1 Introduction

Embedding artificial intelligence (AI) technology into the classroom as a teaching assistant, a tutor, and an advisor has increased and brought opportunities to realize the complementary synergy with human teachers in teaching activities (Kim, 2023; Kim et al., 2022). In fact, much prior work explored and designed AI-based systems or tools to augment teachers’ ability and co-orchestrate complex classroom learning situations between teachers and AI to better support students, rather than AI directly offering personalized support (Holstein & Aleven, 2022; Molenaar, 2022). For instance, AI-embedded dashboards support teachers to understand and monitor students’ learning processes and offer teachers supplementary insights to adequately respond to the needs of the students (Molenaar & Knoop-Van Campen, 2018). Further, AI could possibly facilitate co-orchestration between human teachers to leverage the complementary strengths of each. As an example, Holstein et al. (2018) designed Lumilo, mixed-reality smart glasses, that aid the teacher in describing the actions happening in the classroom in real-time to implement appropriate pedagogical actions. It should be noted, though, that AI cannot effectively address student learning on its own such as students' gaming the system behaviors, or hint avoidance.

The interest and demand for teacher-AI collaboration (TAC) have been increasing to the point of becoming a critical research and design challenge for AI in Education (AIED) (Rodríguez-Triana et al., 2017), however, the implementation of TAC is complex and challenging in K-12 schools. In practice, teachers are used to being the only teacher responsible for the course and the change required to work efficiently in a team of colleagues is challenging even in a human-human collaborative teaching context. This change is profound as it is not just a method or single skill, but it actually changes the whole culture of teaching (Vangrieken et al., 2015). Without formal AIED training, teachers are expected to jointly orchestrate classroom instruction in support of a fully-packed curriculum that involves the planning and real-time management of diverse classroom activities (Dillenbourg & Jermann, 2010) alongside AI without a clear model or best practices to serve as a guideline for that joint orchestration.

Considering that human teachers play critical roles in mediating the effectiveness of AI in K-12 classrooms (Ritter et al., 2016), this study aims to explore the types of TAC in teaching that could be adopted in classroom instruction and to examine both the benefits and challenges of TAC perceived by K-12 teachers. The questions guiding the study are as follows:

  1. 1.

    What types of TAC can be implemented in the classroom instruction?

  2. 2.

    What do teachers think are the main opportunities for TAC in classroom instruction?

  3. 3.

    What are the obstacles to TAC perceived by teachers?

The findings of this study can offer insights into how teachers leverage TAC at the classroom levels and ensure teachers-in-the-loop in AI-supported instruction (Ramos et al., 2020; Renz & Vladova, 2021) whereby AI empowers and not overpowers teachers so that teachers' expertise is respected and teachers are actively involved and have control over the instructional process while balancing the process of human and AI-driven decision-making and mutual monitoring. The study findings can also direct the implications for future developments and practice to support TAC.

2 Literature review

While many research and government initiatives have led to high expectations that AI will save teacher time that could be redirected toward student learning (Miao et al., 2021), others have claimed that AI will make teachers redundant – or teachers’ roles will be at least reconceptualized to be learning designers and classroom orchestrators who manage learner behavior, ensuring smooth interactions between technology and students (Celik et al., 2022; Seldon & Abidoye, 2018).

Amidst the contrasting opinions and perspectives regarding the teacher and AI relationship, a rich line of recent research has proposed the notion of TAC, or human-AI co-orchestration that amplifies the complementary strengths of each participant without simply automating or replacing valuable human-human interactions and limiting the roles of teachers (Holstein et al., 2019; Cukurova et al., 2019; Mavrikis et al., 2021). This shifts the attention that was previously focused on learner-supporting AI, which offers personalized support directly to learners (i.e., AI-powered adaptive tutoring, Du Boulay, 2016), to the design of effective TAC processes and systems whereby teachers and AI are considered equal team members in orchestrating individual and collaborative instructional activities during the class (Dillenbourg et al., 2018). These AI systems are designed to augment human teacher ability to overcome teacher limitations and co-orchestrate classroom interactions and activities (Holstein & Aleven, 2022; Ji et al., 2023). For instance, Kaliisa and Dolonen (2023) designed the Canvas Discussion Analytics Dashboard (CADA) to augment teachers' awareness of relevant features of online learning situations (e.g., students' participation and discourse patterns), and offer some hints regarding what teachers need to do to support students. Teachers can take advantage of the hints provided by the dashboard to facilitate relevant pedagogical decisions and make relevant interventions. That is to say, the dashboard not only helps teachers sense the status quo information about learning but also enhances and expands the range of teachers’ instructional actions. Additionally, some research has explored design approaches in which teachers help and train AI systems, via demonstration, to sense relevant information about students’ broader contexts (e.g., students’ at-home difficulties that may negatively impact their academic performance, Bull & Kay, 2016), recognize instructionally relevant features, and perceive learning situations and information (Lee et al., 2015).

In sum, the discourse around the utilization of AIED shifted from being narrowly focused on automation of teaching tasks and AI-directed teaching to the augmentation of human teacher capabilities linked to learning and teaching (Chatti et al., 2020) and identifying the advantages, scope, and challenges in forming effective TAC. Many studies anticipate and highlight TAC could support teachers in delivering more personalized learning experiences while promoting their motivation, engagement, and immersion in learning, and enhancing self-efficacy, self-organization, and ownership of their own learning (Kim, in press). For example, AI is widely used to create exploratory and gamified learning environments, generating and suggesting personalized learning tasks to help teachers recognize, facilitate, and achieve quality engagement of each individual learner (Miao et al., 2021). The Kid Space, an interactive learning environment, incorporates extensive multimodal sensing and sense-making technologies to engage learners both physically and socially with play-based learning, guided by an animated teddy bear that analyzes and responds to learners' behaviors, emotions, and utterances in real time (Anderson et al., 2018). On the other hand, AI could support teachers in preparing and developing instructional materials by searching and scraping online educational resources (e.g., X5Learn, Perez-Ortiz et al., 2021), analyze and support teacher practices (Wise & Jung, 2019), time management and course/lesson planning (e.g., Chounta et al., 2022) .

Despite these promising use cases, there is evidently widespread concern related to the practice of effective TAC. First of all, it is important to acknowledge that while AI has demonstrated a strong ability to solve a single task in a well-defined scenario it has limitations in understanding various contexts, scenarios, and behaviors. These limitations restrict AI’s ability to address highly complex and open problems in diverse teaching scenarios which serves as a key challenge in TAC (Crescenzi-Lanna, 2023). Second, much research underlines the need to empower teachers' AI-related competencies. This encompasses not just general computational thinking skills, but also the understanding and use of key concepts to create and apply AI tools and AI-empowered pedagogical approaches to facilitate learning while ensuring teachers' agency and accountabilities (Kim et al., 2022; Miao et al., 2021; Touretzky et al., 2019). Particularly, the teachers' lower levels of meta-knowledge in recognizing and assessing complementarities and shortcomings of working AI lead to poor delegation decisions (Fügener et al., 2022) and distrust toward AI (e.g., algorithm aversion, Dietvorst et al., 2015).

It remains unclear how a teacher and AI can form and apply TAC in the classroom context. An active, collaborative relationship where these two heterogeneous instructors are working purposefully, regularly, and collaboratively to deliver substantive instruction, and to meet the diverse learning needs of students has rarely been demonstrated. This study, therefore, aims to explore the diverse forms of TAC in classroom instruction perceived by teachers. Such foresight indicators can enhance our understanding of what complementary strengths teachers and AI systems hold in a given context and can be used to guide the design of systems that combine these strengths.

3 Methods

3.1 Participants

This study adopted a purposeful sampling method to recruit 30 research participants. Each participant’s school level they are currently teaching, gender, education level, teaching years, and subject as well as types of AI used for instruction were examined (see Table 1). Regarding the types of AI used for instruction, participants were asked to choose ones that apply to their own experience among five types: Intelligent Tutoring System (ITS), Exploratory Learning Environments (ELE), AI-based dashboard, Automatic writing evaluation (AWE), and chatbot (Homles at al., 2019a; Lameras & Arnab, 2021). In addition, all participants were required to have at least a year of teaching experience in AIED, either in teaching with AI or teaching about AI (Holmes et al., 2019a, b). The former refers to the use of AI as direct teaching tools such as AI-based mobile apps and adaptive or personalized learning management systems, whereas the latter indicates actual teaching experience in AI skills and understanding (i.e., AI concepts, AI applications, AI ethics, building AI). The study did not require participants to have extensive knowledge or technical skills in AI model development as the study aimed to examine their perspective on TAC and its potential effects, not developing AI modeling techniques.

Table 1 Participants’ characteristics
Full size table

This study obtained ethical approval from the author’s university’s Institutional Review Board (NO. ER-AOFE-11000086620220922083531) and also received informed consent from all participants.

3.2 Data collection

The study conducted Focus Group Interviews (FGI) to allow the participants to express opinions, describe experiences, and provide detailed information (Turner, 2010).

Prior to conducting FGI, the study created scenarios to better facilitate the interviews as well as allow participants to shape their own perspectives and assess TAC in a more critical and nuanced manner within a real-classroom context with their peer group (Zimmerman & Forlizzi, 2017). To explore teachers’ diverse perceptions, the AI that collaborates with teachers in the study embraces all types of AI being used in school instruction, both physical (Physical systems in which AI is embedded such as Robot, AI speakers, etc.) and digital AI (e.g., model, algorithms, generative AI, etc.)

To create scenarios, the study first conducted an online brain-writing activity about TAC scenarios in K-12 classrooms (Linsey & Becker, 2011) among 6 teachers (2 from primary, 2 from middle, 2 from high schools), with an average of 7.8 years (SD=3.5 years) of teaching experience to create technically feasible and relevant TAC in a classroom instruction context. Each teacher wrote down their ideas in response to TAC scenarios using Miro (https://miro.com/brainstorming/), a free online collaborative whiteboard and passed them to another teacher for review and refinement. Once every proposed scenario made a full round, the team shared all the ideas together to review and refine each idea until all designers agreed that the scenarios of TAC were technically feasible and contextualized in the real classroom. These TAC scenarios were further validated through expert reviews. The study conducted individual semi-structured interviews with a total of 5 experts (1 from instructional design and technology, 1 from primary education, 1 in secondary education, and 2 from Human-Computer Interaction) with an average of 10.8 years (SD=7.8 years) of research experience and 8 years (SD=6.2 years) of teaching experience, using a video conferencing platform (i.e., Zoom and Tencent platform) in 60–70 minutes. The questions discussed include, but are not limited to: (1) Can you improve this scenario to make it technically feasible?; (2) can you improve this scenario to have a positive impact on classroom interaction based on your own teaching experience?; (3) Do you have any research ideas that can be used as a new scenario? As a result, prior to conducting FGI with participants, a total of 15 scenarios (Table 2) were developed by reflecting on the experts’ review of the teacher created scenarios and AIED literature which were then grouped into six categories: (1) One teach-One observe (one has primary responsibility while the other gathers specific observational information on students or the instructing teacher); (2) One teach-One assist (one has primary instructional responsibility, while the other assists students with their work, monitor behaviors, or corrects assignments); (3) Station teaching (the co-teaching pair divides the instructional content into parts and the students into groups); (4) parallel teaching (each instructs half the students on the same lesson); (5) alternative teaching (one teaches the main lesson to a larger group of students while the other works with the small group of students on an entirely different lesson); and (6) team teaching (both delivers the same instruction at the same time).

Table 2 TAC scenarios
Full size table

This study formed six focus groups consisting of teachers from different levels of schools (i.e., primary, middle, and high schools), teaching subjects, years of teaching experiences, schools, and gender (see Table 1). A single FGI was conducted per week between 3 March 2023 to 7 April 2023 until all groups had been interviewed with each session lasting approximately two hours. Each group of participants first read each of the scenarios aloud and was asked to critically assess and discuss how TAC could be implemented and contextualized in the actual classroom without being too intrusive and, additionally, describe potential benefits for and obstacles to TAC that could undermine or strengthen classroom teaching and learning with guiding questions (see Table 3). Meanwhile, participants were also asked to choose the forms of TAC that would work well and identify which did not sound desirable to elicit their holistic point of view regarding perceived opportunities and drawbacks of TAC.

Table 3 Interview guiding questions
Full size table

All interviews were conducted in Chinese, the participants' native language, and audio-recorded, transcribed, and later translated into English by research assistants with research experience in translation in both Chinese and English.

3.3 Data analysis

A reflective thematic analysis (Braun & Clarke, 2006) was undertaken to identify and construct emergent themes for each RQ. After going through the data familiarization stage by reading and re-reading the transcripts, three researchers independently generated initial codes related to intriguing statements or phrases in the data. These were then combined into 20 potential overarching themes. The entire dataset was once again thoroughly reviewed to uncover and develop any new emerging codes and themes. These were compared and contrasted with each other until all researchers reached an agreement on every theme.

To ensure the reliability of the qualitative data analysis, the study employed two strategies: (1) respondent validation and (2) employing moderators (Golafshani, 2003). The interview transcripts and analysis results were confirmed with each participant and were allowed to be revised when necessary. In addition, the study conducted three iterative discussions with the team of instructional designers and experts who were involved in the TAC scenario storyboard development to review, organize, define, and name the themes. As a result, a total of 14 themes were identified which included 6 themes with 7 sub-themes under RQ1, 4 themes with 4 sub-themes under RQ2, and 4 themes with 3 sub-themes under RQ3 (see Table 4).

Table 4 Summary of emergent themes
Full size table

4 Findings and discussion

4.1 Types of TAC

4.1.1 One teach, one observe

Teachers anticipated that they would serve as leading teachers while AI could play a key role in learning analytics and teaching analytics. Although AI’s role in learning analytics and teaching analytics share a similar end goal of effectively improving teachers’ instructional processes, there are differences in the target audiences, objectives, and educational data used. First, teachers suggest that teachers would serve as the primary teachers, while AI gathers specific observational data on students’ learning progress, performance, and engagement with educational resources/tools and tasks/activities as well as their peers in learning which otherwise is difficult to gather and track by an individual teacher. In turn, teachers can better understand how the learning process is carried out including challenges that exist within the classroom experience, providing them the information necessary to maximize learning potential and reinforce behaviors while applying interventions that more closely align with learners’ needs and preferences.

Also, teachers perceive that AI’s support in student observation creates a starting point for helpful conversations between both learners and teachers rather than having a broad idea that is unable to target specific problems. Teachers can take this opportunity presented by AI to help students feel connected with classroom learning and to engage students as partners in improving their learning environments.

AI's observable data may not fully pinpoint students' various learning challenges but I can take them as a point of reference to ask students questions about their behavior and interpret and discuss data together and clarify and plan for the next steps. (D18)

By contrast, AI-powered teaching analytics supports the primary teacher in reflecting on the effectiveness of their instructional design and delivery process and in developing an approach for further action and refinement. Hence, AI is directed to teachers and collects data that is relevant to support teachers in reflection on teaching approaches and their engagement with the activities, such as the frequency of teachers' participation and engagement in discussion activities, levels of engagement in the teaching process, analysis of their teaching approach, and any points of modification including the learning content delivery mode.

4.1.2 One teach, one assist

In line with the abovementioned themes, teachers expected TAC would consist of a teacher serving in a lead role while AI served in a secondary role, in this case as an administrative assistant. To be specific, teachers expected AI to carry out routine administrative and clerical tasks that distract them from their core teaching responsibilities. For instance, responding to students’ frequently asked questions (A4, B7); analyzing attendance data (B9, C13); digital device troubleshooting such as offering virtual technical support (D20, E22); processing exam results (A3, F26); and filing, collating, and reporting records of their student’s academic progress and constantly update this information in digital forms (B8, C15). With respect to exam-related tasks, although teachers understood AI could automatically assess learner assignments and exams which would save time, teachers argued that AI’s depth of interpretation or accuracy of analysis cannot excel beyond what one human teacher can offer (Holmes et al., 2019a, b). Also, teachers highlighted that teachers learn about their learners while assessing outcomes that allow teachers to develop insights on what and how to support individual students. Thereby, teachers envision TAC in assessment in a way that AI supports the grading/assessment process (i.e., offering prompts and shortcuts) while teachers do the actual grading.

Furthermore, while teachers take primary responsibility for delivering whole-class instruction, AI could support individual students within the whole group as an assistant tutor by answering student questions (E21, F26), monitoring/correcting students' works (A1, B7), and suggesting a broad range of instructional resources which may include but is not limited to multimedia resources, supplementary readings and task sheets (C12, D17) to encourage, improve, and promote teaching and learning activities during the process of instruction.

4.1.3 Co-teaching in stations

Another form of TAC perceived by teachers occurs in the station teaching approach; the co-teaching pair of human teacher and AI divides the class into workstations or learning centers where learners rotate in and out in a small group, which is either managed by a teacher or AI. Teachers proposed that stations could be designed with different learning activities/tasks and/or different concepts related to the same lesson to utilize both teacher and AI lesson-planning abilities and content area strengths as well as enhance students' deeper understanding of the subject and applied knowledge.

On the other side, learning activities could be designed and conducted with the same concepts and ideas, but in a sequence of different difficulty levels to move students progressively toward stronger understanding and, ultimately, greater independence in the learning process. Students, for instance, can work with a teacher to solve a set of five polynomial equations at station 1. At station 2, students will use laptops to complete a Desmos activity on graphing polynomials with AI and so on (A4).

4.1.4 Parallel teaching in online and offline classes

The TAC can also take place in the form of parallel teaching in which the class is split in half, with each half being taught the same lesson by either a human teacher or AI. Teachers are expected to collaborate with AI to functionalize an effective Hybrid-flexible (Hyflex) model of instruction that could offer equivalent learning experiences for both those attending online and in-person classes to ensure that students present in the classroom and those studying remotely due to a variety of unique situations (e.g. health, location, etc.) can learn together. However, implementing a Hyflex mode of instruction puts increased pressure on teachers to plan and design instruction that engages both in-person and online attendees, while also considering potential tech issues in the online learning environment and fully supporting interaction between in-person students, virtual students, and the teacher during instruction. In this regard, teachers expected TAC could take the form of one class being taught by a human teacher in a traditional classroom environment while also independently having the AI support the online learning environment. TAC in this context could contribute to the development of creative solutions to such challenges by reducing the student-teacher ratio and increasing instructional intensity for an individual student.

4.1.5 Differentiated teaching

Another form of TAC is the differentiated teaching approach, as identified in TAC scenarios of Table 2, in which one instructs most of the class while the other works with a smaller group of students who need specialized attention. Teachers and AI can work to analyze and review students' data to figure out which students need support filling in gaps in background knowledge, which students need remediation, or which students could benefit from accelerated learning because they already know the content or have mastered the skills of the large group lesson and offer a specific instruction accordingly. Teachers anticipated that this approach, whether teachers take a big group or a smaller group, allows teachers instructional flexibility in better managing what students learn, how students learn, and how students are assessed and tailoring instructional strategies to maximize individual growth in the teaching content.

4.1.6 Team teaching

Teachers suggested that the teacher and AI could co-teach at the same time, but take turns, sharing the responsibilities of lead instruction, with equally active, but possibly different roles in a lesson (e.g., the teacher makes a point, AI can jump in and elaborate if needed or vice versa, D16; Teacher leads the class, AI leads the experiment, E25).

In line with this, it was interesting to capture that teachers reflected on the different strengths of humans and AI teachers in a team teaching scenario. For instance, AI is efficient in teaching facts, logic, and theories whereas human teachers can read social cues and strong sensory feelings and emotions, to nurture students' whole-person development, also known as holistic learning, developing students both intellectually (e.g., higher-order thinking like analytic abilities, technical skills — things like C++ programming, data set analysis, or process mapping, and the acquisition of subject-based knowledge) and emotionally (e.g., value, consciousness, self-/social awareness, intrapersonal communication, adaptability, and conflict resolution).

AI teaches with something that it is good at or even excels over humans such as teaching programming since AI automatically detects errors in code that I often miss and even takes so much time during the class. I then support students' soft skills such as defining genuine problems, collaborating with peers, and presentation skills, which are unable to articulate and explain clearly for AI to instruct them to students. (F29)

4.2 Benefits of TAC

4.2.1 Instructional design

Teachers described that TAC can help teachers better design classroom instruction for engaging students in authentic problem-solving while deepening subject content knowledge. In particular, teachers highlighted that working with learning analytics systems powered by AI improves an understanding of their students and better informs them about students’ learning difficulties and dispositions that they may be unable to identify during traditional classroom instruction. Further, learning analytics supports monitoring of the evolution of individual student learning progress and can predict the learning time needed for mastering a subject. In turn, teachers can plan for better customized pedagogical interventions and resources, improve the structure of learned knowledge, and design differentiated and flexible learning opportunities for students in the classroom, hopefully leading to better learning gains (see Table 4). Teachers, however, argued that their efforts in working out the details of individual student academic progress and performance and implementing personalized instruction through TAC are impacted by schools and governments that set the same fixed learning outcomes that are ultimately measured by standardized exams. In other words, the learning path design may be set in personalized ways but not the destination (Holmes et al., 2018). Meanwhile, teachers highlighted that education is not merely about supporting students' cognitive development only. Thereby, they aspired to work with AI systems to develop personalized instruction that can support learners’ broader areas of development including cognition, social interaction with peers, emotional intelligence, well-being, and so on.

4.2.2 Teaching delivery

Another benefit of TAC perceived by teachers was creating more opportunities to implement creative and engaging activities. As F27 said in Table 4, teachers recognized that working with AI supports technology-rich environments whereby teachers and AI can both conduct hands-on activities, (re)create a realistic and authentic real-life event or a set of learning situations to solve problems in the areas of the instruction. Through such activities, teachers expected their teaching delivery of subject-specific knowledge such as facts and definitions could be digested and consolidated by students through presentation, meaning-making, problem-solving, and inquiry activities supported by both teachers and AI.

TAC also allows for more flexible student grouping. Teachers reflected on data analysis proposed by AI and put students into small groups for instruction considering their different readiness, interest, or learning styles. Students can be grouped at the same skill level or with varying skill levels. Flexible grouping also allowed smaller student-to-teacher ratios in which teachers felt more confident in treating and supporting the unique needs of individual learners with learning and attention issues.

It’s simply impossible for one teacher to listen in on the conversations of multiple collaborative groups at once. The AI system can monitor multiple written discussions simultaneously, providing timely alerts to me if a group’s conversation repeatedly goes off-topic, or conversely if a group is progressing at a faster rate than their peers. In turn, I feel less overwhelmed to address lapses in effective collaboration and provide students with more individualized support (C15)

4.2.3 Teacher professional development

Teachers perceived that TAC supports their professional development. First, teachers critically examine their own instructional practices, experiences, and assumptions with data collection and analysis conducted by AI which is largely engaged in three stages: (1) the collection of data about what is happening in the classroom, (2) the analysis and evaluation of this data, and (3) an exploration of how this data relates to and can inform their practice and beliefs as a teacher (Wise & Jung, 2019). Teachers found reflection through TAC is more flexible, less time-consuming, and self-directed compared to ones that involve their colleagues in the reflective process through class observation and consultation (see Table 4). However, teachers expressed that the process of reflection with AI requires high-quality data about classroom practices, student experiences, and outcomes, followed by thoughtful analysis, and contextualized interpretation in an instructional setting. Thereby, there needs to be more support for teacher reflection on their practices through the collection and analysis of sustainable, continuous, and robust collection, analysis, and interpretation of multimodal data (Persico & Pozzi, 2015) while respecting qualitative insights teachers have into their role as instructors and instructional designers (Warr & Mishra, 2021) and promoting dynamic TAC interaction mechanisms necessary for teacher reflection to meet changes in curriculum and the varied needs of their students (Wise & Jung, 2019).

Teachers also identified that TAC requires them to engage in continual learning to stay current in the most effective and efficient uses of AI, a landscape that is constantly changing, to develop an understanding of how to combine and leverage content, pedagogy, and technology knowledge to better support students’ learning (e.g., acquisition of content knowledge, F28; engaging students in authentic problem-solving, D19) while benefiting teaching practices.

4.2.4 Lowering grading load

TAC could reduce teachers' grading load, enabling teachers to provide greater amounts of feedback on students' learning as well as provide students with immediate automated feedback to help students revise and improve their work (Grimes & Warschauer, 2010; Wilson & Roscoe, 2020), although most teachers highlight such an automatic evaluation system works well only on multiple-choice or assignments with well-defined answers such as math and coding.

However, given that many types of assignments (i.e., daily exercises and quizzes) still follow conventional paper-based practices, teachers highlight that there is a need to establish and adopt a learning management system (LMS) that can automate or support data processing tasks such as inputting data, digitalizing records, and tidying up the formats of diverse types of assignment.

4.3 Obstacles to TAC

4.3.1 Lack of explicit and consistent curriculum guidance

The lack of explicit and consistent curricular structure and guidance is a prominent challenge identified by teachers. This deficiency impacts the four fundamental components of classroom instruction, as shaped through TAC: the 'why', the 'what', the 'how' of learning, and 'how well' the classroom instruction achieves its goals (Tyler, 1949).

First of all, teachers expressed that an absence of informed visions and goals of AIED inhibits consistently defining the purpose of TAC for classroom instruction, diversifying the forms of TAC, and aligning TAC with diverse pedagogy. Although personalization of learning has been the strong rationale for the integration of AIED, most teachers described that the dominant goals for AIED tend to be two-fold: (1) Cultivating high-level Talents in AI development and (2) Improving standardized exam results. Reflecting on the former goal, TAC's focus and outcomes prioritize improving students' technical skills such as programming languages, and data science, and developing advanced AI. Meanwhile, the latter goals directed TAC to support students to achieve a higher grade in a standardized examination, such as Gaokao (The national college entrance examination), and prioritize rote memorization over higher-order thinking (i.e., creativity, metacognition), and knowing facts over a critical engagement (i.e., critical analysis and reflection). For instance, B10 expressed,

I think we somehow misuse or underuse AI when it comes to the educational domain since we can’t direct AI to play what significant tasks to be engaged in and what contributions it should offer to education. We simply use such state-of-art technologies just to prepare students for exams.

Her views reflect critical questions discussed by previous literature: (1) ‘what is the right kind of education that we need to pursue and deliver jointly with AI?’; (2) ‘are the AI technologies being used in schools addressing the right educational tasks?’; and ‘are they enhancing learning as an essentially human and social activity, or aiming to simply make learning ‘more efficient?’ (Holmes, 2020; Kim et al., 2022; Miao & Holmes, 2021).

In line with the disjointed goals of AIED, the lack of systematic and authoritative guidance on teaching content and related materials was another significant challenge to effective TAC. Although teachers list numerous examples of teaching approaches in AIED, including teaching AI concepts in ICT/STEM activities, the use of AI-powered apps to support student subject knowledge learning, and interdisciplinary learning under the AI+X framework, teachers argued that existing textbooks and ready-to-use instruction packages are all directed towards the emphasis on technology (e.g., the concepts and inner mechanisms of AI systems). In this regard, teachers expressed concerns that such technology-centric content caters to the needs of students who want to pursue further study or work in AI disciplines, thereby TAC mainly occurs within the domain of STEM and coding education while underserving other domains and fails to accommodate students' diverse characteristics (i.e., learning interests, motivation and understanding of AI). Their views allude that insufficient understanding of AI and pedagogical content and materials could exacerbate the implementation of a broader coverage of TAC and thereby the teaching areas delivered through TAC could also be narrowly defined (Long & Magerko, 2020; Zhou et al., 2020).

Furthermore, limited TAC target outcomes and evaluation mechanisms led teachers to remain unclear about the scope and scale of the TAC; a question of how different activities and relations of collaboration between teachers and AI fit together within a wider educational practice and what tangible outcomes to pursue over long-time periods to make the most sustainable TAC was not addressed well. As E24 illustrated in Table 4, although teachers were informed about a wide range of benefits that TAC could offer, as noted earlier from the literature review, teachers argue that learners' high-level academic performance and achievements from school exams were considered and counted as the key intended outcomes of TAC, while its potential impacts on other areas such as teachers’ decision-making process, empowering students’ cognition or supporting mental health, facilitating classroom practices are not taken into account.

4.3.2 Commercial-driven AI

Most AI tools currently in use in educational contexts have been a commercial version of its AI model, which are not generally developed in consideration of educational usage nor reflective of proven pedagogical need. Although teachers acknowledged the diverse roles of the private sector in AIED and support for AIED in schools (e.g., offering training bases for teachers, D17; supporting students’ AI-related extra-activities such as contests and boot camp, E22 and building AI innovation space in the school, E21), teachers claimed that AI that they currently encounter and work with for the purpose of instruction, a specific application domain, has been designed, developed, and deployed dominated by computer-centric views without sufficient understanding and considerations of underlying education rationale and adequate pedagogical principles (Holmes et al., 2019a, b). Thereby, the educational application of these models may not be aligned with classroom instruction. In turn, teachers argued that such pedagogically insufficient and incompetent AI may scale up poor pedagogic practice, rather than supporting students and teachers in developing innovative approaches to learning and instruction, and thereby its roles would ultimately diminish and remain merely as mundane back-end AI-as-a-service plug-ins, or a specific type of recommender system, although its potential goes far beyond these uses (Williamson & Eynon, 2020).

4.3.3 Absence of clear AI ethics

In contrast with the health sector, where there are long-established ethical principles and codes of practice with regard to the treatment of human subjects, an absence of a commonly accepted model of ethics or a robust ethical framework, approach, and procedures exists for AIED and TAC research which was found to be another critical challenge to further understanding and developing TAC for classroom instruction. In particular, as processes and outcomes of TAC have an unknown and unintended impact on individual learner’s learning experience, and their neurological, cognitive, and socio-emotional development for life (Gottschalk, 2019). As TAC is used at scale, teachers expressed a great need for clear and robust ethical guidelines and interventions that demonstrate AI efficacy and safety in the educational context, including foreseeable misuse, through the entire life cycle of TAC use in classroom instruction.

Meanwhile, teachers argued that the ethics of data (i.e., data bias, privacy, transparency, and trust in data analysis) and computational power (i.e., accuracy in prediction modeling and AI recommendation) are crucial but not sufficient alone. Given that TAC is set to take place in diverse areas and processes in classroom instruction, the ethics of AI also needs to account for the ethics of education including, but not limited to, the ethics of teacher expectations, of resources allocations (including teacher expertise), of gender and ethnic biases, of behavior and discipline, of the accuracy and validity of assessments, of what constitutes useful knowledge, of teacher roles, of power relations between teachers and their students, and of particular approaches to pedagogy, such as behaviorism versus constructivism (Holmes et al., 2022b).

4.3.4 Negative attitude toward AI

Teachers in the study indicated that teachers’ biased and negative attitudes toward AI could be another barrier to effective TAC in the classroom. Many teachers made comments about their colleagues’ attitude toward AI during the interviews and confided that most of their colleagues naively assume that AI would replace human teachers, or they simply consider AI as another kind of sophisticated machine that has been put into the classroom instruction forced by the government. To some extent, their concern was valid; teachers without knowing what AI is, what technical levels are required for teachers to interact with AI, and how AI could work with teachers and students in what ways, could have been overwhelmed with concerns and anxiety. However, one interesting point to capture is that teachers have, in fact, been using technology in classroom instruction and many of those technology tools are powered by AI (Schmidt, 2016). Teachers appeared to have formed a generally positive attitude towards ICT usage in contrast to their critical view AI. For instance, D17 expressed:

Teaching with ICTs has been promoted among Chinese education circles for many years. I am sure all the teachers nowadays use ICTs, more or less. We have reached the point of no return. But they are not fully aware that AI has been embedded into those technologies so our teaching practice has now become convenient.

This quote resonates with previous studies highlighting that it is crucial to understand how people feel about AI, how people are affected by AI presence, and how people’s attitudes toward AI may affect or bias subsequent interaction with AI in order for AI to be integrated and accepted in human societies (Nomura et al., 2006). Accordingly, teachers’ attitude toward AI constitutes an important element for their teaching experience and performance with AI as well as persistence in teaching with technology (Tas, 2016; Wentzel et al., 2010). In addition, existing literature has shown that humans’ incomplete and primitive understanding/conceptualization of AI can influence their interactions with technology, and determine their actions with technologies (Phillips et al., 2011). Also, an incomplete understanding could potentially lead humans to misunderstand AI’s abilities and thereby create a pitfall for under-utilization (e.g., misuse or disuse).

5 Conclusion

The present study explored teachers’ perceptions of diverse forms of TAC as well as the opportunities of and obstacles to TAC pertinent to implementation in a K-12 context. Participating teachers in the study recognized the diverse forms of TAC that could be adopted to contribute to classroom instruction. Also, they described opportunities and obstacles to TAC implementation in the K-12 context. Based on the findings, this study provides implications for the current and future AIED.

First, it must be executed with deliberate long-term educational vision and goals with built-in curriculum systems/structure and pedagogical supports to maximize the potential of TAC and minimize barriers to impactful TAC. Teachers in the study echo the barriers found in other studies related to shallow discussions regarding the long-term visions and goals of AIED (Abdelaziz, 2019; Kim et al., 2022). Although much of the integration of AIED and rationale for TAC has been supported for the sake of personalization of learning, teachers in the study expressed that the current forms of TAC are likely set to improve students’ academic performance in a standardized examination by spoon-feeding pre-specified content, which is driven by the overwhelming attention to performance-driven curriculum design. In this regard, teachers directed us to ponder upon the meaning and purpose of education and the right kind of education we need to pursue TAC. In fact, teachers are expected to realize and implement students’ life-long learning and personalized learning both in the learning process and through outcomes. Additionally, they are expected to incorporate diverse teaching pedagogy such as guided discovery, productive failure, project-based learning, active learning, and deep learning to support students' higher-order thinking rather than rote memorization, while also empowering students' critical and active engagement in learning rather than simply knowing facts in the textbooks. This resonates with the prior research suggesting that we should aspire and focus on leveraging the power of teachers and AI to address genuine education as a ‘wicked problem’, such as promoting students' engagement and inclusion alongside knowledge delivery and formative assessment and feedback (Holmes et al., 2022a). It is, therefore, crucial to acknowledge the varied beliefs, values, and assumptions in the nature of education and establish informed public discussions around why, what, how, and how well TAC reflects a bigger picture/definition of education that takes account of contemporary and future learner needs and context should come into first, otherwise the roles of AI will passively remain ancillary to classroom instruction and scale up the poor pedagogical practice (Abdelaziz, 2019; Holmes et al., 2022a).

In line with this, the findings revealed that the AI technologies that are meant to collaborate with teachers in the educational domain are mainly developed and given by the commercial sector. Although teachers acknowledge the support from and collaboration with the commercial sector, they bring with it fundamental ethical questions over what constitutes appropriate use and rationale by a private company of learner-AI, teacher-AI, and learner-teacher-AI interactions, and what information and interventions they suggest or do not suggest to teachers. These further link to the question of AI loyalty- for whom does the AI system work (Holmes et al., 2022a). In this regard, earlier research argued that it is crucial to set underlying educational rationale for the design and development of AI applications, design and implement robust ethical guidelines, and avoid ethics washing- a growing instrumentalization of ethical language by private tech companies that have been often used to justify their self-interest in design, test, and implement their AI technologies (Knox, 2020; Bietti, 2020). To increase the transparency and trustworthiness of the process and effects of TAC, system developers should be obliged to explicitly align the loyalty of their AI systems and governance structures with the best interests of the learners and teachers as well as others affected by the system. This should include measures to involve multiple stakeholders that fall outside of the commercial sector in the AI tool’s design, procurement, and deployment including learners, teachers, parents, researchers, tech providers, and policymakers. This can serve to ensure multivocality on ethical oversight and clear communication of the expectations of what will be done with both the main and incidental impact of TAC. For the same reasons, there is a critical need for appropriate professional development for teachers as well as for administrators and policymakers, so that they can make informed decisions about which AI systems might be appropriate for their classroom, how those systems broadly work with teachers and students, what they might achieve, their challenges and risks, and what unintended consequences there might be.

Furthermore, teachers in the study characterized and anticipated the diverse possibilities and forms of TAC by taking into account mutual/shared interaction in which teachers and AI systems can augment each other’s strengths, not a unilateral adaptivity in AI systems or teachers separately (Holstein et al., 2020). In this regard, emerging lines of research are beginning to explore the design opportunities of human-AI co-orchestration systems that optimize the complementary strengths of human intelligence and AI, so that the two together behave more intelligently than the two separately (Dellermann et al., 2019), and it envisions collaboration between human and artificial intelligence in a social, technological ensemble (Cukurova et al., 2019). For instance, while teachers work with students, AI systems follow along with teachers’ instruction and adaptively present education resources (e.g., relevant readings, videos, or practice materials) that support their current goals. Alternatively, a system may respond during or after the teacher by adaptively providing feedback on the quality of the instruction (e.g., the clarity of a particular explanation the teacher provides), to help the teachers adjust and improve over time (Holstein et al., 2019; Walker et al., 2014). Meanwhile, teachers can augment the set of instructional moves available to an AI system by either customizing or creating new actions for the system (Heffernan & Heffernan, 2014; Holstein et al., 2019). Yet, there are numerous challenges in conceptualizing and developing such systems. Four challenges are set out by Akata et al. (2020): (1) how do we develop AI systems that work in synergy with humans, (2) how do these systems learn from and adapt to their environment, (3) how do we ensure ethical and responsible behavior, and (4) how can humans and AI share and explain their awareness, goals, and strategies to each other? Developing collaborative, adaptive, responsible, and explainable hybrid intelligent systems is important across many application domains but is especially useful for our thinking about the role of AI in education. The focus on the interplay between learners, teachers, and AI to optimize learning will be central to the successful application of AI in this field. Future research thereby takes a step forward and towards the development of richer instructional theory, practice, and technology for TAC, rather than simply limiting the value of AI development and applications in education to accessing, organizing, and managing data to support teachers’ decision-making. More research needs to be conducted to understand the nature of TAC and the interaction between teachers and AI based on learning and teaching theories, to develop feasible instructional models and strategies to assist teachers in navigating and formulating more meaningful and impactful TAC and to advance AI models and technology to support TAC.

Although the present study presented important findings and implications for structuring and implementing TAC in the context of K-12 instruction, the study recognizes some limitations that should be addressed in a future study. To begin with, the current study was conducted by reviewing the storyboards that have been developed by instructional experts, rather than directly interacting with AI systems, although participating teachers mostly have extensive teaching experience in AIED. This might not have fully revealed participants’ perceptions about TAC. If participants have direct and longitudinal collaborative interactions with the AI systems in the real classroom environment, their perceptions may change over time. Future studies can examine teachers’ experiences and views of actual collaboration with AI systems on various teaching and administrative tasks. One possible way is to conduct a virtual experiment/simulation using AI-embedded virtual and immersive technologies or the Wizard of Oz method, a prototyping method that involves interaction with a mock interface manually controlled by operators (e.g., Kasepalu et al., 2022) to explore how teachers build a sequence of actions TAC on a task given to map a variety of types of TAC. Furthermore, a total number of 30 participating teachers in the study may somewhat limit the generalization of the findings. Future research thereby needs to increase the number of research participants and consider quantitative data collection and analysis.