"I want it to talk like Darth Vader": Helping Children Construct Creative Self-Efficacy with Generative AI

Creativity scholars agree that creativity refers to the combination of the elements of novelty and usefulness [3, 4, 8, 11]. However, they also emphasize the importance of socio-cultural contexts in defining these criteria [3, 43]. Children’s limited understanding of social norms can significantly impact their creative expression [53]. For example, Rosenblatt and Winner [79] distinguished between preschool and older elementary school children’s drawings. They called the preschool stage the “pre-conventional phase," characterized by children expressing their own preferences without conventional influence. In contrast, older elementary school and early middle school children are in the "conventional phase," aiming for more realistic art, influenced by their understanding of cultural norms and social expectations. Similarly, the “fourth grade slump,” initially introduced by Paul Torrance, describes a decline in original thinking when children reach fourth grade (around 9 years old) [94]. This phenomenon, attributed to experiences in formal education emphasizing social norms [80], is accompanied by a decrease in creative self-efficacy, or a person’s belief in their ability to create [93], as grade levels and age rises [12]. Furthermore, there is also indication that creativity plays a role in learning as it contributes to identity formation [8] and cognitive development [57], as well as serves as the foundation for creativity acknowledged by others [11, 52]. Beghetto and Kaufman label learning moments “mini-c creativity,” which they define as creative moments meaningful to the individual [11]. Considering the vital role of creativity in children’s development, fostering social understanding, and learning, there arises a significant question: How might GenAI either amplify or harm children’s creativity? More specifically, we focus on child-GenAI interactions as a potential producer of the moments of “mini-c creativity,” that lay the foundation for learning about creativity and AI [11].

Constructionism, associated with the work of Seymour Papert [75], is a learning theory that posits that people learn best through creating artifacts. Tools such as Papert’s "Logo Turtle" and Resnick’s Scratch¹ have been used both to evaluate knowledge of programming concepts [35] and assess attitudes and computational thinking in children [63, 76]. Research has also shown that creating and programming custom models can help children shift their attribution of the agency of AI systems from the devices to the programmers of those devices [31].

There are examples of educators applying constructionism in AI Literacy curricula [1]. There is also ongoing development of tools aimed at augmenting children’s creativity, such as StoryDrawer [107] or The Invention Workbench [14]. However, many AI Literacy curricula rely on the use of widely available AI tools [95], not ones that are made to directly support constructionist learning environments. AI tools, such as Google’s Teachable Machine or ChatGPT are not always designed to augment children’s creativity. A systematic review conducted by Casal-Otero et al. [21] identified two main approaches to AI literacy for K-12: Learning experiences focused on understanding AI and Proposals for implementation of AI learning at the K‑12 level. Ng et al. [70] also identified three dimensions of AI literacy for children that include AI Concepts, AI Applications, and AI Ethics/Safety based on the work of Touretzky et al. [95]. As AI literacy has developed, curricula aimed at older children (ages 11 - 17) have included activities related to computer science proficiency, such as understanding decision trees or adapting systems [55]. For younger children (10 and below) activities generally include exploring AI through storytelling or activities of play to help build an understanding of computational thinking [89, 90, 99, 103]. While we have an understanding of how constructionism can support learning for computational thinking, there is still question as to how GenAI tools differ from tools such as Scratch or AI literacy tools such as Cognimates [30] (a platform for building games and training AI models). Similarly, as creativity is a crucial skill needed to support children’s knowledge of themselves and society, we also need to explore how to GenAI might function as a creative support tool within the framework of constructionism. Exploring at home with families [32, 61], or in other learning environments such as public spaces or museums, has also been shown to offer potential opportunities to help children build their AI literacy and creative confidence [44, 60]. Therefore, we focus our study on the potential of GenAI as a tool for supporting constructionism of not only computational thinking, but also creativity, within a non-formal setting.

Within HCI, researchers have highlighted the potential of GenAI for Human-AI co-creation, emphasizing the development of models and design principles that prioritize human-centered AI (HAI). HAI comes in a variety of forms, but one prevalent concept is the collaboration of humans and AI working as a team [19]. This includes tasks that involve creativity, in which the human creator use AI to produce a creative output such as a story, a song, or a visual artwork. Frameworks such as the Human-AI Co-Creation Model demonstrate ways in which co-creation with AI may not only help to support the creative strengths of the human utilizing AI, but also allow them focus their energy on the “most creative part” [101, p.177]. In this way, many consider the AI system to embody the role of “AI Agent” [24], meaning that the AI acts as a collaborator during the creative process, with special attention paid to its potential in the ideation process [50, 58, 83]. Hwang [49] additionally argues that certain AI tools are better suited for certain types of tasks, presenting four categories of co-creative AI tools including Editors, Transformers, Blenders, and Generators. Each co-creative tool has its own strengths in facilitating creativity during the creation process [49]. Yet, using AI effectively requires understanding of the model and one’s goals. According to Chang et al. [22], creators often use highly specific language and templates to craft prompts to achieve desired outcomes with GenAI.

Frich et al. [37] noted while the HCI community has seen a diversification of creative support tools (CSTs) across the creative process, many CSTs never leave the labs in which they were created. They also mention that general CSTs are more widely created and shared than ones for experts. Children’s needs are similar to the complex needs that expert users face in that they differ from the general public’s view. As the adoption of GenAI tools grows, we are gaining a better understanding of how AI can co-create with people. Despite this, with children’s differing creative needs, there remains a question about the specific ways in which these tools may need to be adapted to fit the unique creative and learning needs of children. Overall, we understand the ways in which children’s creativity differs from adults, the possibilities of constructionism to help support learning, and the potential of CSTs to support human creativity, yet, much less is known about how GenAI might function as a CST that helps support children’s learning and creativity. Therefore, we decided to explore the potential of GenAI by centering the child and their experience with these tools.

For this study, we employed a participatory design (PD) method called Cooperative Inquiry  [33, 106]. PD is a method that allows both designers and users to co-design new technologies, democratizing the design process by directly involving users. Cooperative Inquiry (CI) focuses on design partnerships between children and adults [33, 40, 106]. CI allows us to gain insight into children’s learning and creative processes as well as empower the children to share their thoughts about AI in creative endeavors. Research within HCI has shown that PD techniques like CI allow children to more concretely express abstract, sensitive, and complex ideas on topics such as security and privacy [54, 105], gender and sexuality [56], and family finances [104]. It has also been used to understand children’s creativity [2].

The co-design group, KidsTeam UW, consisted of both adult design researchers (investigators, masters students, and undergraduate students) as well as child participants (n = 12) as seen in Table 1. For this paper, we use pseudonyms for all children’s names. All the child participants were recruited through mailing list, posters, and snowball sampling. All children had parental consent and child assent. The university’s Institutional Review Board for Ethics approved all research conducted. Participants engaged in six design sessions over the course of four months (February to May 2023). Five to eight adult facilitators acted as design partners for each 90-minute session.

Each design session consisted of snack time (15 minutes) to build relationships with the children, circle time (15 minutes) — a warm up activity in which we ask a “Question of the Day” to help adults and children get ready for the design activity — design time in groups (45 minutes), and finally, discussion time (15 minutes), where groups present their final designs and the whole team reflects on the design experience. We organized our sessions to focus on children’s general perceptions of GenAI tools within creative activities (DS1/DS2/DS4) and how they created and adjusted their process using the systems (DS3/DS5/DS6). We selected the GenAI tools based on ubiquity and public accessibility.

We recorded all design sessions via Zoom video chat on three to four computers depending on the number of design groups; we collected a total of sixteen hours and thirty minutes of video. The researchers also collected the designed creative artifacts of the sessions via screenshots (if digital) and pictures (if analog). Additionally, for each of the groups the adult researchers recorded the group’s collective ideas and thoughts on a Google Slides deck after discussing and summarizing with children at the end of each group session.

We employed a combined inductive and deductive qualitative approach for data analysis [88]. The first author initially created a codebook through inductive open coding of co-design videos, covering topics like "Context of AI - Ethics" and "What is AI? - Limitations." We then created Analytical Memos (AMs) through consolidating the videos and design artifacts. The AMs were written by one member, and then validated by a second member. The research team iteratively discussed and updated codes, refining the codebook three times to include categories like "Understanding of AI - AI Outputs" and "Expression with AI - Modifying the Tool." The final codebook can be seen in Table 2. After agreements had been made and all AMs were completed, the first author compared the codebook to work on AI literacy [32, 59, 70], constructionism [42, 75], and creativity [3, 4, 52]. Using the updated codebook, a primary coder applied the codes to each AM, as an example, applying the code Evaluating Outputs to the quote “Damian says it is ‘disappointing’ because the images have been inaccurate.” A second member reviewed each code, citing agreement or disagreement. After all secondary coding was completed, research team members discussed all instances of code applications with disagreement to reach a consensus [46].

Children in our study perceive GenAI tools as capable of producing creative outputs, but required external assistance, either from adult co-designers or peers, to develop a mental model that supported a tool’s creative uses. This emphasizes the importance of scaffolding to reshape children’s cognitive frameworks, encouraging a perspective of GenAI as an active, collaborative partner in the creative process. As Jimin expressed during DS6: “You could be creative [with Generative AI] but that doesn’t mean it would sound better,” where she later suggested that this was due to her feeling of a lack of “control.” Our analysis identified three factors influencing a child’s mental model of GenAI for creative uses:

In our study, children grasped how GenAI works by comparing their expectations with system outputs. They observed that GenAI frequently combined pre-existing content to generate responses, exhibited repetitive behavior in the responses, and had a capacity for learning and adapting through interactions. For instance, Cyno noted that ChatGPT generated Ed Sheeran song lyrics by blending existing songs (DS6). Similarly, Diago felt like ChatGPT just “kind of made a pattern” when generating results for creative writing (DS3). Jimin in response to the group’s prompt about what player ChatGPT would rather have on its sports team, predicted the typical disclaimer, “As an AI language model, I cannot…” (DS3), demonstrating children’s capacity to learn and anticipate typical AI behaviors. This understanding shaped the children’s perception of GenAI as a creator of pre-existing and ready-made works, limiting their creative ideas to what they believed the system could complete. They saw GenAI as a tool to execute rather than a collaborative tool for fully exploring their ideas.

Moreover, children shaped their mental model of GenAI by comparing it to other technologies. For example, they drew parallels to Google (Diago, Cyno, Damian, and Zane) and Siri (Alex). Diago highlighted a distinction by stating “with Google, people put the information in [beforehand], whereas GPT makes up answers as it goes" (DS1) and Alex pointed out that “you don’t really have to go online for Siri since you can just press a button without having to login" (DS1). Zane differentiated ChatGPT from a calculator, emphasizing “the calculator only does math, but ChatGPT does everything" (DS1). This comparative assessment extended to the way children phrased prompts, reflecting a mental model akin to familiar information-seeking technologies, as observed with Matt prompting ChatGPT with“How do I hack Google” (DS2) or the prompt “How do I know if a kids is scamming people with a lemonade stand?" as shown in Diago’s comic about a bad use of ChatGPT in Figure 2. These comparisons indicate that children construct their mental models based on more familiar technologies.

Furthermore, children tended to generalize ChatGPT’s affordances to all GenAI systems, leading to challenges with different tools. As an example, when creating with DALL·E, Cyno expressed frustration when DALL·E could not remember what her character Lisa looked like without writing the description each time (DS5). It was not until an adult informed her, unlike ChatGPT, DALL·E does not remember what the user previously prompted. Children formed incorrect assumptions of GenAI when they were not supported in identifying the specific affordances of a tool. This suggests that encouraging children to re-establish their model of GenAI as a tool with a specific purpose can help them to form more accurate mental models. Consequently, when their ideas did not align with the system’s expectations, children were more likely to abandon their original creative ideas, indicating a decrease in confidence and satisfaction.

Despite children’s increasing understanding of how these GenAI tools functioned, they frequently encountered moments of frustration when the tool’s outputs did not align with their expectations, resulting in issues with the "gulf of evaluation" [71]. These moments most often occurred when children felt the system was unable to represent their ideas of a topic they were excited about, such as when Damian was frustrated while asking ChatGPT to write a sad poem about rocks, in which the system continually reproduced the same poem. Additionally, they became frustrated with the lack of transparency of the systems, such as when Zia and Cyno were confused as to why ChatGPT combined Japanese and Korean together when translating a story they wrote (DS3) or when Alex and Eiko did not understand why all the pictures DALL·E created were creepy (DS4). When children felt the tool did not meet their expectations, they would engage in one of three behaviors:

If the task or topic was something the child liked, they were more likely to engage in the first two behaviors. If they either did not care about the task, or had tried several times to get their desired result to no avail, they would resign themselves to trying something new. We have listed examples from our sessions for each adjustment in Table 3.

These findings reveal that children go through a process of trial and error when using GenAI for creative tasks. These iterations are marked by evolving comprehension of what the tools can do as well as readjustments from the children in the description of their creative goals. Children were more likely to overcome challenges during interactions when the topic aligned with their interest. This suggests that children feel more confident when they believe the system can adapt to their interests, allowing them to adjust prompts and enhance their creative understanding. Systems that support iteration help children develop coping strategies that supported their creative goals.

The children found GenAI tools to be too formal and limiting in their language, causing challenges in navigating their creative processes and feeling confident in their creative goals. Throughout our sessions, children frequently critiqued the formal language of the GenAI tools. Damian, Diago, Zia, Alex, and Eiko noted the language used by ChatGPT sounded “scientific” and lacked their desired casual conversational tone (DS1). Similarly, some frustration arose when the AI would formally apologize when being corrected by a child (Zane), as well as when GenAI did not understand what they were saying, as evidenced by Matt’s exclamation: “It doesn’t understand what I’m trying to say!”

Instances of “formal language” often resulted from the AI assuming domain-specific terms, as seen when ChatGPT provided information about herding behavior rather than dinosaur friendships when Diago asked about “dinosaur relationships” (DS1). This misalignment also occurred when the children considered themselves experts (e.g., in Pokémon or a specific musician such as BTS), leading to additional frustration when the systems could not accurately capture their understanding of a concept.

Creative experiences for children are non-linear [92], and thus, language code-switching becomes a crucial skill for effective interaction with systems [13]. Damian, for example, while asking ChatGPT about Star Wars, suggested that the system should talk to them like characters from the series in order to help him write a story about the space franchise (DS1), stating “I want it to talk like Darth Vader.” This was further highlighted in our music session (DS6) where Zane, who did not have a musical background, struggled to evaluate the generated chord progression’s fit for his song. Zane and an adult co-designer noted that they did not know how to evaluate the chord progression generated because they did not know enough about Bebe Rexha’s music (the musician they were trying to replicate). For children, the systems’ lack of customizable language, tailored to their domain knowledge or personal interests, posed a barrier during creative experiences. Their enthusiasm and confidence were higher when the system aligned with their domain knowledge. Children suggested potential solutions, such as Cyno proposing a visual style picker for DALL·E to maintain style consistency (DS4) and Diago suggesting that DALL·E should allow children to provide other images as examples (DS5). In fact, domain-adaptive pre-training of language models has been shown to increase task performance [41] and may offer one way of supporting children’s creative experiences.

In sessions four through six, children utilized multiple tools, transposing ideas across contexts, adapting content to meet each system’s requirements. This included summarizing parts of a story generated by ChatGPT so that it could be used in DALL·E or deciding if the music style generated by the Multi-track Chords tool fits their generated lyrics. This interchangeability led to a reliance on preferred mediums. In DS5, when Zane and Diago exhibited reluctance to incorporate ChatGPT, as it did not seem to understand their idea of adapting the story of Back to the Future, the decided to independently create the narrative, subsequently employing DALL·E exclusively for visual aspects of the story (3a). Contrarily, Jimin and Alex commenced their creative progression with ChatGPT, then leveraged the resultant narrative as a foundation for generating prompts for DALL·E (3b). Their choices resulted in distinct outcomes—the first being a concise narrative with text more akin to prompts and the second a more sprawling composition, with 12 pages of text and visuals (Figure 3).

Children developed personalized mental models emphasizing GenAI tools’ potential limitations for creative use which were closely associated with their experience with the creative domain. Certain tools were perceived as more useful, emphasizing medium preferences in children’s creative processes, which was heightened when they felt the tool did not understand as much about their preferences as they did. Children’s suggestions highlighted the need for customization, aligning system language with users’ domain knowledge. This suggests that supporting children’s creative experiences with GenAI requires a balance between customizable language, code-switching, and tool adaptability to individual preferences and domain expertise.

Children’s perception of GenAI for creative works was significantly influenced by the framing of the creative task, indicating a nuanced relationship between the nature of the task and their ability to express creative intentions. Notably, when children could align the task with their creative interests, there was a discernible shift from information-seeking to expressing their ideas.

This transition was evident during sessions, such as DS3, where children directed prompts more closely aligned with their creative interests. This was exhibited when Diago, Zane, and Matt wrote two poems, one about Minecraft and one about Zelda. After the poems were generated, they asked ChatGPT to combine them, in which Diago noted after that "I feel like it didn’t really combine them." The group discussed what they wanted to do, and subsequent interactions involved adjustments to the poem’s length, feedback on the generality of the poem, suggestions to incorporate elements related to Bokoblins (a monster from Nintendo’s Zelda franchise), and attempts to prompt the system to more equally address Minecraft and Zelda. During this exchange, Diago also reflected back remembering that he had asked earlier about "Kirbo" (an online spoof of the Nintendo character "Kirby") and noticed ChatGPT had placed Kirbo into the poem. This prompted him to ponder, "why [ChatGPT] put the Kirbo in if you just asked for Minecraft and Zelda… I wonder if it can’t forget after sometimes.” These interactions ultimately lead to more exploration and culminated in a version of Rick Astley’s Never Gonna Give You Up as seen in Figure 4. In this case, Kibro was also further referenced in the bridge. This highlights that when children perceived GenAI systems as having a capacity for learning and adaptation through interactions, they developed a more nuanced understanding of the system’s capabilities, influencing their future creative experiences with the tool.

In addition to personal creative intentions, ethical considerations also played a role in how children engaged with GenAI tools. When asked about when it is good or bad to use GenAI to help make something, Zia and Diago noticed that ChatGPT could write essays about anything, prompting Jimin to suggest that tools like ChatGPT made it too easy to write essays and GenAI tools might actually be harmful since “you can’t trust [students] to…use it properly” (DS1). Cyno, Diago, and Alex all suggested that cheating might also be a reason why their schools had banned the use of ChatGPT.

When engaging with situations involving personal creative endeavors, a more introspective dimension surfaces as a determinant for their ethical considerations. When prompted about their opinions about a friend utilizing ChatGPT to compose a birthday card for them, responses were contingent upon contextual nuance. A prevalent sentiment among the children was one of sadness and disappointment, rooted in the perception that the utilization of GenAI undermined the authenticity and effort invested in crafting a heartfelt communication. Conversely, Diago articulated that the use of ChatGPT could be acceptable, provided the AI-generated content served as a scaffold upon which the friend constructed a more elaborate and personalized message (DS2). Similarly, when discussing how they might feel if their favorite artists or books were written completely with GenAI, children often also qualified how the AI would be used in creating, stipulating that as long as it did not make the whole work of art, it was possibly fine. An assertion from Damian encapsulated this sentiment as he expressed that a completely AI-authored book would erode the sense of author-reader connection, describing that an AI-authored book would “dismantle" some of the joy of reading. Zane agreed that he would not like if AI were writing the stories he read (DS5). Paradoxically, a counterargument posited by Diago again, emphasized that an AI-authored book might have more information, so he was not sure if he cared if it was written by GenAI.

The creative intention, encompassing the creative environment and the child’s expression of their goals, significantly influences how they perceive and navigate the system. Using GenAI tools for personal expression prompts questions about authenticity and effort, as children recognize the potential for these tools to affect the genuine sentiment behind their creative endeavors. Ethical considerations also come to the forefront, especially in formal educational settings, where concerns about misuse, cheating, and a perceived responsibility for proper use are often considered. These discussions highlight the multifaceted nature of children’s engagement with GenAI, emphasizing the need for nuanced conversations and considerations in both educational and personal creative contexts.

We developed a model (Figure 5) based on our results that contains three layers. In the model’s center, “mini-c moments” are where children derive meaning from creative experiences, such as expressing satisfaction during end-of-session discussions, or demonstrated a moment of learning about a tool. The supporting contexts layer identifies four key contexts, viewed through the lens of context by Duranti and Goodwin [34], such that the focal event (e.g., mini-c moment) is dependent upon and interpreted through the phenomena of the context around it. Arrows indicate the dynamic relationships between these contexts and mini-c moments. Scaffolds around these contexts signify observed actions enhancing these contexts during design sessions. Each section contributes like a "slice of pie," emphasizing main takeaways from our findings.

We believe that developing GenAI systems and educational resources that scaffold these supporting contexts can help children clarify the role GenAI has on creativity and help children construct creative self-efficacy while using these systems via supporting moments of “mini-c.”