llm_wiki/raw/教育AI研究/outputs/celik_pages.json

{
  "page_1": "Co-constructing adaptive lesson plans with GenAI: Pre-service \nteachers ’ Intelligent-TPACK and prompt engineering strategies\nIsmail Celika,*\n, Sini Kontkanenb\n, Jari Laruc\n, Alanur Ahsen Dalyancic\naAcademy Research Fellow, Faculty of Education and Psychology, University of Oulu, FI-90014, Oulu, Finland\nbUniversity of Eastern Finland, Faculty of Philosophy, Joensuu, Finland\ncFaculty of Education and Psychology, University of Oulu, FI-90014, Oulu, Finland\nARTICLE INFO\nKeywords:\nAdaptive learning\nTeacher education\nIntelligent-TPACK\nPrompt engineeringABSTRACT\nGenerative Artificial Intelligence (GenAI) technologies present new opportunities for teachers to \ndesign adaptive and student-centered instruction. However, the educational value of GenAI de-\npends not only on technical usage but also on teachers ’ ability to formulate pedagogically \nmeaningful prompts. Prompting strategies are not isolated from teachers ′ prior knowledge and \nskills. Less is known about how pre-service teachers ′ AI-related knowledge influences prompt \nengineering strategies, in turn leading to meaningful adaptive lesson plans. Considering this gap, \nwe design an instructional task for pre-service teachers to generate adaptive lesson plans with the \nhelp of GenAI. Prior to the task, we collected data about their AI-related skills, namely AI literacy \nand Intelligent-TPACK. The prompts were qualitatively analyzed based on the phases of the \nKnowledge Construction (KC) Framework. Then, we explored the pedagogical value of adaptive \nlesson plans through a rubric in terms of three indicators: student agency, adaptive strategies, and \nflexible tools. PLS-SEM analysis revealed that as long as pre-service teachers have AI-specific \ntechnological and pedagogical knowledge, they formulate higher phases of prompts based on \nthe KC framework. Our analysis showed that prompts from higher phases generated more \nadaptive lesson plans in terms of student agency, adaptive strategies, and flexible tools. We also \nfound an indirect effect of Intelligent-TPK on adaptive lesson plans. This study highlights that \neffective prompt engineering is a pedagogical act shaped by teachers ’ knowledge, not merely a \ntechnical command. It also underscores the importance of embedding AI-specific pedagogical \ntraining in teacher education. By conceptualizing prompts as epistemic moves, we offer new \ninsights into how teachers and GenAI can collaborate to produce responsive and inclusive \nlearning experiences.\n1.Introduction\nThe growing emphasis on adaptive learning in education highlights the critical role of teachers ’ knowledge and skills in responding \nto diverse student needs (Wang, Christensen, et al., 2023 ; Wei et al., 2025 ). It is a complex instructional task for teachers to design and \nimplement adaptive learning process (Bernacki et al., 2021 ). This is because it traditionally requires high levels of pedagogical \nexpertise, real-time decision-making, and the ability to offer personalized support ",
  "page_2": "the advent of Generative Artificial Intelligence (GenAI) technologies, it has become increasingly feasible for teachers to design more \npersonalized and responsive learning experiences (Hsia et al., 2025 ). This is because GenAI tools can support differentiation, scaf-\nfolding, and learner agency (Yang & Markauskaite, 2025 ). However, teachers have a responsible role in the ethical and pedagogical use \nof GenAI technologies in designing adaptive teaching practices (Celik, 2023 ; Henderson et al., 2025 ; Tagare et al., 2025 ).\nGenAI tools such as ChatGPT, CoPilot, and Google Gemini, which are now widely used by teachers and students, are considered \nintelligent systems (Ariza et al., 2025 ; Guggemos, 2024 ; Hong et al., 2025 ). For teachers to effectively integrate these tools into their \npractice, it is essential not just to understand how to use but also how to interact with them meaningfully (Guggemos & Seufert, 2021 ; \nWang et al., 2025 ). A key component of this interaction is prompt formulation, which serves as the foundation for successful \nGenAI-assisted teaching (Hsu, 2025 ). The quality of AI-generated outcomes is largely determined by the user’s ability to craft effective \nprompts (Cain, 2024 ; Lee and Palmer, 2025 ). This has led to the emergence of prompt engineering, defined as the strategic design of \nprecise and purposeful inputs to guide AI tools in producing relevant and valuable outputs (Knoth et al., 2024). Well-crafted prompts \ncan significantly enhance GenAI’s ability to support instructional tasks such as lesson planning. However, there is still limited un-\nderstanding of how to optimize prompt use in educational contexts. To strengthen teacher and GenAI collaboration, further research is \nneeded to explore prompt engineering and identify strategies that provide the most effective support for teachers’ professional work \n(Celik et al., 2022 ; Cress & Kimmerle, 2023 ; Park & Choo, 2025 ).\nMoreover, there is limited understanding of how these strategies intersect with teacher professional knowledge to ethically inte-\ngrate GenAI into education. Indeed, ethical GenAI integration requires the teacher’s knowledge not only for technical use of GenAI \ntools but also pedagogical reasoning and ethical considerations (Guggemos, 2024 ). In this regard, teacher education institutions must \nequip future teachers not only with AI literacy but also with pedagogical competencies specific to AI-integrated instruction (Celik et al., \n2024 ; Cheah et al., 2025 ). Pre-service teachers need structured opportunities to practice prompt engineering and critically evaluate \nGenAI outputs (Laru et al., 2025 ). In this way, they can align such outputs with learning goals and student needs. To better understand \nand support pre-service teachers’ ethical and pedagogical use of GenAI, the use of theoretically validated frameworks can provide \nessential guidance in this process.\nAn updated framework of Technological Pedagogical Content Knowledge (TPACK) (",
  "page_3": "In sum, teachers should consider student agency, scaffolding strategies, and technology use as flexible tools when designing \nadaptive learning environments. These three elements work together to ensure that learners are both supported and empowered, while \ninstruction remains responsive to diverse needs. The current study addresses adaptive learning through these concepts by examining \nhow lesson plans incorporate opportunities for student choice, the presence of scaffolding strategies, and the integration of flexible \ntechnological tools.\n2.2. Intelligent-TPACK\nThe Intelligent TPACK framework is a contemporary extension of the TPACK framework developed by Mishra and Koehler (2006). \nIt was proposed in response to the growing integration of AI technologies into educational settings. This framework aims to define the \nspecific professional knowledge for teachers to ethical and pedagogical use of AI-enhanced tools (Celik, 2023 ). Fig. 1shows TPACK and \nIntelligent-TPACK frameworks.\nThe Intelligent TPACK framework emphasizes the importance of teacher competencies not only in using AI tools pedagogically \nmeaningful and ethically responsible. It has five AI specific following components: Technology Knowledge (Intelligent- TK), Intelligent \nTechnological Content Knowledge (Intelligent-TCK), Intelligent Technological Pedagogical Knowledge (Intelligent -TPK), Intelligent \nTechnological Pedagogical and Content Knowledge (Intelligent-TPACK).\nIntelligent-TK , which refers to knowledge about how AI systems operate and how to use them effectively; Intelligent-TPK , which \nfocuses on the pedagogical application of AI tools to support student-centered instruction, differentiation, and formative assessment; \nand Intelligent-TCK , which involves the use of AI tools to represent and explore subject-specific content. The integration of these do-\nmains results in Intelligent-TPACK, which represents a teacher ’s ability to design, implement, and evaluate instructional activities that \neffectively combine AI technologies with appropriate pedagogical strategies and subject matter knowledge. Importantly, the frame -\nwork addresses the need for critical AI literacy, including awareness of ethical concerns such as bias, transparency, data privacy, and \nthe role of human judgment in AI-supported decision-making. Teachers are expected to engage in human-AI co-agency, where the \nteacher retains pedagogical control while leveraging AI tools to augment learning experiences.\nOverall, the Intelligent TPACK framework provides a comprehensive model for preparing educators to navigate the complexities of \nAI-enhanced education by aligning AI capabilities with pedagogical goals and learners ’ needs in ethically grounded ways (Celik, \n2023 ).\n2.3. AI literacy\nTo navigate the increasing presence of AI-based technologies in society and education, individuals must possess a broad and in-\ntegrated set of knowledge, skills, and values that enable both effective use and ethical engagemen",
  "page_4": "recognize various forms of new literacies, including media, information, digital, and AI literacies, which demand additional cognitive, \nethical, and operational skills (Griffin & Care, 2014 ; Kong et al., 2021 ; Laru et al., 2025 ). Among these, AI literacy is increasingly seen \nas essential for full participation in a society shaped by intelligent systems.\nDespite its importance, AI literacy remains an evolving construct with limited consensus around a comprehensive definition (Chiu \net al., 2024 ). Nonetheless, prevailing views emphasize that AI literacy involves not only understanding how to operate AI tools but also \nrecognizing their ethical implications, such as fairness, bias, transparency, and accountability (Steinbauer et al., 2021 ). This is \nespecially pertinent because ethical decisions in AI development are often made by technical experts without sufficient attention to \nend-users’ perspectives, capacities, or values (Holmes et al., 2021 ). Hence, fostering ethical sensitivity among AI users is crucial for \nensuring responsible use and promoting socially inclusive outcomes.\nTo provide conceptual clarity, Ng et al. (2021) conducted a systematic and exploratory review using Bloom’s taxonomy as an \norganizing framework. Their model defines AI literacy as comprising three progressive levels: knowledge and comprehension, \napplication, and evaluation and development. This model emphasizes the integration of data science and computational thinking \nwithin a multidisciplinary framework, highlighting that ethical awareness, particularly regarding fairness and transparency in AI \nsystems, is a fundamental component of AI literacy (Ng et al., 2021 ). Importantly, these skills are not only relevant for developers but \nare also essential for educators, students, and everyday users who interact with AI in personal and professional contexts (Long & \nMagerko, 2020 ). AI literacy also requires an awareness of AI’s limitations, such as its black-box nature, potential for biased outputs, \nand susceptibility to misinformation (Steinbauer et al., 2021 ). Thus, individuals must be equipped to critically interrogate \nAI-generated information and avoid overreliance on its outputs. As Casal-Otero et al. (2023) argue, such competencies empower users \nto adopt a reflective and informed stance toward the expanding ecosystem of AI-driven tools, including large language models.\nTo guide this study, we draw upon the AI literacy framework proposed by Wang et al. (2023) , which defines AI literacy as the ability \nto recognize, use, evaluate, and ethically engage with AI technologies. This framework delineates four key dimensions: Awareness, \nwhich involves recognizing the presence and relevance of AI technologies and developing a conceptual understanding of how they \nfunction. It is viewed as a prerequisite cognitive process that underlies meaningful AI use (Wang et al., 2023 ). Usage, which refers to the \npractical and operational ability to utilize AI tools for v",
  "page_5": "teacher–GenAI collaboration. We recognize that such interactions are not epistemically equivalent. In other words, unlike a peer, \nGenAI has no capacity to intentionally produce, justify, and take responsibility for knowledge claims, which is described as epistemic \nagency (Celik et al., 2025 ;Stroupe, 2014 ). However, functionally, GenAI responses can play a role similar to peer contributions by \nintroducing new information, generating dissonance, and prompting negotiation of meaning (An et al., 2025 ; Zhang & Wang, 2025 ). In \nthis sense, although AI cannot be considered an epistemic equal, its outputs serve as dialogic triggers that scaffold pre-service teachers’ \nknowledge construction processes. Thus, the KC framework remains useful for capturing the human-driven but AI-supported cycles of \nrefinement, testing, and integration of knowledge.\nIn this study, the “knowledge” constructed through prompting refers specifically to the co-construction of instructional knowledge in \nthe form of adaptive lesson plans. While pre-service teachers relied on their existing pedagogical and technological knowledge (e.g., \nIntelligent-TPACK, AI literacy), the KC framework was employed to examine how this knowledge was enacted and transformed \nthrough dialogic interaction with GenAI. Thus, the framework allowed us to move beyond viewing prompts as technical commands \nand instead conceptualize them as epistemic moves through which teachers and GenAI collaboratively shaped lesson design. The \noutcome of this process was not only the final lesson plan but also the iterative refinement of pedagogical reasoning. In their in-\nteractions with AI, teachers negotiated meaning, resolved dissonance, and tested synthesis.\nThe KC framework outlines five phases: (1) the sharing and comparison of information; (2) the discovery of dissonance; (3) the \nnegotiation of meaning; (4) the testing and modification of proposed synthesis; and (5) the agreement and application (Gunawardena \net al., 1997 ). These phases progress from surface-level exchange to deep cognitive engagement and knowledge co-construction. We \nfocused on four phases of the KC framework, excluding the final phase. The rationale behind this is that the final phase requires the \nimplementation of co-conducted knowledge (i.e. lesson plan) in a practical context.\nSharing and comparing information (Phase 1) involves participants introducing ideas, facts, or observations without attempting to \nreconcile differences. Discovery and exploration of dissonance (Phase 2) marks the point at which learners notice contradictions, gaps, \nor inadequacies in the information. Prompts in this phase may question the relevance or appropriateness of AI-generated responses (e. \ng., “This doesn’t seem age-appropriate. Is there a simpler way to explain it?”). Negotiation of meaning and co-construction of \nknowledge (Phase 3) entails efforts to resolve dissonance by integrating perspectives, elaborating ideas, or formulating interpretatio",
  "page_6": "research suggests that these knowledge domains —technical, pedagogical, and ethical —are foundational for responsible and effective \nGenAI use in education (Celik et al., 2022; Ng et al., 2021 ). However, little is known about how these competencies interact with \nprompting behavior and influence GenAI outputs. By integrating AI literacy (Wang et al., 2022) and Intelligent-TPACK frameworks \n(Celik et al., 2022), this study seeks to uncover whether more knowledgeable pre-service teachers can more effectively utilize GenAI \ntools toward the generation of contextually adaptive and pedagogically sound learning experiences. To achieve these research ob-\njectives, we defined the following research questions (RQs). \n≡RQ1-Obj1: How do pre-service teachers employ prompt engineering strategies to generate adaptive lesson plans, and what \ninstructional features emerge from these plans?\n≡RQ2-Obj2: What are the associations of pre-service teachers ’ AI literacy and Intelligent-TPACK skills with their prompt engineering \nstrategies and AI-generated adaptive lesson plans?\nThese research questions are tested through the research model illustrated below in Fig. 2.\n4.Methods\n4.1. Participants\nA total of 120 pre-service teachers were in the first phase of their five-year MA degree in a two-stage teacher education programme \n(Bachelor 3 years, Master 2 years) in Finland. Among these, 100 were enrolled in the classroom teacher education program and 20 \nwere enrolled in the special education teacher education program.\n4.2. Task: generation of adaptive lesson plan\nIn order to investigate the ability of pre-service teachers to collaborate with GenAI tool for adaptive lesson planning, participants \nwere asked to complete a structured task that simulated a real-life instructional scenario (see Appendix A). They were asked to imagine \nthat they were primary school teachers preparing a 45-min science lesson for a 5th-grade classroom consisting of 28 pupils (15 boys \nand 13 girls with an average age of 10). The learning objectives were to explain the stages of the water cycle (evaporation, \ncondensation, precipitation, and collection) and to emphasize the importance of water in the environment. Participants were \ninstructed to use a GenAI tool (Copilot) to generate teaching activities tailored to learning goals and pupil characteristics, to support \ntheir planning.\nCrucially, the task emphasized designing adaptive or personalized instructional strategies by encouraging pre-service teachers to \nrequest support from the AI that considered pupil diversity, engagement, and differentiated learning needs. For this task, participants \nworked in pairs, resulting in 60 lesson plans being submitted in total. Each pair was given 15–20 min to interact with the AI and refine \ntheir instructional approach using iterative prompts. Afterwards, participants submitted a final AI-generated lesson plan justifying \ntheir choices. To better understand this justification, we also asked participant",
  "page_7": "4.3. Data collection tools\n4.3.1. Intelligent-TPACK\nThe Intelligent-TPACK Scale (Celik, 2023) was used to measure pre-service teachers ’ professional knowledge for pedagogically and \nethically integrating AI-based tools into instruction. The scale consists of 27 items distributed across five dimensions: Intelligent \nIntelligent-TK (5 items, α ˆ.856), Intelligent-TPK (7 items, α ˆ.858), Intelligent-TCK (4 items, α ˆ.868), Intelligent TPACK (7 items, \nα ˆ.895), and Ethics (4 items, α ˆ.864). Each item was rated on a 7-point Likert scale ranging from 1 (strongly disagree) to 7 (strongly \nagree). All subscales demonstrated strong internal consistency, indicating the reliability of the scale for assessing AI-specific peda-\ngogical and ethical knowledge in teacher education contexts.\n4.3.2. AI literacy\nTo evaluate participants ’ understanding of, and ability to use, recognize and evaluate AI technologies, we used the AI Literacy Scale \n(AILS), which was developed by Wang, Rau and Yuan (2022). The scale is grounded in a four-factor model of AI literacy, encompassing \nawareness, usage, evaluation, and ethics. The final version of the AILS consists of 12 items (three per construct), which are rated on a 7- \npoint Likert scale. Reliability scores for the subscales were satisfactory: awareness (α ˆ.73), usage (α ˆ.75), evaluation (α ˆ.78), and \nethics (α ˆ.73). The overall scale demonstrated strong internal consistency (α ˆ.83). Confirmatory factor analysis supported the four- \nfactor structure, and model fit indices (e.g., CFI ˆ0.99, RMSEA ˆ0.01) confirmed the robustness of the theoretical model. This scale is \na validated instrument for measuring individuals ’ general AI literacy across cognitive, operational, evaluative, and ethical dimensions.\n4.3.3. Adaptive lesson plan rubric\nThe final lesson plan selected by each pair of participants was evaluated using an adaptive lesson plan rubric. For this rubric and its \ndimensions, we used prior validated rubrics to technologically and pedagogically assess teachers ′ instructional activities (Harris & \nHofer, 2011 ; Koh, 2013 ; Kopcha et al., 2014 ). The rubric consists of three key dimensions: (1) adaptive strategies and scaffolding, (2) \nstudent agency, and (3) flexible content and tools. The first dimension focused on how the plan incorporated responsive instructional \nstrategies and provided scaffolding for learning based on pupils ’ varying needs. The second dimension addressed the level of student \nautonomy and opportunities for personalized learning pathways. The third dimension evaluated the flexibility and appropriateness of \nlearning materials and tools for supporting differentiated instruction. This rubric enabled us to systematically assess the adaptive \nqualities embedded in the AI-generated content, as influenced by the pedagogical framing in pre-service teachers ’ prompts Scoring of \nthe lesson plans ranged from one to four (1: Limited, 2: Basic, 3: Proficient, 4: Excellent).\n4.3.4. Prompt engine",
  "page_8": "4.4. Data analysis\nTo answer RQ1, two different coding approaches were applied. First, the Copilot outcome selected by pre-service teachers from \nCopilot was evaluated using the adaptive lesson plan rubric. Second, all prompts generated by pre-service teachers during the task \nwere coded according to the KCP. Thus, while only the final Copilot output was analyzed with the rubric, each teacher prompt was \nclassified based on the KC framework. To make sure the results of the coding were reliable, the first and second authors of the study \nrandomly picked about 15 % of the lesson plan (N ˆ4) and prompt engineering (N ˆ4) data from the whole dataset and coded them \nseparately (Fleiss et al., 2013 ). We calculated the Cohen ’s Kappa score between the two coders as 0.65 for prompt engineering and 0.72 \nfor the lesson plan. According to Landis and Koch (1977) , values between 0.61 and 0.80 indicate substantial agreement in inter-rater \nreliability. Therefore, the coding consistency between the two raters was sufficiently strong to support further analyses.\nTo examine the relationships among the research variables (RQ2), we applied partial least squares structural equation modelling \n(PLS-SEM) (Ringle & Sarstedt, 2016 ). This method is well suited for testing complex models and the interconnections between con-\nstructs, which are operationalized through observed variables (Henseler et al., 2016 ). In addition, it places fewer restrictions on sample \nsize. The PLS-SEM procedure follows two main stages. First, the measurement model is assessed to establish the reliability and validity \nof the constructs (Ringle & Sarstedt, 2016 ). Next, the structural model is analyzed to investigate the relationships among the constructs \nwithin the research model (Henseler et al., 2009 ).\nThe measurement model was built if convergent validity was ensured. To check this validity, three parameters were examined \nagainst established thresholds: (i) item reliability, evaluated through factor loadings (F0.70), (ii) composite reliability for each \nconstruct (F0.70), and (iii) average variance extracted (AVE F0.50). Among these, the average value extracted is considered a \nparticularly important criterion for convergent validity (Fornell & Larcker, 1981 ).\nIn the research model, the relationships were tested by estimating standardized regression weights (betas, β), including both direct \nand indirect effects among the constructs. Second, the structural model was examined by assessing path coefficients, their significance \nthrough 1000 bootstrapping, the coefficient of determination (R2), effect sizes (f2), and predictive relevance (Q2). The model ’s \napproximate fit was further evaluated using the standardized root mean square residual (SRMR), which was below the recommended \nthreshold of 0.08, indicating acceptable fit. The analyses were conducted using SmartPLS 4.\n5.Results\n5.1. The analysis of prompting strategies and lesson plans for adaptive learning (RQ1)\nAfter the task w",
  "page_9": "identified gaps or misalignments with suggestions generated by AI. Finally, KCP 4: Modification of Proposed Synthesis (15.1 %) was \nthe least represented but nonetheless present, with prompts focused on adapting, improving, or evaluating previously suggested \noutputs. These findings suggest that while pre-service teachers initially used GenAI to retrieve general information, they increasingly \nmoved toward more nuanced, goal-directed instructional design interactions as the prompt sequence progressed. During the coding \nprocess, the highest KC phase of prompts (HKCP) was also defined.\nTable 2shows the analysis of adaptive learning dimensions in lesson plans created by pre-service teachers. The highest overall score \nwas for adaptive strategies/scaffolding dimensions, indicating that pre-service teachers frequently incorporated moderate to high \nlevels of instructional support into their lesson plans. For example, the lesson plans included group work and visual aids. Yet, there \nwere some cases showing hands-on experimentation and teacher-guided station rotations. The student agency dimension indicates \nthat, although many lesson plans included some level of pupils ’ engagement and participation, only a few encouraged pupils to make \nautonomous choices or reflect on their learning. High-agency examples allowed pupils to select learning formats and presentation \ntools, demonstrating greater ownership of the learning process. The ’flexible content and tools ’ dimension had a similar average, with \nmost lesson designs using at least two content formats (for example, drawing and video) or different learning materials. The best plans \nlet students use different types of media to show what they understand, while the more traditional plans only let students use the media \nthat the teacher chooses.\n5.2. The interplay of AI literacy, Intelligent-TPACK, prompt engineering strategies, and lesson plans for adaptive learning (RQ2)\nPrior to the PLS-SEM analysis, we defined the highest KC phase of prompts (HKCP) among four consecutive KC phases. Next, \nPearson ’s correlations among AI literacy, Intelligent-TPACK, prompt engineering strategies and lesson plan for adaptive learning were \ncalculated. Table 3displays the results.\nAs presented in Table 3, all components of the Intelligent-TPACK framework correlated with AI literacy dimensions. A positive \ncorrelation was found between the number and highest phase of prompts. We observed a moderate and positive correlation between \nthe usage dimension of AI literacy and number prompts and the highest phase of prompts. Similarly, TK, TPK of Intelligent TPACK were \npositively correlated with number prompts and highest level of prompts. Further, the factors of adaptive lesson plan were correlated \nwith number prompts and highest level of prompts.\nTable 4presents the results of the measurement model analysis. All factor loadings exceeded the recommended threshold of 0.70, \ndemonstrating satisfactory item reliability. The ",
  "page_10": "Table 3 \nBivariate correlations among the research variables.\n2 3 4 5 6 7 8 9 10 11 12 13 14\nTK (1) 0.72** 0.76 \n**0.69 \n**0.81** 0.66** 0.53 0.61** 0.50** 0.42** 0.34** 0.12 0.18 0.15\nTCK (2) – 0.78** 0.71** 0.64** 0.51** 0.49** 0.41** 0.39** 0.33** 0.22** 0.13 0.11 0.10\nTPK (3) ​ – 0.69** 0.71** 0.50** 0.61** 0.59** 0.48** 0.34** 0.43** 0.31** 0.29** 0.33**\nTPCK (4) ​ ​ – 0.68** 0.52** 0.43** 0.58** 0.61** 0.22** 0.18** 0.10 0.11 0.09\nEthics (5) ​ ​ ​ – 0.47** 0.74** 0.78** 0.71** 0.03\u00000.01 0.03 0.08 0.11\nUsage (6) ​ ​ ​ ​ – 0.78** 0.70** 0.67** 0.45** 0.43** 0.02 0.05 0.10\nAwareness (7) ​ ​ ​ ​ ​ – 0.54** 0.61** 0.31** 0.28** 0.10 0.09 0.05\nEvolution (8) ​ ​ ​ ​ ​ ​ – 0.52** 0.07 0.03 0.02 0.07\u00000.09\nEthics (AL) (9) ​ ​ ​ ​ ​ ​ ​ – 0.01 0.03 0.02 0.02 0.05\nNP (10) ​ ​ ​ ​ ​ ​ ​ ​ – 0.51** 0.39** 0.41** 0.46**\nHKCP (11) ​ ​ ​ ​ ​ ​ ​ ​ ​ – 0.40** 0.47** 0.53**\nAdaptive (12) ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ – 0.56** 0.64**\nAgency (13) ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ – 0.61**\nFlexible (14) ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ –\n**Significant correlation at the 0.01 level (two-tailed).\nHKCP: Highest KC phase of prompts; NP: Number of prompts.I. Celik et al.                                                                                                                                                                                                            Computers  & Education  241 (2026)  105485  \n10 ",
  "page_11": "Larcker, 1981 ).\nWe conducted a PLS-SEM analysis to examine both direct and indirect effects among the research variables. The initial model \nincluded all paths from AI literacy and Intelligent-TPACK to prompting strategies and adaptive learning dimensions. After removing \ninsignificant paths, the final model demonstrated acceptable fit, with an SRMR value of 0.043, below the recommended threshold of \n0.08 (Hair et al., 2022). The R2 values of the endogenous constructs ranged from 0.32 to 0.54, indicating moderate to substantial \nexplanatory power. Predictive relevance (Q2) values were greater than zero for all endogenous variables, confirming the model ’s \npredictive validity. Fig. 5illustrates the final model.\nThe number of prompts was positively associated with Intelligent-TK (⠈0.27) and Intelligent-TPK (⠈0.26). Additionally, the \nnumber of prompts showed a positive relationship between participants ’ awareness (⠈0.23) and usage of AI (⠈0.29). The phase of \nprompts was also significantly and positively associated with TK (⠈0.30), TPK (⠈0.32). Further, AI-related awareness (⠈0.28) \nand usage (⠈0.30) are related to the phase of prompts. Moreover, the number and phase of prompts were found to be interrelated (β \nˆ0.35).\nPhase of prompts are associated with flexible tools (⠈0.31), student agency (⠈0.27), and adaptive strategies (⠈0.25). Lastly, \nthe PLS-SEM analysis yielded an indirect effect between TPK and both adaptive strategies (⠈0.17; 95 % CI [0.06, 0.29]) and student \nagency (⠈0.20; 95 % CI [0.09, 0.33]). This indirect effect was through the phase of prompts.\n6.Discussion\nThe integration of GenAI into education has given rise to a number of critical questions concerning the knowledge and strategies \nthat teachers will require in order to collaborate effectively with AI systems. While emerging literature highlights the technological and \npedagogical potential of GenAI tools to support personalized learning, differentiation, and instructional design (Zhang et al., 2025 ), far \nless is known about how pre-service teachers ’ prior knowledge shapes the quality of AI interaction, particularly through prompt \nengineering. Existing research frequently treats prompting as a procedural or syntactic task, neglecting its pedagogical and epistemic \ndimensions (ElSayary et al., 2025 ; Lee & Palmer, 2025 ). This creates a substantial gap in understanding how prompt formulation \nreflects deeper cognitive engagement and professional reasoning in educational contexts. Our study addresses this gap by offering a \nnovel perspective, which is to prompt engineering to teachers ’ AI-specific pedagogical and technological knowledge and skills. The \nlatter is termed Intelligent-TPACK and AI literacy.\nIn accordance with this, an evaluation was conducted to ascertain how these knowledge domains inform the construction of \nadaptive, student-centered lesson plans with GenAI. The application of the KC framework to the analysis of prompt strategi",
  "page_12": "demonstrates that prompts are not neutral inputs, but epistemic moves grounded in teachers’ instructional goals. The findings of our \nstudy provide empirical evidence that higher-phase prompts, reflecting negotiation and synthesis, contributed to the generation of \nmore adaptive lesson plans. In turn, these higher-phase prompts are predicted by pre-service teachers’ AI-related knowledge. This \nstudy contributes to current debate by conceptualizing prompt engineering as a pedagogically grounded practice and emphasizing the \nnecessity of integrating it into teacher education programs.\n6.1. The features of prompting strategies and lesson plans\nThe findings indicated that pre-service teachers frequently employed prompts at the negotiation of meaning (KCP3) level, sug-\ngesting a strong orientation towards pedagogical inquiry as opposed to mere information retrieval. This finding suggests that when \nconfronted with a task necessitating adaptive design, many pre-service teachers instinctively endeavour to co-construct meaning with \nGenAI by posing clarifying questions, requesting elaborations, or seeking instructional alternatives.\nIt is noteworthy that prompts in the sharing information (KCP1) and discovery of dissonance (KCP2) stages were also common, \nreflecting early stages of interaction, where participants probed GenAI for foundational explanations or identified gaps in AI-generated \ncontent. However, the comparatively limited utilization of KCP4 (Modification of Synthesis) and KCP5 (Application) indicates a \nrequirement for the scaffolding of pre-service teachers to facilitate deeper engagement in iterative and evaluative prompting.\nPrompt engineering has recently emerged as a widely discussed skill in both educational and professional contexts, especially with \nthe increasing accessibility of GenAI tools like ChatGPT and Copilot. However, despite this growing interest, many end users, \nparticularly novice educators, lack a conceptual understanding of what prompt engineering entails as a cognitive and pedagogical \nprocess (Knoth et al., 2024; Walter, 2024 ). This disconnection frequently leads to surface-level interactions with GenAI, where users \nfocus on command syntax rather than strategic input formulation. In this regard, the KC framework offers a valuable lens for examining \nhow novice users engage with GenAI. By situating prompts within progressive phases of epistemic engagement, the KC framework \nenables educators and researchers to evaluate the instructional quality embedded in prompt sequences (Gunawardena et al., 1997 ; \nLucas et al., 2014 ).\nOur study also demonstrates that the analysis of knowledge construction phases in the prompting can reveal not only the extent to \nwhich content is merely being retrieved, but also the presence of more sophisticated forms of pedagogical reasoning, such as the \nformulation of questions, the integration of ideas, and the refinement of concepts. In the present study, the KC framework was found to \nbe",
  "page_13": "The findings of this study demonstrate that prompting is not a neutral or mechanical act; rather, it is a pedagogically embedded \nprocess. Teachers who have a firm grasp on both the capabilities of AI and the instructional goals to be achieved are better positioned to \nco-construct effective and adaptive learning designs with GenAI. This finding lends support to the emerging calls for the integration of \nprompt engineering into teacher education curricula as a component of AI literacy and Intelligent-TPACK.\n7.Conclusion\nPre-service teachers’ prompting with GenAI worked as an ongoing process shaped by what they already knew about teaching. \nInstead of using one-time commands, they adjusted their prompts based on the AI’s earlier answers. This shows that prompting is a \nthoughtful activity connected to planning and teaching goals. As their interaction with GenAI continued, their prompts became more \nfocused and detailed. This pattern suggests that teachers were not just asking questions but building a kind of conversation with the \ntool to improve their lesson ideas.\nThe prompts provided by pre-service teachers using GenAI functioned as a continuous process, shaped by their previous teaching \nexperiences. In lieu of utilizing one-time commands, a decision was made to adjust the prompts in accordance with the AI’s prior \nresponses. This finding indicates that prompting is a deliberate activity associated with the formulation and execution of teaching \nobjectives. As their interaction with GenAI continued, their prompts became more focused and detailed. This pattern suggests that \nteachers were not merely posing questions; rather, they were employing the tool to facilitate a form of dialogue, thereby enhancing \ntheir lesson plans.\nFurthermore, the technological knowledge of pre-service teachers in artificial intelligence clearly affected the quality of their lesson \nplans. Those with stronger knowledge in using AI for teaching created prompts that led to more flexible, engaging, and personalized \nplans. Their effective use of GenAI shows their understanding of technology helps guide the tool in useful ways. Prompting became a \nway to apply their knowledge to practice. This underscores the significance of incorporating AI-related pedagogical competencies into \nteacher education curricula, thereby equipping future educators with the skills to utilize these tools in a meaningful and responsible \nmanner.\n8.Limitations and future research\nWhilst the present study provides valuable insights into the prompt engineering strategies employed by pre-service teachers and the \npedagogical use of GenAI tools, it is important to acknowledge the study’s limitations. Firstly, the sample consisted exclusively of \nFinnish pre-service teachers, which may limit the generalizability of the findings to other cultural or educational contexts. Future \nstudies could expand the sample to include teachers from diverse educational systems to explore potential contextual differen",
  "page_14": "CRediT authorship contribution statement\nIsmail Celik: Writing – review & editing, Writing – original draft, Visualization, Validation, Methodology, Investigation, Formal \nanalysis, Data curation, Conceptualization. Sini Kontkanen: Writing – review & editing, Formal analysis, Resources, Jari Laru: \nWriting – review & editing, Writing – original draft, Formal analysis. Alanur Ahsen Dalyanci: Writing – review & editing, Writing – \noriginal draft.\nAcknowledgments\nThis study was conducted as a part of first author ′s academy research fellow project entitled “Power T/A” with the number 363521 \nsupported by Research Council of Finland. This work was also supported (third author) by the Strategic Research Council (SRC) \nestablished within the Academy of Finland under Grants #352859 and #352871\nAppendix A \nTeacher-GenAI Collaboration Task\nTask Description: Imagine you are a primary school teacher, and you need some support from GenAI-based tools (e.g., ChatGPT, \nCopilot) for your teaching.\nBefore your instruction, you will use Copilot for planning your lesson. Copilot will assist you in “organizing teaching activities \nconsidering students ′ characteristics ” for achieving a learning goal during the lesson.\nLesson Duration: 45 min\nClassroom Context: 28 students (15 male, 13 female)\nStudent age: 10 years (Average)\nGrade Level: 5th grade\nLearning Goals: By the end of the lesson, students should be able to: \n≡Understand and explain the stages of the water cycle (evaporation, condensation, precipitation, and collection).\n≡Identify the importance of water in the environment and how it moves between different stages.\nTo have effective teaching experience, you need suggestions from Copilot for some teaching activities. \nWe kindly ask you to use Copilot to get some support. You can ask a couple of questions (or prompts), till you are happy with the final suggestion. \nFor this task you will have 15–20 min. After you complete your task, please answer the following questions.\nWhich questions (prompts) did you use to communicate with Copilot?\nPrompt1: \nPrompt2: \nPrompt3: \nPrompt[N]:\nWhat was the final outcome you liked from Copilot?\nPlease just copy the final outcome: \n——Please elaborate why you have chosen this outcome:\nData availability\nData will be made available on request.\nReferences\nAriza, J.˘A., Restrepo, M. B., & Hern ˘andez, C. H. (2025). Generative AI in engineering and computing education: A scoping review of empirical studies and educational \npractices. IEEE Access . https://doi.org/10.1109/ACCESS.2025.3541424\nArthars, N., Dollinger, M., Vigentini, L., Liu, D. Y. T., Kondo, E., & King, D. M. (2019). Empowering teachers to personalize learning support: Case studies of teachers ’ \nexperiences adopting a student-and teacher-centered learning analytics platform at three Australian universities. Utilizing learning analytics to support study success , \n223–248. https://doi.org/10.1007/978-3-319-64792-0_13\nAn, S., Zhang, S., Guo, T., Lu, S., Zhang, W., & Cai, Z",
  "page_15": "Celik, I., Gedrimiene, E., Siklander, S., & Muukkonen, H. (2024). The affordances of artificial intelligence-based tools for supporting 21st-century skills: A systematic \nreview of empirical research in higher education. Australasian Journal of Educational Technology, 40(3), 19–38.\nCelik, I. (2023). Towards Intelligent-TPACK: An empirical study on teachers ’ professional knowledge to ethically integrate artificial intelligence (AI)-based tools into \neducation. Computers in Human Behavior, 138, 107468. https://doi.org/10.1016/j.chb.2022.107468 .\nCelik, I., Dindar, M., Muukkonen, H., & Jarvel a, S. (2022). The promises and challenges of artificial intelligence for teachers: A systematic review of research. \nTechTrends, 66(4), 616–630. https://doi.org/10.1007/s11528-022-00715-y .\nCelik, I., Muukkonen, H., & Siklander, S. (2025). Teacher –Artificial Intelligence (AI) interaction: The role of trust, subjective norm and innovativeness in \nTeachersacceptance of educational chatbots. Policy Futures in Education , 14782103251348551 .\nCheah, Y. H., Lu, J., & Kim, J. (2025). Integrating generative artificial intelligence in K-12 education: Examining teachers ’ preparedness, practices, and barriers. \nComputers and Education: Artificial Intelligence, 8, Article 100363. https://doi.org/10.1016/j.caeai.2025.100363\nChiu, T. K., Ahmad, Z., Ismailov, M., & Sanusi, I. T. (2024). What are artificial intelligence literacy and competency? A comprehensive framework to support them. \nComputers and Education Open, 6, Article 100171. https://doi.org/10.1016/j.caeo.2024.100171\nCress, U., & Kimmerle, J. (2023). A systemic and cognitive view on collaborative knowledge building with generative AI. Computers in Human Behavior, 147, Article \n107861. https://doi.org/10.1016/j.chb.2023.107861\nElSayary, A., Kuhail, M. A., & Hojeij, Z. (2025). Examining the role of prompt engineering in utilizing generative AI tools for lesson planning: Insights from teachers ’ \nexperiences and perceptions. Human Behavior and Emerging Technologies, 2025 (1), Article 9986139. https://doi.org/10.1155/hbe2/9986139\nFleiss, J. L., Levin, B., & Paik, M. C. (2013). Statistical methods for rates and proportions . John Wiley & Sons. \nFornell, C., & Larcker, D. F. (1981). Evaluating structural equation models with unobservable variables and measurement error. Journal of Marketing Research, 18(1), \n39–50. https://doi.org/10.1177/002224378101800104\nGriffin, P., & Care, E. (Eds.). (2014). Assessment and teaching of 21st century skills: Methods and approach . Springer . \nGuggemos, J. (2024). On the predictors of computational thinking and its relationship with artificial intelligence. In E. Sharples, C. L. Saxena, & R. Kumar (Eds.), \nArtificial intelligence for supporting human cognition and exploratory learning in the digital Age (pp. 179–201). Nature Switzerland: Springer . \nGuggemos, J., & Seufert, S. (2021). AI literacy and its importance in education: Research findings and implications. Computers and Educ",
  "page_16": "Robertson, J. (2024). Game changers: A generative AI prompt protocol to support strategy and design in human–AI collaboration. Business Horizons, 67(5), 521–531. \nhttps://doi.org/10.1016/j.bushor.2024.02.003\nScardamalia, M., & Bereiter, C. (2006). Knowledge building: Theory, pedagogy, and technology. In R. K. Sawyer (Ed.), The Cambridge handbook of the learning sciences \n(pp. 97–115). Cambridge University Press. https://doi.org/10.1017/CBO9780511816833.008. \nShemshack, A., Kinshuk, & Spector, J. M. (2021). A comprehensive analysis of personalized learning components. Journal of Computers in Education, 8(4), 485–503. \nhttps://doi.org/10.1007/s40692-021-00188-7\nStahl, G. (2000). A model of collaborative knowledge-building. Proceedings of the Fourth international conference of the learning sciences. Mahwah, NJ: Erlbaum. https:// \npublic.websites.umich.edu/~icls/proceedings/abstracts/ab70.html . \nStahl, G. (2017). Group practices: A new way of understanding knowledge building. Springer. https://doi.org/10.1007/s11412-016-9227-5\nSteinbauer, G., Kandlhofer, M., Chklovski, T., Heintz, F., & Koenig, S. (2021). A differentiated discussion about AI education K-12. KI-Künstliche Intelligenz, 35(2), \n131–137. https://doi.org/10.1007/s13218-021-00724-8\nStroupe, D. (2014). Examining classroom science practice communities: How teachers and students negotiate epistemic agency and learn science-as-practice. Science \nEducation, 98(3), 487–516. https://doi.org/10.1002/sce.21112\nTagare, D., Karki, T., & Yu, W. (2025). K-12 teachers’ ethical competencies for AI Literacy: Insights from a systematic literature review. Computers & Education. , \nArticle 105435. https://doi.org/10.1016/j.compedu.2025.105435\nVan de Pol, J., Volman, M., & Beishuizen, J. (2010). Scaffolding in teacher–student interaction: A decade of research. Educational Psychology Review, 22(3), 271–296. \nhttps://doi.org/10.1007/s10648-010-9127-6\nWalter, Y. (2024). Embracing the future of Artificial Intelligence in the classroom: The relevance of AI literacy, prompt engineering, and critical thinking in modern \neducation. International Journal of Educational Technology in Higher Education, 21(15). https://doi.org/10.1186/s41239-024-00448-3\nWang, S., Christensen, C., Cui, W., Tong, R., Yarnall, L., Shear, L., & Feng, M. (2023). When adaptive learning is effective learning: Comparison of an adaptive learning \nsystem to teacher-led instruction. Interactive Learning Environments, 31(2), 793–803. https://doi.org/10.1080/10494820.2020.1808794\nWang, B., Rau, P. L. P., & Yuan, T. (2023). Measuring user competence in using artificial intelligence: Validity and reliability of artificial intelligence literacy scale. \nBehaviour & Information Technology, 42(9), 1324–1337. https://doi.org/10.1080/0144929X.2022.2072768\nWang, L., Song, R., Guo, W., & Yang, H. (2025). Exploring prompt pattern for generative artificial intelligence in automatic question generation. Interactive Learning \nEnvironments, 33(3), 2559–2584. https"
}