates. The participants were shown various visual examples of text prompts and the pictures that Midjourney generated from those prompts. The basic principles of prompt-crafting and iterative improvement were presented. After the researchers had shown examples of how to use the Midjourney interface, the participants were asked to split into three small groups. Each group was given a computer and asked to collaboratively explore and experiment with Midjourney in practice. Over the next 90 min, the participants produced a large number of various kinds of images (see examples in Figure 4), and they were shared in the joint Discord channel. On the channel, all the groups were able to observe what the other groups were doing and what kinds of prompts they had tested and further crafted.
After the exploration phase, the small groups were asked to share their thoughts, ideas and experiences through guiding questions: (1) What kinds of opportunities, challenges and ethical questions does artificial intelligence pose for the process of design and its teaching in schools? and (2) What things can be automated in the process of design or in teaching it? What can't be automated? What should not be automated? These insights were shared in a joint discussion at the end of the workshop.

Data collection
Similar to the study by Lupton and Watson (2022), the use of creative making in the workshop ultimately aimed to prompt our participants into sharing their insights and developing new ideas rather than simply answering a set of interview questions. Accordingly, the data collection was embedded in the joint making and discourses, which were recorded using two Go-Pro cameras and two audio recorders. Moreover, two researchers observed the emerging activities and provided on-demand support.

Data analysis
Audio data from three group discussions and joint reflections were transcribed, pseudonymized and loaded into the Atlas.ti qualitative analysis software application. The data were analyzed using thematic analysis, which has become a widely-used method for analyzing qualitative data (Braun and Clarke 2006). First, the data were read several times to gain a general sense of the nature of collaborative discourses. Second, meaningful units relating to the research question were marked and given preliminary, descriptive names. Thirdly, these data-driven interpretations were reviewed in terms of previous research, that is, whether previous research has identified similar themes (e.g. key phases of craft design process, impact of AI). Fourtly, the coded meaning units were then subjected to a further analysis, in which different codes were compared and contrasted in order to identify how they could relate with each other and form an overarching, candidate themes. The final phase involved reviewing potential themes in which the codes were aggregated into two main themes and sub-themes within them, as elaborated in the result section. To increase the reliability of the data analysis, the interpretations were discussed between the two researchers, and several data examples are provided in the results section for readers to evaluate our interpretations (see Hammer and Berland 2014).

Tensions and tradeoffs in the process of design
Ideation. On a positive note, the participants saw many opportunities for enhancing idea generation in the process of product design. Jaana, from the teaching staff, reflected, 'In a way, it's like combining something new, like seeing things again from a new perspective.' The visualization of impossible ideas triggered pre-service teachers' reflections. Aapo (preservice teacher) stated, I'm interested in things or possibilities that are not really true or where there is an impossible shape, or some pattern is in itself outside of reality. But even if it's not in the real world, it can be further developed.
Externalization. In addition, Aapo further saw opportunities for scaffolding that the externalization of mental images could open up for children as well as for teachers through gradual prompt-crafting: I would see that if AI would help the child [iterate and develop ideas] in the designing: well, okay, … the yard of dreams: I want a single-family house there, a red house, then there will be … no no, I want that bigger house, okay now that's it, I want it to be made of wood, a big red wooden house. Then, it starts to do it little by little and thus helps the child in a way. It's okay; it's starting to look like that. I also want Pokémon in the yard (laughs), so something like this. So after this, the child and also the teacher will have a fairly similar idea of what it can become. Then, it can also help the teacher and the child as well.
Moreover, his peer, Siiri, elaborated on Aapo's idea by stating that the text-based generation of visualizations may help children in wording vague ideas: 'Like with small children, if we're planning something, then they have to put their ideas into words. So, then, you learn something like giving words to things that you are thinking about or what you could write there.' Liisa, from the teaching staff, reflected on the potential benefits of new kinds of externalization: 'Is it like a more playful one, it [Midjourney] did not understand now what I meant by this, so I will try this word. Would this kind of frustration go away?' Design constraints. The teaching staff were concerned that ideas that were visualized using AI might not stay within the design constraints, such as limitations regarding materials, available resources and actual craft skills and the knowledge of students. Jaana stated, You can create such really cool visual ideas with that. But then there's the question of how to make them concrete. It can be distressing for someone who is not an experienced maker or actor. Or, when a student brings it to the teacher, saying that he or she will make this because it is so cool. But then, he or she does not have any idea of what it really takes to make it, how it can be made, or if he or she has the skills to do that … do we have the materials? Or is it just some ideal of clothing? Does it really fit when worn, or is it aesthetics before ergonomics, and so on?
Taking an alternative perspective, teachers also pondered the opportunities of using AI-generated visualizations for analyzing situative design constraints, like Vuokko proposed: 'It could guide [learners] to analyze what would be required from me as a learner or craftsman to implement that design, and of course the teacher's help would be so important.' In addition, teacher educators also reflected on the positive feelings and wellbeing that may emerge from digital design. Jaana said, 'Many people like to make clothing designs. They get that good feeling from the fact that they are just designing.' The creation of digital designs was also associated with clothing avatars, which could potentially decrease material consumption in real life. Vuokko elaborated, 'It can even reduce production then this way, because you just dress yourself in the digital world.' Making. Many participants emphasized the embodied nature of making that is missing from digital design. Anita, from the teaching staff, reflected, 'What should not be automated, in a way, the embodiment … physical presence or doing, modifying the materials … or at least I want to physically modify the materials. I can feel the tools and materials in my hands.' Likewise, the pre-service teachers reflected on how touching, forming materials, and making with your own hands gives rich sensory experiences and activates brain areas that are crucial in child development. Timo stated, Humans have hands and need a sense of touch and holding things in their hands that develop the brain, for example, that frontal lobe. And then there is persistence, and so forth. So there can be AI assistance and designing made by AI, but it does not exclude the fact that there is and should be such an analog side as well.
In addition, pre-service teachers reflected on the difference between the digital and the material making of things in terms of feelings of well-being. For example, Anna said, 'Because humans are so strongly primates […] crafts and baking and gardening and such things in which you are doing with your hands produces endorphins in a completely different way as well as a different feeling of satisfaction.' Liisa also reflected on how AI-powered design may hinder opportunities to learn how to deal with difficult emotions that are an inherent part of creative making: With a few clicks, it quickly produces absolutely wonderful visualizations. This might perhaps circumvent some feelings of frustration at some point, but at the same time, it presents a missed opportunity for learning exactly what is learned through that frustration. And yet that is so essential for development, growth, and learning.
Moreover, the teaching staff also reflected that learning crafts and craft-related skills in designing and making is a slow process that takes years of practice. There again, AI-powered design is characterized by dynamism, as Katri critically stated: If the machine does everything, then it is like handing your brain over to an external source. I don't need to be a person anymore; the machine does everything for me. But then, making crafts is slow, it's done slowly, but this [AI] is making designs in just a few minutes, and using your own brain is completely gone with that system. Assessment. The use of AI-powered design tools prompted few questions about the assessment of creative learning, especially concerning the design skills under assessment. Vuokko said, 'About that assessment. If a pupil or student uses this [text-to-image generative AI], what are we assessing in terms of his or her design skills? Are we evaluating whether he or she has invented good prompts?' Future skills. Exploration with AI prompted the teaching staff to contemplate the long-held fears of automation and its relationship to craft making and teaching. Anita stated, Already in the 1980s, we were afraid that when the industry mass produced clothes, craft as a school subject would be pointless and not needed anymore. So, it was already discussed when we were students. But, it didn't disappear anywhere, and the hobby boom of crafting has increased. However, pre-service teachers also discussed how developments in technology are, in fact, rapidly transforming work life and society, which makes it difficult to predict the future skills that the next generations will need. Anna reflected, Artificial intelligence will surely replace certain work tasks. It's just like that. I don't think anyone doubts it. But when we talk about school education and future skills, what should be taught to children … And the fact that, well, children are not, in a way, taught for the existing world where we are today, but to where the children will be in ten years or in fifteen years. In a way, if this technological development continues at that pace, and it will also probably continue to accelerate exponentially, even choosing their future skills can be quite a challenge, to know what we can and are able to teach.'

Unforeseen tension and concerns
Biases. Making with AI prompted teachers to engage in deep reflections on creativity. The teaching staff recognized some similarities between the way AI, on the one hand, and human designers, on the other hand, rely on what they have seen before. People's design ideas rarely appear out of thin air, but reflect what people have seen and experienced. Sometimes, people actively look for ideas. Anita stated, In a way, we do the same thing that machines do. We surf on social media; we see pictures all the time on Pinterest, and we look at pictures on Instagram. Our brain is doing it all the time, and then all of a sudden this [AI] produces it. But it doesn't appear out of nowhere. It comes from a stock of images.
The teaching staff and pre-service teachers alike pondered whether too strong a reliance on existing ideas and styles could lead to a vicious circle of repeating the same ideas ad nauseam and reinforcing existing biases. Liisa asked, Somehow, I'm thinking about stylistic trends and reinterpretations of them, and are they not going to be done? So, more material is produced for us, so that picture looks like this, and it looks like this, that it has these certain features, but the data is still limited. And we just talked about those biases, aren't they repeated all the time in the new material that is produced with those certain words … aren't we still left with a biased view of certain things?' Copyright. AI-generated images also prompted much discussion on copyright. Silja (pre-service teacher) deliberated, But isn't it the case that the data that the artificial intelligence uses to create those images, isn't it that it uses everything that is available on the Internet? In other words, it is art made by other artists, and it uses it to form its own image. So, it actually, in a way, it's like robbing other people's artwork. Without asking what permissions, like hey, I'm now using this. For example, just what you created, that steampunk picture [ Figure 5] … It might actually have just the layout, the way it has those characters, just like in someone else's artwork.
Black-boxing creativity. Exploring AI led teaching staff to debate the definitions of creativity. Vuokko said, I was thinking about that concept of creativity. That isn't one definition of creativity that you know how to make or combine existing or old things in a new way. That's what this definitely does. But then I thought about what were the things that it pulled together: wouldn't it be great to make it visible? But we cannot see those traces. We only see the words that were put in it. It got those words and tips, and by combining them, it created a new product. But is it creative or not? Moreover, the pre-service teachers also recognized the difficulties in locating human agency in an opaque, black-box system, particularly whether and to what extent the output is based on one's own or somebody else's choices.
Marjut reflected, 'Who programs that artificial intelligence? How does it affect our decisionmaking and designs and others? And you should probably be well aware of that.' She further recognized that platform owners hold a significant degree of power in the process of design: 'There is power behind this, but who uses that power and what does it want?' Behaviour engineering. Students also raised other concerns with regard to the power of algorithms, affecting not just what people can create and how these systems are used but, more profoundly, how AI-generated images with personalized styles can be used for marketing and behaviour engineering. Liisa said, 'Think about those styles. Everyone probably has a favorite aesthetic style that appeals to us, and if someone knows what it is, then someone could feed us some material that suits our aesthetic taste for brainwashing.' Apo reflected on how AI-generated images can affect our feelings of trust and even alter our cultural memory: Then the misinformation starts. I don't remember the term for that word, but the fact that an image is created of an event, an image is created from it, and what kinds of emotions it evokes in that moment, five years from now, ten years from now … then the question is what about the memory history of mankind if we start to trust the images produced by artificial intelligence, and soon we no longer recognize what is the real image and what is fake?
While the teachers were critically reflecting on the potential risks and harms of AI-generated images and data-driven design, there was recognition that such issues should become an integral part of teaching. Liisa was concerned that we probably cannot put this back into the bottle anymorethese images generated by artificial intelligence and this visual world. We can't get it there anymore, but it's just that in our teaching, a really essential part would be to understand how it can be biased and what biases exist.

Discussion
The aim of this study was to shed light on Finnish pre-service craft teachers' and teacher educators' insights into the use of text-toimage generative models (images generated by AI to match users' text prompts) in craft design and learning processes. Through creative making, this study aimed to stimulate discourses and engage workshop participants to develop new insights into the potential and limits of generative AI.
The analysis of emerging discourses revealed how reflections on generative AI were highly situated in terms of the holistic craft process (Pöllänen 2009) that forms the core of the Finnish National Core Curriculum. Teachers talked about how AI may assist the ideation process by providing new perspectives and visualizing the possible and the impossible. AI was also envisioned to help children externalize vague design ideas in a manner that makes them visible to teachers (Seitamaa-Hakkarainen and Kangas 2013).
At the same time, the participants also pointed out various limitations of generative AI for craft education. In terms of design constraints (Seitamaa-Hakkarainen and Hakkarainen 2000), they suspected that the use of generative AI would create completely new skill gaps where AI-created visualizations and the available material resources or actual skills and knowledge of the students do not meet. In addition, digital design was perceived to lack the embodiment and materiality that characterize the essence of craft making and craft education in Finland (see Seitamaa-Hakkarainen and Kangas 2013). Participants associated touching and forming different materials with one's hands with children's brain development, indicating a view of what should not be automated. Likewise, AI was discussed in terms of emotions and embodied interactions that were also seen as crucial for human learning, development and well-being. Indeed, neuroscientific research has shown that the somatosensory, motor, and visual areas of the cortical surface are activated during craft making, and the stimulation of these areas is crucial in childhood (Huotilainen et al. 2018). According to Huotilainen et al. (2018), neuroscientific research has also confirmed the potential of arts and crafts as a means of enhancing well-being. In other words, creative making with generative AI prompted teachers to strengthen their views on the unique nature of handmade crafts and the skills that they aspire to cultivate in their teaching, and how these aspirations relate to tensions and tradeoffs when adopting generative AI in craft practices. In addition, making with AI prompted teachers to ponder the assessment and development of skills, especially in terms of future-oriented teaching.
The analysis also showed how collaborative making with generative AI stimulated critical discourses on new topics in the context of craft education. Introductions to and hands-on exploration with data-driven design prompted participants to deliberate over many concerns evidenced in critical media studies, such as new kinds of power relationships, algorithmic bias, hybrid influencing and behaviour engineering (e.g. Zuboff 2015; Hendricks and Vestergaard 2018;Kramer, Guillory, and Hancock 2014;Valtonen et al. 2019). These concerns were localized in making design decisions and feelings about data-driven futures.
Firstly, of the many types of bias in machine learning models, the workshop type facilitated discussion of three readily identifiable types of bias in text-to-image generative models. Those three arise from the training data and are related to misrepresentation (e.g. harmfully stereotyped minorities), underrepresentation (e.g. eliminating occurrence of one gender in certain occupations) and overrepresentation (e.g. defaulting to Anglocentric perspectives) (Bommasani et al. 2021). The participants' concern with the possible homogenization of perspectives is justified and real: The case Greg Rutkowski (Heikkilä 2022) showed how one artist's style can get to dominate a whole genre through prompt engineering (thousands of people prompting AI to draw 'in the style of Greg Rutkowski' and making the outputs available online, which further amplifies the volume of an already popular style in the training data). The workshop participants experimented with niche Peruvian designs that might suffer from underrepresentation or misrepresentation due to possibly limited volume in training data, but the results did not lead to discussion.
What is more, due to how they are trained, the models are likely to default to the preferred representations of a narrow segment of the society. Every data set-even those that contain billions of image-text pairs scraped from the Internet-incorporates some world view, and slices the world into categories that may be highly problematic (Bowker and Star 2000;Keyes 2018). That makes Marjut's question relating to politics and power-'Who programs that AI? [And how does that programming affect everyone else's decision-making?]'extremely important. In the politics of AI, neither the creators of the material that the AI is trained with nor the people who appear in countless photos learned by AI models have any power or agency (Crawford 2021). Unbiased as they may seem at first, AI models are politically and ideologically laden ways of classifying the rich social and cultural tapestry of the Internet-which itself is a pale reflection of human diversity.
Secondly, concerns about new kinds of behaviour engineering based on aesthetic styles were also raised, along with concerns about the production of fake images that may ultimately alter our cultural memory (are all AI-generated images 'fake,' and if so, what is 'genuine'?). By critically weighing the societal implications of generative AI, the participants' reflections addressed not just the potential, but also the imperfections, fragility and subjectivity of these technologies that are becoming part of the world's fabric (Audry 2021).
Thirdly, explorations on generative AI prompted teachers to re-evaluate the concept of creativity (Oppenlaender 2022), and also the ways in which data-driven design and blackbox algorithms could lead to the repetition of ideas and styles, to maintaining existing biases, and even to copyright violations. While there is a legitimate concern that generative AI could end up robbing artists of their livelihood, these easy-to-use technologies may also provide means to foster imagination, self-expression and networked co-creativity in a new way (Fiebrink 2019;Audry 2021;Oppenlaender 2022). Generative AI provides a new medium for expressing one's design ideas; a new medium for creativity. Yet, the question of copyright would have deserved more time for a facilitated discussion, as the participants chiefly considered the copyright question from the intellectual property viewpoint but not in terms of the fair use doctrine of the copyright law. While popular training sets contain images that are protected by intellectual property laws, it has been argued that the fair use doctrine, which has been interpreted as sanctioning the automated massive scale scraping of the Internet for training data, also ultimately aligns with the goal of creating fairer AI systems (Levendowski 2018).
While creating new designs with AI and engaging in shared discourses, the teachers also agreed that many of the issues and concerns that were raised should at least be taken into account when introducing AI topics in craft education. Creative making may be an important way to raise awareness and generate alternative or resistant imaginaries (Lupton and Watson 2022). However, as the tensions captured in this article are based on craft teachers' first encounters with generative AI, it remains to be seen whether and how these technologies and experienced tensions will be actualized in educational practice. This study is also limited in that it focused exclusively on text-to-image generative AI and did not explore the benefits or shortcomings posed by other developments in AI. While the findings of this exploratory study cannot be generalized, early results on teachers' insights and experienced tensions are important for the development of craft, design and technology education research, as well as for providing future teachers with the ability to prepare their students to become creative and critical makers in the age of AI. By introducing perspectives related to text-to-image generative tools, we also wish to initiate discussions and encourage researchers, teachers and pre-service teachers to further explore and critically evaluate these ambivalent avenues in education.

Notes
1. By imaginaries this article refers to socially constructed realities and visions of desirable futures, that are animated by intersubjective understandings of social reality attained through and supported by technological advances (Jasanoff and Kim 2009). Sociotechnical imaginaries project visions of what is societally desirable as well as warn against risks, downsides, and hazards intertwined with technological innovation, as originally theorized by Jasanoff and Kim (2009). 2. The non-profit organization laion.ai provides several database options: LAION-400M contains 400 million English image-text pairs and LAION-5B contains 5.85 billion multilingual image-text pairs.