Teachers, Parents, and Students' perspectives on Integrating Generative AI into Elementary Literacy Education

The viral launch of new generative AI (GAI) systems, such as ChatGPT and Text-to-Image (TTL) generators, sparked questions about how they can be effectively incorporated into writing education. However, it is still unclear how teachers, parents, and students perceive and suspect GAI systems in elementary school settings. We conducted a workshop with twelve families (parent-child dyads) with children ages 8-12 and interviewed sixteen teachers in order to understand each stakeholder’s perspectives and opinions on GAI systems for learning and teaching writing. We found that the GAI systems could be beneficial in generating adaptable teaching materials for teachers, enhancing ideation, and providing students with personalized, timely feedback. However, there are concerns over authorship, students’ agency in learning, and uncertainty concerning bias and misinformation. In this article, we discuss design strategies to mitigate these constraints by implementing an adults-oversight system, balancing AI-role allocation, and facilitating customization to enhance students’ agency over writing projects.


INTRODUCTION
In early January 2023, The New York Education Department announced a ban on using generative AI chatbots (ChatGPT) in school districts' networks and devices over concerns about potential misuse and safety [90].By May of that year, however, the department dropped the ban, announcing plans to explore whether there were potential possibilities to use the technology in the classroom [60,91].When new technology is introduced in educational settings, perceptions often swing between excessive optimism and skepticism, largely due to the uncertainty surrounding the actual usage of these systems in real-world scenarios [27,87].The ongoing discourse in education around generative AI (GAI) emphasizes the need for comprehensive research into its integration within educational contexts [98].
GAI, also known as Generative Adversarial Networks (GANs), GAI systems have gained signifcant attention within the HCI community [75,109,124].The advances of generative AI (i.e., ChatGPT, Dall.e 2, Midjourney) open up a new horizon of open-context conversation with an AI chatbot [4,20,74,78], including generating novel outputs-such as images, text, music, or video-based on patterns it learned from large datasets during its training [15].The HCI research community has started to examine utilities and interaction techniques with these systems [63,116], focusing on new interaction styles [52,125], Large Language Models' (LLM) capacities [63], and how to adapt the systems to creative activities for adults [43,68].While the advancements in GAI have captivated the HCI community with their ability to foster novel forms of open-context interaction, applying these technologies in educational settings, especially for elementary school students, presents a diferent set of challenges and opportunities.
Technology integration in education requires understanding practical realities rather than relying solely on technological advancements, which call for balanced approaches that recognize the complexities of teaching and learning [87].Recognizing the role of storytelling in child development [53] and its impact on critical skills like imagination and comprehension [69], it becomes clear that integrating such advanced technologies in education demands a careful balance.This approach should respect both the potential of GAI and the intricate nature of teaching and learning processes, ensuring that technological advancements are meaningfully and efectively aligned with educational needs and realities.Considering the need to underscore the applicability of leveraging GAI in writing instruction for students, we conducted a study to examine the diferent perspectives of stakeholders in K-6 education (i.e., teachers, parents, and students) regarding the integration of GAI in elementary school literacy education.Our objective is to understand stakeholders' aspirations and concerns regarding the use of new systems in academic settings in a holistic manner by including both teachers and learners so that the HCI research community can use these insights to design and develop GAI-powered educational applications that are safe and productive for elementary school students writing.
In this study, we sought to answer the following questions: • How do stakeholders in elementary school settings-parents, teachers, and students-perceive AI to support teaching and learning writing projects, and what are their opinions of the potential benefts and limitations of leveraging it?How do values and motivations towards GAI systems difer among stakeholders in education?
• In what ways can GAI systems be designed so that they are efective, engaging, and safe for teaching literacy for 2nd to 6th graders?
To answer these questions, we conducted workshops with families with children ages 8-12 (i.e., in 2nd through 6th grade) that included semi-structured interviews with students and parents during and after the workshop.Also, we carried out 1:1 semi-structured interviews with 16 teachers to better understand teachers' motivations, perspectives, and strategies for leveraging GAI in writing projects.In total, we report on insights from 40 participants who present unique perspectives on GAI from three groups of stakeholders in education (i.e., 16 teachers, 12 parents, and 12 students).
From the study, stakeholders' perceptions towards GAI systems and their opinions of potential benefts and challenges related to writing surfaced three major themes: 1) teachers' view as a part of digital citizenship development, 2) parents' perception of new types of toys, games, and screen time, and 3) students' perceptions as smart and helpful companions.In addition to these major themes, we highlight possible obstacles and concerns regarding authorship and ownership issues over writing outputs, challenges examining students' agency in learning, and difculties in controlling bias and hallucinated content created by GAI systems.Based on the fndings, we provide design implications to mitigate the shortcomings of these systems in educational settings.This discussion includes: 1) navigating the complexity of authorship in AI-assisted writing systems through examining a child-AI interaction chatlog, 2) enhancing student agency through role allocation and curating AI personas in GAI systems to promote independent writing and cultivating conversations aimed at fostering students' unique voices, and 3) balancing fexibility and control with teacher-in-the-loop GAI-LLM systems that allow teachers to curate child-AI interaction.
We aim to contribute to the HCI community by highlighting the practical applications and limitations of GAI in education and by ofering insights that can guide the design and implementation of GAI tools in a way that aligns with the needs and concerns of various educational stakeholders.
Two main contributions are made by this work: • Our study provides a qualitative investigation of the efcacy of generative AI for writing projects, surfacing potential benefts and challenges in using LLM-driven chatbots in educational settings.Our fndings demonstrate that GAI systems ofer opportunities for creating adaptive teaching materials tailored to students' unique competencies in writing, broaden ideation and timely interaction through dynamically generated learning resources, and provide individual, culturally relevant feedback.At the same time, using GAI systems in writing carries signifcant limitations regarding authorship, agency, and potential misinformation.
• We present design implications by investigating ways to harness generative AI in writing projects safely and efectively.We surface the challenges and difculties from stakeholders' perspectives and provide insight into designing new systems.We propose design suggestions to enhance safety by balancing fexibility and control through teacher-in-the-loop systems where teachers can prompt to curate AI agent capacity with prompt bank interfaces, designing the AI agent persona as coach or/peer rather than an assistant, and designing role-allocation among AI and students of which students have the freedom to write independently, edit, customize themselves instead of having the AI agent generate on their behalf.

RELATED WORKS
In this section, we examine research literature related to the implications of artifcial intelligence for education in HCI research, as well as educational research related to artifcial intelligence applications for learning and teaching in educational settings.

Tracing the Evolution of Technology in Education: Implications for Modern AI Integration
Refecting on the past usage and integration of new educational technologies in real-world educational settings can ofer valuable insights for predicting and enhancing their efectiveness in learning environments [87].To contextualize our investigation of the potential applications and benefts of emerging GAI systems in educational contexts, we trace the impact of Massive Open Online Courses (MOOCs) and Intelligent Tutoring Systems (ITS).These technologies have been pivotal developments in the history of scalable learning with implications for the educational sector.Despite rapid technological advancements, the anticipated radical transformation in education by innovative educational technology companies (e.g., Khan Academy, Udacity) has largely fallen short of expectations.Personalized learning platforms claim to tailor education to individual student needs, but they often fall short in practice due to the complexities of learning processes, efective pedagogies, and the constraints of algorithmic customization [87].Therefore, Reich 2020 argues that educational innovations must be deeply rooted in the realities of teaching and learning.
Reich [87]'s four dilemmas highlight the complexities of learning at scale platforms, emphasizing the need for a critical reassessment in the context of emerging Generative AI (GAI) technologies.These dilemmas include the preference for familiar tools, the unequal benefts of new technologies, the challenge of nuanced assessment beyond binary right or wrong answers, and the issues of data privacy and equity [66].As GAI systems ofer more natural and adaptable human-AI interactions, they present an opportunity to address these challenges, making AI-based educational tools more accessible and equitable for diverse learners.
Recent advancements in Large Language Models (LLMs) enhance their ability to assess human reasoning in writing, moving beyond the traditional right-or-wrong evaluation methods of current Intelligent Tutoring Systems (ITS).This progress ofers a more nuanced understanding of student logic and thinking, enabling personalized and adaptive feedback.Studies, such as those by Steiss et al. [98], are beginning to explore GAI's potential in analyzing and understanding the nuances of students' written work and reasoning processes, which pose potential capabilities to integrate algorithmic guided instructions fexibly.

Integrating Artificial Intelligence in Education.
The use of artifcial intelligence in education (AIED) has been explored through the application of intelligent tutoring systems, conversational agents (CA), and chatbots.These technologies have enhanced teaching and learning [11,21,55,76,84,101,117,121], yet little of this prior work addresses directly how AIED integrates holistically into educational settings [21,54].An exception to this is Chiu et al. [22] systematic review of AI's roles in education, which surfaces potential benefts of AI for learning, including providing adaptive learning by assigning tasks based on individual abilities that enhance academic performance and facilitating human-machine conversation to motivate and engage students.However, Chiu.[22] pointed out the need for further studies that examine students' educational outcomes with AI-based systems (such as chatbots or conversational AI).
The HCI community has also provided insights into the perception of AI systems [106] among educational stakeholders, including teachers [65,85], children [6,16,115,118,122], and parents [39,40,105,120].To design AI tools and curriculums that align with the values and contexts of stakeholders in education (i.e., teachers, parents, and students), Lin and Brummelen.[65] conducted co-design workshops with K-12 teachers to develop design recommendations for creating AI curriculums and tools aligned with teachers' needs.Their fndings revealed how teachers value learning outcomes, student engagement, ease of use, and collaboration when incorporating AI in the classroom.Design recommendations from the study emphasize the importance of designing AI tools to be adaptable to diverse contexts (e.g., diferent grades and subjects).
Outside the classroom, parents see technology (including AI) as a way to enhance parent-child interactions by selecting content for their children, showing a preference for customized content [127].Children's views on AI agents difer based on age and their performance in AI experience and interaction.Younger children often perceive AI agents as intelligent toys, while their older counterparts perceive them more as humanoid entities with lesser intelligence [115].Additionally, Xu and Warschauer.[122] reported that most children view conversational agents (CAs) as having cognitive capabilities via continuous communication but possess fewer psychological entities (i.e., having emotion).The fndings suggest possibilities of designing CA as a learning companion, incorporating social interaction and emotional feedback [122].
Despite this body of recent research, there is still a lack of clarity regarding the role of artifcial intelligence (including generative AI) from educators, parents, and students' standpoints.Additionally, further research is needed to investigate whether and how these emerging technologies can improve the learning process of literacy development in elementary school settings.

Emerging
Trends and Challenges in Generative AI Applications for Education.The rapid advancement of GAI, such as large language models (LLMs) and Text-to-Image (TTI), learn patterns and structures from existing data and generate new content [113].These breakthroughs have led to a new generation of dialog systems that enable the possibility of leveraging the system to facilitate openended discussion and generate educational content for teaching and learning [79].Ahmad et al. [2] examined the implications of ChatGPT in the education sector, emphasizing the need to develop skills for using LLMs and GAI to be prepared for future job markets.This requires students to know how to prompt AI systems efectively and to be able to analyze the quality, originality, and accuracy of the results [2].
Research on AI systems in literacy education (reading and writing) focuses on LLM-based chatbots for language learning [1,128], scientifc writing [42], creative writing [24,97,126], and creative storytelling [50,127].For example, Gero et al. [42] studied how LLM-powered co-writing platforms can enhance engagement and idea generation with STEM graduate students.Yuan et al. [126] studied adult hobbyist writers' sense of ownership over AI-assisted writing and found that AI integration does not undermine writers' feeling of ownership because writers use AI-generated text as an inspiration rather than taking it verbatim.Lee at al. [63] also conducted studies with adult participants to understand the afordance of large language models (LMs).The authors aimed to guide the design of LLM applications and developed a CoAuthor system, which focuses on capturing and analyzing user engagement data.This system tracks how users collaborate and construct stories, providing valuable information on user interactions and narrative development within the context of LLM applications.The fndings showed that CoAuthor enhances writing productivity, increasing the text writers produce.But Yuan et al. [126] and Lee et al. [63] also raised questions about writer's feeling of ownership over their writing outputs and indicated the results were uncertain.
Recent GAI-powered educational applications ofer potential opportunities to leverage GAI systems in teaching and learning (GPT-3, TTL) [17].For example, Speak [5] uses GAI systems (GPT-3) to simulate smooth verbal conversation with learners to improve English speaking profciency without age limit.Also, web applications and conversational agents (CA) have been developed to support students' reading comprehension through story creation (i.e., Wanderly, OnceUponABot, AlexaBedtimeStory) mainly for families with children ages 5-12 [14,71,96].MagicSchool.ai[3] is a web application that uses GAI systems to support efcient lesson plans for teachers by suggesting and generating quizzes and scafolded lesson materials.Khan Academy recently launched an LLM-based AI agent, Khanmigo, that carries a text-based conversation with students as a tutor, as well as facilitating teachers' versions as teaching assistants, which assist teachers in creating lesson plans for a wide range of subjects (history, language arts, math, foreign language) across K-12 [61].
However, it is still unclear how these new interactions, user experiences, and learning engagements afect learning outcomes [10].The current story creation apps powered by GAI systems produce whole stories for students, which raises a question about whether it could promote language learning or undermine creativity [50].Hence, further research is needed to ensure GAI-powered learning tools are efective and age-appropriate.

Child-AI interactive systems
Nowadays, an increasing number of children interact with AIenhanced products daily.Researchers have explored the perspectives of various stakeholders, including teachers, parents, and children.Findings reveal that parents desire CA to foster children's social engagement and involve parents in in-home learning [41].However, researchers raised concerns about the lack of open-ended and extended back-and-forth dialogue while considering CA to support children's language development [119].Enhancements are also needed for human-AI collaboration to relieve the repair burden on families during their communication breakdowns with CA [11].As for children's perspective, research eforts have been made to investigate children's perception of their data utilized online [110], children's autonomy over the technology [111], and AI technologies' infuence on child development.With the recent advancement of GAI, such as LLM, daily life AI-enhanced products have largely extended their power of human-AI collaboration, including children-AI co-creation.This is also aligned with the rising desire for AI literacy education outside of the computing domains [94], and in turn, challenges AI literacy education by equipping children with some basic AI literacy in both classroom and family scenarios [33,99].These all require a deeper understanding of stakeholders' needs and concerns around child-AI co-creation.
Existing child-AI co-creative systems encompass interactive storytelling [126], creative writing [35], and drawing [127].Wordcraft [126] is a text editor fostering collaborative engagement between users and LLM in storytelling.It facilitates open-ended conversations related to the narrative, responds to users' natural language queries, and ofers suggestions to assist writers in overcoming creative hurdles.The study with adult participants suggests incorporating real-time requests and predefned controls to amplify the co-creative experience.In the intersection of drawing and creative storytelling, "StoryDrawer" aims to support children in creating oral stories during visually immersive storytelling episodes [127].Results from the evaluation with children highlight the importance of encouraging collaboration and co-creation between children and the AI system rather than solely relying on the system to generate stories.CreativeBot is a robot designed to stimulate children's creativity through co-creative storytelling [35].The robot's ability to generate unexpected and surprising story elements proved particularly efective.Findings imply fexibility, adaptability, collaboration, and surprise as crucial factors for the CreativeBot.Besides such conversational, drawing, or robotic interactions, researchers have developed diferent LLMs as supports for collaborative creative writing [77,100], where creativity requires writing with a relevant purpose, understanding, judgment, and evaluative abilities in ways that are deemed original and valuable to a community [26].However, by this defnition, by relying primarily on summation, LLMs lack the intention to write and do not possess the self-feedback loop necessary to intentionally deviate from conventions, hindering their capacity [38].Therefore, specifc interface elements need to be designed to compensate for such limitations of LLMs.Beyond such inspection from the technical perspective, research is needed to develop a more in-depth understanding of children's, parents', and teachers' needs and concerns around child-AI co-creative systems.

METHODOLOGY
In the previous sections, we outlined the adoption of AI technologies in education and addressed a research gap resulting from a lack of holistic understanding of teachers, learners, and their caregivers' opinions.Considering potential stakeholders' perceptions of GAI systems may provide design implications to help guide the development of GAI tools and systems for elementary school students.To elicit stakeholders' perspectives on potential possibilities and limitations of a GAI-LLM chatbot system for writing, we conducted a workshop with families with children ages 8 to 12 (parents, N=12 and children N=12) that focused on how they used a text-to-image generator (i.e., Stable Difusion [31]) and a chatbot powered by LLM (i.e., ChatGPT [19]).Following the workshop, we conducted 1:1 interviews with teachers who specialized in teaching writing in elementary school settings (N=16).In total, we reported on insights from 40 participants who interacted with both tools.Participants were recruited from our researcher's network (mailing list and contacts) and snowball sampling.We sought to identify teachers, parents, and students' motivations, challenges, and opinions with the new systems, elicit their concerns, and identify their perceptions and strategies in writing using GAI platforms.

Study Procedure
3.1.1Workshop with Families.In April 2023, we conducted a workshop with families with children ages 8 to 12 (2nd and 6th graders) (Table 2) in order to better understand students' strategies and struggles when interacting with the current state of LLM-based chatbots and text-to-image generators.Their parents' and guardians' opinions and perceptions regarding using the systems for writing projects were also considered.We focus on the 8 to 12 age group, recognizing the critical importance of this phase in developing reading and writing skills [48].This period is pivotal as children transition from learning to reading to reading to learning, a fundamental shift highlighted in Loveless's 2023 article.Early interventions during this stage can greatly infuence a child's educational path and future opportunities [47].Given this, our study aims to investigate how enhanced engagement with writing and literacy activities facilitated by GAI platforms can positively or negatively impact learning in these formative years.Parent participants (n=12) completed a screening survey before the workshop to ensure they were 18 or older and lived with children ages between 8 to 12 years old.The average age of parent participants was 39.8 years old at the time of the workshop, of whom (10/12) were female and (2/12) were male.According to parent reports, the mean age of the student participants was 9.8 years old, and (5/12) were girls.Eleven children (11/12) were identifed as Asian American, four children (4/12) spoke only English at home, and the remainder were bilingual (6/12) or spoke English as a second language (2/12).All children possessed sufcient oral English profciency for daily conversation.The median household income of the twelve families is $118,749, with a range from a minimum of $29,999 to a maximum of $200,000.Given the socioeconomic standards of the West Coast, USA, this income bracket is typically classifed as upper-middle class [108].It was the frst time the students had used GAI-LLM chatbot and Text-to-Image generators (TTL), while seven parents (7/12) reported already using them.Family participants were compensated $25 for their time and efort.
The 2-hour, 1-day workshop was conducted in a library in a Southern California metropolitan city.Accompanied by their parents, children were required to create a visual story using a textto-image generator (i.e., Stable Difusion) and a chatbot powered by LLM (i.e., ChatGPT).During the writing project, we sought to understand the students' strategies and their interactions with the system through observation by taking feld notes and voice recording youths' verbal expressions and semi-structured interviews [88].Given the California State Standards in elementary literacy education, we chose narrative writing activities for students [46] instead of giving students a specifc topic to write about; students wrote creatively without limitations.The topic of the visual story was open-ended, and students picked a topic based on their own interests.To assist, several prompt examples were provided (e.g., "I would like to write a topic of the story, how can I start?", "Can you list fve story ideas?") before they began writing.Students worked individually without their parents' intervention unless they needed to access a required platform (i.e., Google Classroom, Google Docs).Students used the systems under the supervision of researchers.
We created a Google Classroom for the workshop that served as an information resource as well as a repository for participants' fnished visual stories.Students were allowed to use the Text-to-Image generator and LLM chatbot to develop their stories.One of the researchers ran the workshop, and the other researcher observed, took feld notes and conducted semi-structured interviews with children during and after the workshop.While students worked on generating their visual story, one of the researchers conducted semistructured interviews with the parents.Interviews were recorded using a voice recorder, no videos were taken during the workshop, but pictures were taken, and students' artifacts were collected.We sought to understand students' opinions and their perceptions of AI by posing the following questions [88]: What do you like or dislike about using ChatGPT and Stable Difusion for your creative writing and visuals?, Have you found AI useful?, How can artifcial intelligence help you?To understand parent's opinions and their perspectives on using the systems for their children, one of the researchers conducted semi-structured interviews with the parents while students worked on creating visual stories.With parents, we discussed the following topics: How do you think AI impacts your child's learning?, Do you want your child to use AI or learn about AI?, What is your overall impression of using AI for your child?Interviews were recorded using a voice recorder, no videos were taken during the workshop, but pictures were taken and students' artifacts were collected.
3.1.2Teachers' interviews.Teacher interview data collection was conducted online between June to August 2023.Teachers were recruited using similar snowball recruitment eforts as the families, with the only criteria for eligibility being that they were either current or former K-12 teachers.The teachers we interviewed (n = 16) were elementary classroom teachers from 1st to 7th grades, most of whom (14/16) work in public schools.Thirteen teachers (13/16) specialized in teaching writing and were afliated with the National Writing Project (NWP) network.Teaching experience averaged 13.3 years (min=1.7 years, max=32 years).More detailed participant information can be found in Table 3.The majority of the teachers (8/16) are located in the United States (California and Pennsylvania), and four of them are in Asia (South Korea and China).
The majority of the teachers (14/16) work in public schools, with only two working at private schools (see Table 3).The teacher interviews were conducted individually for up to an hour via video conferencing due to geographical distances, with an average length of approximately one hour.We sought to elicit their current teaching practices, struggles, and motivations when teaching writing to their students.Afterward, we introduced GAI systems (i.e., features and functionalities) and asked about their  (Midjourney), whereas the rest were unfamiliar with these systems.Teachers who have used GAI systems continue to use it in their teaching practices since they frst tried, and their years of teaching experience are averaged at 5.8 years, compared with 22.5 years for teachers who have never used GAI systems.
In the interviews, the following topics were discussed: • The interviewee's general practices, difculties in teaching, and concerns for their students (e.g., "What is the hardest part in teaching writing in your class?"), • Their experiences the state-of-the-art GAI systems (i.e., Chat-GPT, Stable Difusion) (e.g., "What is your level of familiarity with Generative AI systems like ChatGPT and Stable Difusion?"Have you ever used or willing to use the GAI systems in your class or for yourself?"), • Their opinions of their intended usage of the GAI systems, and their opinions and concerns about them (e.g., "Can you tell me your thoughts about the GAI systems as students use them for writing?", "Can you share your opinions on whether or not GAI systems are benefcial or harmful for students?","How do you envision these systems being used by teachers or students?")A recording of all interviews was conducted with the consent of the participants, and teacher participants were compensated $25 for their time and efort.Our study was approved by the authors' institutions' institutional review boards (IRBs).

Data Analysis
The interview data was frst transcribed using an automatic transcription program (Otter.ai) that maintained the original audio and aligned it with the transcript.After thoroughly reviewing the transcript, we transferred the transcript to a qualitative data analysis software (Atlas.ti)ensuring that the original audio was preserved and accurately aligned with the transcripts.Following this, we utilized qualitative data analysis software for an initial round of open coding, adhering to established qualitative research methodologies [93,103].We conducted an inductive approach to analyze interview data [102].Following the inductive approach, two researchers independently read the transcripts and identifed key themes and patterns within the text.This collaborative and iterative process of theme identifcation and analysis was instrumental in reaching theoretical saturation [70].Each researcher assigned the frst round of low-level codes guided by our research questions (e.g., participants' opinions (stance) of the potential benefts and limitations of leveraging GAI; how their values and motivations difer) into each theme.In order to reduce overlap between themes, we repeated discussions with researchers.We categorized the low-level codes into higher-level themes.The researchers regularly discussed (every week for two months for an hour each) and iterated to construct the themes.By systematically coding the data and constantly comparing emerging themes, we were able to ascertain when no new themes were emerging from the data, indicating that theoretical saturation had been achieved.We organized our results around the main theme of the advantages and challenges of using LLM chatbots for educational purposes in K-6 settings, which emerged from this coding.We categorized codes into four high-level themes (i.e., perception, positive opinions, negative opinions, and suggestions).The analysis contained nine mid-level themes (i.e., teachers' perception of digital literacy development, parents' perception of toys and games, students' perception as helpful companions, creating adaptive teaching content, timely interaction and broadening ideation, personalized and culturally relevant feedback, lack of context for students, problems with authenticity and authorship, hard to distinguish students' agency, difcult to control biased and misinformation) and 34 codes under each theme.

Limitations
Our study focused on the context of educators and families in one of the metropolitan cities on the West Coast, United States, as well as mid-high socioeconomic families.It is possible that our fndings do not represent the perspectives of all populations on LLM-based education chatbots for writing.Additionally, the majority of families in the study were multilingual, primarily Asian-immigrated families (7/12) whose children were born on the West Coast of the United States and attended public schools.Since our samples lack a diverse cultural background, some of their perspectives and opinions might be limited.The majority of parent participants were mothers (11/12), and eight mothers (8/12) were stay-at-home with an average age (of 39 years old); hence, their views and opinions from the interviews are hard to represent all parents' perspectives towards GAI systems for their children's writing project.Additionally, considering the majority of teachers we interviewed are from high-SES school districts, their teaching practices, motivations, and concerns are likely to difer from those of other teachers, so generalizing their views is problematic.A future study should also consider interviewing school district administrators, whose voices are central to systemwide policy decisions.
Additionally, during the workshop, we missed the opportunity to collect chat logs to investigate students' interaction techniques with a chatbot.Similarly, while we reviewed the fnal output of the students' writing pieces, it would have been better to check the history of their editions in Google Docs in order to understand their contribution to the writing better, whether they simply copied and pasted from AI-generated text, or how much they wrote by themselves.An analysis of the student's perception of ownership and the actual percentage of contribution to the piece would be valuable, as well.It may be worthwhile to investigate in the future if there are diferent ways to assess and measure students' learning in AI-students co-writing projects in the classroom.

FINDINGS
By analyzing qualitative interviews and observational notes, we uncovered multiple perspectives regarding the use of GAI in literacy education.In this section, we report major fndings regarding our participants' opinions and experiences with GAI.We outline the values and perceptions of multiple stakeholders (see Figure 1), then elaborate on the fndings in the advantages and constraints of GAI for literacy education (see Figure 2).The fndings are categorized by each stakeholder's viewpoint to highlight how their values and perspectives difer.Following that, we categorize the themes into teaching and learning and integrated stakeholders' opinions, as stakeholders often have insight into other stakeholder perspectives (e.g., teachers' perspectives on students; and parents' perspectives on their children).
We report major themes in our stakeholders' perspectives and opinions about using GAI in literacy education, particularly teaching and learning writing.GAI is perceived diferently by each stakeholder, including 1) teachers' view as a part of digital citizenship development, 2) parents' perception as new types of toys, games, and screen time, and 3) students' perceptions as smart and helpful companions.

Multifaceted Views on the Role of GAI in
Literacy Education Nine teacher participants (9/16) expressed willingness to promote the use of GAI to foster safer and healthier ways of using the systems.Specifcally, T3 noted: "I do think that instead of rejecting it, we need to fgure out how it works for us and what we need to do with it.I mean, our students are going to be using it, our co-workers are going to be using it, right?It's going to be in the world.So I do think we're better of to fgure it out than to reject it for sure." While over half of teachers tried to embrace the GAI systems into their practices, (7/16) considered them as an essential part of the digital citizenship development for both teachers and students, agreeing to teach students about GAI systems as another tool that they will need to learn how to use.
Teachers pointed out that GAI systems can also be used to support educational processes [23], nine respondents (9/16) emphasized that GAI systems like ChatGPT and Text-to-Image generators can be integrated into their instructional processes: "I think it has a lot of potential.I think there's lots of excitement for potential teachers in lesson planning.I don't think it's kind of replacing any existing curricula.But I think it can be a tool to extend the teaching as a part of the process." For instance, one respondent noted that the current GAI-LLM chatbot lacks the capacity to be fully integrated into human conversations but can be useful for brainstorming ideas: "I don't think Al has been adapted to fully understanding or answering questions yet, but I have used it a ton as a student and a professional to brainstorm ideas.It's like a friend with a wealth of information, like someone I bounce back ideas from." Our fndings indicate that teachers are willing to integrate new systems (GAI) into their teaching pipeline along with digital literacy development.In addition, they stressed the importance of equipping their students with the ability to use GAI systems to develop their digital citizenship.
4.1.2Parental Caution: Atitudes Toward GAI Systems in Children's Literacy Education.On the whole, parents expressed more conservative attitudes, with (11/12) of respondents expressing skepticism about the use of GAI systems in their children's education.In spite of the fact that all participants in the parents' interview (12/12) agreed that AI will be a part of their children's lives as they grow, it is still important to know how to use it properly.For parents of children ages 8 to 12 years old, it is more important for their children to learn how to use GAIs responsibly and safely, which makes them more cautious about potential harm.
Seven parents expressed concern over uncertainty and data privacy when their children played games or watched videos with real-time chats with anonymous strangers on the internet; they found AI such as Alexa or Google Play to be safer.According to P01, "My kids also play with Anonymous.I'm so worried because of the anonymous player, we don't know if the person is good or bad.So, if my kids are going to play with anonymous players, I would choose to play with AI because I think AI is at least safer than those harmful people." Also, we identifed a confict between their values and their perception of GAI systems.It is important for parents to prioritize their children's overall well-being and well-rounded development (i.e., soft skills, emotional, physical, and intellectual), not just hard skills and academic success (i.e., test scores and grades).Eight parents (8/12) emphasized their focus on literacy education and their willingness to support it through child-centered approaches and interest-driven experiences (e.g., purchasing books their children are interested in reading).However, these parents perceive GAI systems like ChatGPT and Text-to-image generators (TTL) as other types of games and toys that will increase their children's screen time.P1 said: "I mean, for kids, ChatGPT and Stable difusion are just another type of toys.It's like they play Roblox or Minecraft or AI graphics." There also appeared to be a generation gap between parents and children over AI perception, mirroring the lack of confdence for parents to introduce new technology to their children that has existed for decades [86].Most parents (8/12) perceived the GAI systems as new to them, so they had difculty imagining how it would afect young minds.For example, P03 and P04 mentioned: "I have no idea.Because I don't know AI exactly, Because I didn't learn it when we were young, it's hard to say it's unnecessary because we don't know it well.That's the problem.So the parents like us from the generation that we don't even have AI." While such expressions of distrust are rooted in a lack of knowledge and experience, some parents identifed that learning the new system with their kids could serve as a learning opportunity for them both.P08 highlights, "So things are maybe an opportunity for parents to learn with a kid at some time.Okay, so they get to know what AI is like and how to use AI." As such, even though all parents acknowledged that their children need to learn how to use GAI systems properly, most parents prioritized promoting critical thinking and problem-solving instead of introducing GAI systems to their children.Moreover, parents (n=11, mothers) presented anxiety over adapting GAI systems for their children's writing projects, which could limit their children's creative thinking.Hence, they were curious about fnding a way to leverage GAI systems for themselves as adults and using it for their children instead of directly giving them to their kids (i.e., creating word quizzes for their children).

Creative Allies with Caveats:
Students' Mixed Perceptions of GAI Systems in Literacy Projects.For students, data from the workshop revealed that they (9/12) regard chatbots and TTLs as creative, smart, and helpful companions in the process of creative visual story writing, as S1 mentioned: "I initially thought that artifcial intelligence wouldn't be able to do creative things because it doesn't have a brain or mind, but it turned out more diverse and creative than I expected, which surprised me." The vast majority of students (11/12) were optimistic about using the GAI in the process of creative writing, with (10/12) of students pointing out the efciency of using the GAI-LLM chatbot and TTL generator to enable rapid prototypes, which broadened their choice of ideations.S7 highlighted, "I can use this to test out as many as my ideas.I think it's really efcient." We observed two primary difculties encountered by students when they started the systems: 1) initial user prompts and 2) defcient AI responses.Many students had difculty fguring out what to do due to the blank interfaces and lack of instruction and context on the website.Once we provided guidance on how to start (i.e., an example prompt included "Can you generate fve story ideas for a children's book?"), they began testing them and learning how to use the system.Half (6/12) of the students also complained at times that GAI had not generated the content they intended.As a result, we concluded that instructing and teaching prompt writing would enhance efciency and adaptability [67].Second, we found that the randomness of the output generated by GAI systems can be a double-edged sword.Despite the possibility of unexpected, sometimes inappropriate results (e.g., generating a dead animal), Seven students (7/12) saw these moments as chances to expand their ideation, as they are likely to view even unexpected outcomes as part of the divergent process of their conception.

Delineating Advantages: GAI's Contributions to Literacy Education
To elaborate on the fndings about the advantages of GAI in literacy education, we categorized the themes from our interviews and observations into teaching and learning aspects.In each section, all stakeholders' perspectives are incorporated since stakeholder perspectives represent other stakeholder perspectives (parents concerned about their kids' privacy, teachers' views about their students).Findings demonstrate that the advantages in teaching include enhancing efciency in teaching by enabling fast and easy construction of scafolded materials and content, including preinstruction (by developing diferent levels of materials tailored to each student's abilities), during instruction (by facilitating questions and quizzes), and post-instruction (by developing a rubric).In terms of how this afects user learning, GAI enables personalized experiences that provide immediate feedback to support the needs of diverse learners (i.e., by facilitating a real-time GAI-powered tutoring system).Further, interacting with GAI encourages students to generate ideas around topics, add details, and apply culturally relevant approaches (see Figure 2).

Enhancing Pedagogical Eficiency: GAI in Crafing Customized
and Scafolded Mentor Texts.The teachers (16/16) all afrmed that GAI systems can be used to create adaptive teaching materials as part of their lesson planning.In particular, the majority of the teachers (13/16) who specialized in writing education highlighted the potential for GAI systems to generate scafolded mentor texts (i.e., texts that model for students what good writers do) that allow students to adapt and learn from the authors' writing style (i.e., words, sentences, or paragraphs).T7 highlighted, "A lot of the craft of writing comes from looking at examples and fnding out what the experts did and using what we've learned in our own pieces.Let's say we've studied this particular sentence deeply, and then we won't just imitate it; we fnd it out on our own and then try it on.Then, the kids change that for themselves.I use a ton of mentor texts." However, nine teachers (9/16) pointed out the difculties of fnding and incorporating mentor texts that can be seamlessly integrated into their curriculum at the appropriate level for all students.T3 mentioned, "Using mentor text is really a lot of teacher work to design it and fgure it out.And what if I could generate mentor sentences and have everything ready to go.I would love that.That is one of the ways that we can use it to help us develop some of the mentor texts that we have spent hours looking for." The elementary school classroom teachers (14/16) stated that their students have diferent literacy levels and interests, so a standardized curriculum makes it hard to tailor learning materials to each student's unique abilities.In response, teachers imagined leveraging GAI systems like ChatGPT to generate scafold vocabularies and sentence levels tailored to each student's unique level.According to T13, "Can I use Generative AI to develop reading materials at diferent levels for kids to read?I would love to be able to put in a topic and get information coming out, such as climate change.What would be even much better if you could layer on phonics?I can now do phonics instruction and help support within the realm of the science of reading.Having such a tool would be a tremendous time-saver, simplifying the lengthy process of sourcing and summarizing appropriate materials for diverse classroom needs." Other teachers emphasized that they can use GAI systems to generate mentor texts because they can evaluate the quality of the texts and ensure the content is accurate.As one instructor pointed out, teachers are able to determine whether the GAI-generated content is appropriate or not.As T6 pointed out, "Recently, I used Generative AI to create a mentor text, saving a lot of time.Since teachers have a solid understanding of the topic, we can verify the facts and integrate them into our teaching process.There's defnite learning potential in this approach." This implies the potential opportunity for teachers to use the GAI systems to generate scafolded mentor texts and teaching materials for diferent levels of students' capacity.

Scaling Individual Atention: GAI in Providing Timely and
Tailored Writing Feedback.Elementary school teachers pointed out their unique challenges as public school teachers.Due to the large number of students in a single class and with only one teacher to deal with the class, teachers pointed out the difculties of providing immediate and helpful feedback that support students in writing.T9 emphasized, "I think providing individual feedback is a really time-consuming thing.It is difcult to individualize education for all subjects." One of the teachers (T1), a director who has specialized in teaching writing in the writing center at one of the California school districts for the past 30 years, stressed the importance of developing ideas and adding details.T1 stated, "I think for me, it seems like the area where kids need the most support is actually generating ideas for writing and adding details.Students might give you a sentence or two and say I'm done.But if teachers or AI ask them to add more details, that could enhance their writing.Such as asking, 'Can you tell me more about this?'-we can encourage them to expand their writing.Students frequently fnd it challenging to elaborate on their own without such guidance." Our fndings suggest that teachers can leverage GAI to provide immediate feedback regarding students' writing progress from ideation, grammar checkers, and adding detail.For example, T4 highlighted, "I would love for AI to be able to do this for my students.Could AI give high-quality feedback on the spot to student writing?So I would love for the AI assistant to say, oh, you only used the word pretty.Is there another way to explain it?Can you provide some examples of your opinions?Can you explain more about your character?. " In this regard, interacting with GAI systems (i.e., LLM chatbot, TTI generators) helps students expand their ideas by enabling rapid prototypes that broaden their options.As one of the students (S10) stated, "Since AI provides many options, I can pick the one I like best.I think it is good for me to come up with more ideas because AI has given me suggestions I never thought of, even when I get unexpected results, which actually makes me think of better ideas.Thanks to AI, I think the process went much faster." According to our fndings, using GAI systems would beneft teachers and students.Teachers can reduce the efort they need to provide individual attention to students, and students will be able to receive feedback on their story creation through GAI systems conversation.

Culturally Inclusive
Pedagogy: GAI's Capabilities for Culturally Relevant Literacy Feedback.The other aspect of using GAI for personalized learning is to provide culturally relevant feedback and ideas [80].In our workshop, one of the parents shared that she used Chat-GPT to generate word problems for her child's home language learning, which was Japanese.As a parent of an immigrant child, she wanted her daughter to remain fuent in her mother language.Also, parents who immigrated from Asia mentioned that they are willing to use GAI systems to create culturally salient fable stories that ft their children's interests.According to P04, "So maybe parents will ask to know something about some traditional stories about their own culture, but they don't have the actual book or the graphic reference, like some Asian stories in China, about maybe a dragon or something, maybe parents will ask, do you know how to draw a Chinese dragon?And AI will say the Chinese dragon looked like a really long snake with some hair on the head.Also, they speak diferent languages.I think language translation will also be another activity, like my kids having Korean friends from Korea.So they want to share some Korean as well." The fnding indicates that the potential advantages of using GAI systems are to help teachers create lessons using culturally relevant materials, such as songs, videos, and images (i.e., traditional stories by countries' traditional holidays).By doing so, teachers can create a more culturally inclusive classroom and foster cross-cultural understanding.Additionally, parents, especially those from multicultural families, could bridge the communication gap between each other and encourage a sense of belonging and a strong family relationship through a better understanding of each other's cultural values.

Navigating the Gray Areas: Challenges and Constraints of GAI in Literacy Education
Our research indicates that GAI systems in academic settings may pose challenges related to academic integrity, such as issues of authorship, authenticity, and originality.Additionally, there are concerns about how these systems may impact student agency and autonomy in writing processes.A notable risk is the potential for GAI systems to generate biased or inaccurate content stemming from their inherent randomness and uncertainty.

Ethical
Qandaries, and Accountability in GAI-LLM Writing systems.Nine teachers (9/16) expressed concerns about introducing GAI systems to their students due to the possibility of afecting the originality of their students' work.To teachers, AI-generated work can be a problem for kids to misrepresent themselves.As T06 stressed, "So it's like, if you are using this as a tool, you're taking this work from somewhere, right?Make it your own and claim it your own.I think that the problem is that you took AI, and you didn't give AI the credit.If you're going to use AI, then that's who should be credited for the work of GPT because there's almost a moral issue for me, looking at Chat GPT.And thinking about where that information comes from." By extension, teachers are anxious about GAI systems because if students use ChatGPT to generate their own work, it could undermine students' reasoning.For instance, T16 emphasized, "I mean, teachers are particularly anxious about maintaining the quality of writing and are worried about students' work ethic and creativity.Additionally, there's a signifcant concern regarding plagiarism and cheating." Some teachers, in response, suggested using ChatGPT rather than generating text as an output for students' writing, asking students questions to promote the students' thought processes.As T12 mentioned, "Here's one thing is, instead of writing the whole next part.It asked me, you know, like, choose your own adventure?Do you want it to be this kind of problem or that kind of problem?What comes next?" As far as implementation plans were concerned, seven teachers (7/16) emphasized the need for the school districts and educators to establish a new framework for adapting GAI systems to students' learning, with (6/12) teachers also pointing out the necessity to establish diferent assessment methods.
The fndings demonstrate the importance of designing the GAI systems to promote students' reasoning by providing students with the opportunity to use their own critical thinking skills and creative solutions.Additionally, educators must develop a new means of assessing and evaluating students' writing projects.For instance, teachers can focus on students' learning processes rather than their outcomes, asking their thoughts and opinions instead of asking them to write a certain number of words or paragraphs.This can help to identify areas of strength and weakness in the students' writing and help them to develop their writing skills.

4.3.2
The Agency Dilemma: Unpacking Student Agency in the Complex Role of GAI in Student Literacy.There has been difculty determining the level of agency students have over their writing outputs when using GAI, particularly when it comes to disambiguating how much students write (i.e., the ideas, the sentence, the paragraph, the word choice) versus what GAI suggests and generates.From the writing workshop, we observed that many students (8/12) just copied and pasted directly from GAI-generated outputs into a Google Doc (i.e., "I'm done, I like the story, so why should I change it?"),raising the question of how to design the system to promote the craft of writing, such as idea generation, voice and style, audience awareness, revising, and more.Perceiving cutting and pasting as a refection of a lack of agency by their children, parents were skeptical about the impact GAI would have on their children.Most parents (9/12) pointed out the importance of establishing fundamental knowledge frst (i.e., comprehension and critical thinking skills) before introducing such automated systems as ChatGPT.As P07 mentioned, "How do my kids learn if AI generates everything for them?And do they know enough about the content of what they're asking the AI?I think learning is trial and error by doing things by themselves, and kids need to have the foundation to be able to build upon to access that new AI." Other parents consider that AI system access should determined by age-appropriate standards, as P11 stated, "I think the current version is defnitely not for kids age 8 or 9, it's too open-ended, my kid is too young and it's more important to learn foundation knowledge frst, I think that there is learning that has to happen with that." In accordance with the previous section, one of the key questions raised by adult participants was aspects of student autonomy (their ownership and agency over their writing project).The issue raises the challenge of designing child-AI interaction so that children can control their own learning processes, not just be led by AI.Hence, it is essential to develop AI-driven systems that respect children's autonomy, provide them with appropriate guidance and support, and ensure that the systems are suitable for children's age groups.

Erratic Outputs: Limitations and Concerns in Deploying GAI
for Literacy Education.Like any generative AI chatbot or voice assistant-such as Siri, Alexa, or Microsoft's ill-fated Tay [107]some individuals intentionally try to corrupt or manipulate GAIproduced responses, particularly in online settings.This behavior can take various forms, including providing chatbots with inappropriate or harmful content to elicit inappropriate responses, pushing the boundaries of what the chatbot can understand or respond to by inputting nonsensical or unusual queries to see how the chatbot reacts, or intentionally feeding chatbots with biased or false information to manipulate the responses and promote a particular agenda, ideology, or misinformation.The potential for such student-AI interactions was not lost on our teachers.T12 stressed, "I can imagine there will be kids who want to test the limits and get the chatbot to say inappropriate things back to them.So, I mean, there's that part of it." For instance, in our workshop, we observed students generating images around inappropriate political scenes (i.e., a Hitler statue), pointing to the need for developers of educational chatbot systems to implement safeguards and moderation mechanisms to minimize the impact of such intentional abuse.These safeguards may include content fltering, moderation of user inputs, and continuous improvement of the chatbot's response mechanisms to detect and handle inappropriate or harmful content [36,81,92].Less malicious but still disruptive are instances where a GAI system produces surreal or nonsensical responses to user prompts.GAI hallucination, also known as AI-generated hallucination or AI-induced hallucination, refers to a phenomenon where generative models produce content that may resemble hallucinations in humans, including images, text, or other sensory data that are typically unintended and often nonsensical (i.e., a dead animal without a head).AI hallucination occurs when a machine learning model generates content that doesn't align with the intended output [13,18].It can result from the model's overftting to its training data, exposure to unusual or biased data, or other factors that cause the model to produce strange or distorted outputs.A student (S02) pointed out an unexpected result had been generated from the GAI systems and stated, "If I do it without artifcial intelligence, I can do it with my hands exactly as I thought, but if I use artifcial intelligence, I think it can be seen as a disadvantage in that it is expressed slightly diferently than my intention." In instances where GAI-produced content is inaccurate but seemingly plausible, parents (10/12) argued it is important to consider whether or not students know AI-provided information is accurate.Several parents cited the need for educational AI deployments to be prefaced with fundamental education to develop critical thinking, comprehension, and problem-solving skills so their child can critically analyze and scrutinize information: "And do they know enough about the content of what they're asking the AI?How do we know if kids ask the right question, and how do we know if the information provided by AI is correct or not for students?I think kids frst learn through credited resources and develop that fundamental knowledge, at least by middle school." Based on our fndings, we identifed several challenges with current GAI systems, including the originality of students' writing projects (academic integrity), the agency of students in writing processes (learning), and the generation of misinformation due to the randomness of the GAI systems.These challenges are not distinctive from one another; rather, they are interconnected and need to be addressed collectively.In section 6, we discuss design implications that address the challenges mentioned above.

DISCUSSION
From the study, we examined the potential advantages and challenges of using GAI systems for literacy education in K-6 settings from multiple stakeholders' perspectives.We discovered how each stakeholder's views difer: for teachers, generative AI systems are a new type of digital citizenship development; for parents, these GAI systems are another type of toys or games; for students, these are smart, helpful companions.
In our discussion, we delve into the complexities of integrating cutting-edge educational technologies into learning settings, scrutinizing their impact on the design of GAI learning systems.Additionally, we outline three key design considerations essential for developing efective GAI-based educational applications.

Unpacking the Complexity of Technology Integration in Education
Despite substantial investments in educational technology, there is often a notable gap between the anticipated and actual usage of these tools in classroom environments [27].Teachers' varying levels of comfort and profciency with technology signifcantly infuence its application in teaching.Resource limitations also pose signifcant challenges, with issues like inadequate training, support, and access to current and functional technology impeding efective utilization [27].Reich (2020) underscores the importance of addressing the broader social, cultural, and pedagogical complexities in education, which he deems more crucial than mere technological advancement [87].
The recent LLMs have brought breakthroughs of open-ended conversational systems, which perform open-domain dialog with any topics [51] and it ofers the capability to be fne-tuned [82], enhancing its performance to align with specifc domains and instructional objectives [112,129].Unlike traditional MOOC platforms, which rely on human-guided instructions, educators can now train the LLM with specialized datasets and employ prompt engineering techniques [114] to enable AI to construct instructional content autonomously.Furthermore, students' educational behavior data, which includes their challenges and areas of profciency, can be fed back into the LLM for evaluation.This allows for algorithmically-guided decisions about where to begin instruction based on each student's capabilities.Eventually, educational systems will likely converge three distinct approaches within an integrated system-combining direct instruction, algorithm-guided learning, and AI facilitation.This system will not only instruct and guide but also foster open-ended exploration and collaboration between students and AI agents.Therefore, it is important to examine the possibility that these GAI systems can be integrated with new pedagogical approaches.
Consequently, new breakthrough systems like GPTs [79] will require thorough evaluation in terms of safety, efectiveness, and their ability to foster trust and community integration before they gradually become embedded in societal norms.Organizations such as the Institute of Education Sciences (IES) and Digital Promise, among others, are beginning to form communities of educators to explore the possibilities these systems ofer and to critically examine their applicability for teaching and learning [7,56].Consequently, it is anticipated that these technologies will be integrated into educational systems gradually rather than afecting a radical transformation in teaching and learning methodologies immediately.

Double-Edged Sword of GAI in Education.
From the study, we found educators were drawn to use the GAI systems for instruction and in the way that creating lesson plans (e.g., pre-, during, and post-instruction) can be made easier using AI-scafolded content creation.Meanwhile, students found they could leverage the systems to receive individualized and timely feedback.At the same time, parents pointed out the GAI systems' capabilities to facilitate interest-driven learning, particularly about culturally relevant approaches in writing projects [9,83].
The use of GAI in educational settings presents a complex blend of benefts and drawbacks, which are not mutually exclusive but rather exist simultaneously, refecting a double-edged nature.GAI facilitates open dialogue and free-form conversation, enabling the exploration of culturally diverse topics and translation capabilities.This openness enriches the educational experience by fostering a broader understanding of various cultures and languages.On the other hand, the same openness of GAI systems can lead to potential challenges, including the development of biased perspectives and the generation of inaccurate or 'hallucinated' results [32].Such issues underscore the critical need for careful moderation and strategic oversight, such as the implementation of customized models [30] (e.g., incorporating more diverse races into the image data set to train TTL) so that the system does not generate a particular ethnicity or race.Such precautions are crucial to harness the benefts of GAI while minimizing its risks for educational settings.

Recommendations for System Designers and Developers
As part of this discussion, we propose the design considerations of GAI-powered writing platforms to inform the designing of safe and accessible GAI systems for elementary school settings.To capitalize on the perceived benefts of educational uses of GAI while mitigating the concerns from our stakeholder groups, educational GAI platforms should: 1) provide guardrails to protect students' authorship issues in GAI-powered writing, 2) aford appropriate role allocation to AI and students, and 3) support customizable teacher-in-the-loop systems to enhance the trustworthiness and content-focus of GAI systems.

5.2.1
Navigating the Complexity of Authorship and Ownership in AI-Assisted Writing Systems.Our fndings highlighted that teachers are concerned about their students' authorship and integrity of their writing output, particularly when GAI generates the majority of the content for students [25,28,64].Even though studies have examined GAI-LLM-powered writing systems, such as Gero et al.
[42], Lee et al. [63], and Yuan et al. [126], focus on investigating language models' capacity rather than users' capabilities and their perspectives (including those with diferent cognitive levels, abilities, and ages).Furthermore, there is a lack of studies focusing on educational settings for K-6, which aim to mitigate specifc problems they face (i.e., authorship, plagiarism, assessment) [104].Gero et al. [42], and Lee et al. [63] have identifed that there is no onesize-fts-all solution when it comes to users' sense of ownership and authorship over AI-assisted writing processes due to uncertainty over authorship of language model-generated texts itself.Consequently, there is a need for further research into writers' ownership, authorship, and plagiarism, in addition to developing new methods for assessing and measuring writers' progress [29,58,59].
To better understand what guardrails and guidelines need to be implemented into the development of GAI-LLM-powered cowriting systems for K-6 students, future research on students' capacity, especially on measuring learning processes and assessment of the writing (e.g., how they interact with GAI-LLM like ChatGPT), would be benefcial.
To navigate the authorship and ownership of AI-assisted writing systems like ChatGPT, we propose building a system based on the LLM that facilitates cloud-based infrastructure.The database stores students' utterances in separation from AI-generated texts.To diferentiate between student-generated content and machinegenerated text, the platform will employ text-similarity analysis [57].This method allows educators to compare student writing with AI output, ofering insights into the extent of AI reliance on student work.

Enhancing Student
Agency through Role Allocation in GAI Systems Design.We observed that when students encountered openended GAI systems' interfaces (i.e., ChatGPT and Stable Difusion) without context, they had difculty writing prompts in a way that produced appropriate results.Hence, we argue for designing GAI-LLM co-writing platforms that mimic natural conversation, providing students with concrete context at the beginning of their interaction and ofering options for choosing topics of choice and characters to support child-centered and interest-driven learning experiences [34,83].
According to the workshop with students and teacher interviews, promoting students' agency as writers is essential [62], especially for enhancing learning experiences.As a result, students should be given opportunities to participate in writing projects and promote independent writing actively.This can be accomplished by allowing students to customize and edit their own writing.To facilitate safer and more efcient GAI systems in education without compromising their integrity, system developers and Edtech designers need to establish a division of tasks, setting up boundaries of roles between the AI agent for educators and students.By designing an AI agent persona and curating Child-AI conversations, this can be achieved by encouraging idea generation, adding story detail, and elaborating from the perspective of students.AI agents should be designed to help students think critically and creatively and to encourage them to ask questions through conversation [6,123].For instance, system developers allocate AI's persona as a coach or/peer rather than an assistant-that means rather than having AI generate writing on students' behalf, designing AI agents that encourage students to write their own creative ideas, giving students control over the writing process.Nguyen.[76] discusses the benefts of designing prompts that enable chatbots to foster systemic thinking (such as idea generation and questioning).Specifcally, Nguyen.[76] examined textual conversational agents' (chatbot) role design (personas) and its impact on students' system thinking process in group discussions.The fndings suggested more transactive exchanges with less knowledgeable peer agents (versus interacting with expert agents) as students felt more social and engaging.This fnding suggests that designing an age-appropriate agent role/ persona can impact conceptual understanding, enhancing learning outcomes.The current capacity of prompting LLMs ofers possibilities to optimize the free-form LLM-based chatbot dialogues for that purpose.

5.2.3
Balancing Flexibility and Control GAI-LLM Systems for Educator and Parent Oversight.Our fndings indicated that teachers and parents expressed concern about students' interaction with misinformation and biased content due to the system's randomness.To mitigate the uncertainty associated with GAI LLM systems, it is essential to design a system that balances fexibility and control with adults-in-the-loop systems [49,72,128].Yuan et al. [126] examined some of the methods that oversee the writing processes by providing suggestion options for users and ofering prompt design features from the back end.However, deciding and accepting the suggestions and writing prompts could be challenging for a certain age group and intellectual level or English profciency [69].
Hence, we propose designing an 'educators' view' that allows educators and/or parents to easily 'prompt' and curate GAI-based chatbots' conversation to facilitate a secure mode of student-AI interaction for writing.For example, the new systems will allow educators to prompt GAI systems to carry on their lessons, similar to the current tool that designs a chatbot with fow-based interfaces, such as Voicefow [8].Our suggestion is to develop fow-based interfaces [37] (or block-based interfaces [12]) for educators, where each node or block can translate into a prompt, which will create dialogue as teachers intend, continue writing project instructions, and construct conversation for students.By doing so, the system will provide educators control over a certain level of uncertainty the current GAI-LLM-based chatbot might have and provide openended fexibility, with low foors and high ceilings [89].
The majority of teachers (12/16) we interviewed expressed difculty adapting to new tools and AI applications (due to their heavy workload).Therefore, interfaces should be as simple (and easy to use) as teachers already know.To design the system, we recommend actively collaborating with teachers, co-designing the processes and interfaces through multiple steps of studies starting with needfndings and card-sorting [95] to understand their unique languages and mental model to create an appropriate conceptual model that aligns with educators' goals [65].With that series of user tests and gathering feedback from teachers and students, it is possible to refne the system and optimize its functionalities for educational purposes.

Directions for the future work
For future research directions aimed at broadening the scope and generalizability of our fndings, we advocate for an expanded investigation into GAI utilization.This should involve a comprehensive analysis of system logs and behavioral data within GAI platforms.This includes leveraging GAI platforms for collecting back-end educational data to analyze students' learning progress such as their reliance on AI, writing quality, and the nature of AI-student interactions.By engaging a wider participant base and adopting a longitudinal study approach, we can deepen our understanding of how GAI tools infuence user interactions, experiences, and learning outcomes over time.
To promote accelerated learning through GAI-powered learning tools, further research could also include A/B testing, using multidimensional metrics to evaluate student writing.These metrics include Production: the amount of writing users generate over time and per session within the system, Narrativity: the extent to which a text tells a story with characters, events, places, and things, Syntactic Complexity: the complexity of the text's syntactic structure, Vocabulary: sophistication and concreteness of students' word choice, Grammatical Correctness: the extent to which students' texts adhere to grammar norms [44,45,73].By integrating these AI and database systems, designers and researchers will be better equipped to understand the details of student interaction with AI in writing, aiding in the development of more efective educational tools.This approach enriches insights into AI's educational applications and also sets a foundation for future studies focused on the nuanced dynamics of AI-assisted learning.

CONCLUSION
In this paper, we explored the stakeholders in education's perceptions and opinions regarding the advantages and limitations of leveraging GAI systems in literacy education for elementary school students.Through qualitative studies, conducting workshops and interviews with teachers, parents, and students of 40 total participants, we found that the GAI systems can be used to generate adaptive lesson plan materials such as mentor text for teachers for them to tailor according to each student's needs and skill level (through scafolding and their interests).The GAI system afords culturally relevant and timely feedback that broadens ideation for writing projects.We also discovered the limitations of the systems in determining the authenticity of students' writing projects, difculties determining students' agency over their writing outcomes, and concerns regarding the safety and accuracy of the content.Based on the fndings, we provide implications for future studies to navigate authorship and ownership of AI-assisted writing projects that students produce.We also drew design suggestions to mitigate the concerns regarding the safety and accuracy of content.First, we recommend promoting student agency through role allocation over AI and humans, allowing more room for students to customize and edit their own writing.Second, we propose facilitating teacher-in-the-loop systems where educators and parents can control the lessons by prompting AI to carry on their lessons based on their design.Our study highlights an opportunity to foster collaboration between researchers in the HCI, Education, GAI, and NLP communities to design a GAI-powered platform for literacy education.

4. 1 . 1
Adapting Digital Transformation: Teachers' Perspective on Integrating GAI in Digital Literacy Development.Our results indicate that teachers acknowledge that their students will grow up in a society where emergent digital technology is an integral part of life.

Figure 1 :
Figure 1: Summary of each stakeholder's perspectives and opinions of GAI systems (top: teachers, middle: parents, bottom: students).

Figure 2 :
Figure 2: Summary of our fndings of potential afordances and limitations of GAI systems for writing projects in elementary school settings Teachers and parents were particularly intrigued about the possibility of translating languages and providing examples of diferent cultures with GAI systems.Teachers intend to utilize GAI systems to generate culturally tailored examples they might not be familiar with during lesson planning.T15 stated, "If I'm giving an assignment, and I'm trying to give examples, I only know the examples I know.And I have my cultural bias, I have my background, my limited experience.But if I get to ChatGPT to generate more examples of active and passive voice, it's gonna save a lot of time.And again, I can incorporate things from diferent interest levels, cultures, and vocabulary levels."

Table 1 :
Workshop schedule

Table 2 :
Participants' information for the family workshop

Table 3 :
Participants' information for the interview study experiences and opinions about adapting them in educational settings specifc to writing activities with their students.Most teachers (10/16) already have experience with ChatGPT and relevant GAI systems