A conceptual framework for teacher-AI teaming in education: Harnessing generative AI to enhance teacher agency.

As generative artificial intelligence (GenAI) technologies rapidly permeate education, this session offers a conceptual analysis of how AI is reshaping teacher agency and argues that strengthening this agency is paramount in an AI-driven era to ensure educational practices remain human-centred, ethically grounded, and conducive to the ongoing development of teacher competence. It introduces a framework that distinguishes between: (i) replacement, (ii) complementarity, and (iii) augmentation of teacher competence. Building on the latter two, the chapter proposes a five-level teacher-AI teaming framework comprising transactional, situational, operational, praxical, and synergistic modes of interaction. It highlights the unique affordances of GenAI to help us achieve more praxical and synergistic teacher-AI interactions, as well as GenAI’s potential to enhance transactional, situational and operational teaming in a diverse set of tasks. The session concludes that while the replacement paradigm can indeed yield productivity gains, these benefits may come with costs that must be acknowledged and discussed to support informed decision-making.

The growing presence of generative artificial intelligence (GenAI) technologies in educational contexts presents a paradox of empowerment and concern. On one hand, AI promises to relieve teachers of burdensome administrative tasks, provide personalised learning insights , and complement instructional capabilities including for lesson planning, for classroom implementation and for assessment. For example, AI could help teachers plan their lessons by supporting their review of the knowledge to be taught, by providing ideas on alternative pedagogical strategies or by defining students’ needs and familiarising them with such needs. AI could also support classroom instruction through immediate feedback to students or to teachers about their practice or about their interventions; and AI can be used for modelling students’ mastery, to generate assessment items, supported essay scoring  or feedback on certain aspects of their teaching. On the other hand, there are significant concerns regarding potential threats, including the erosion of teacher autonomy, de-professionalisation of teaching, and ethical pitfalls if AI is misapplied. Many education experts are also concerned that over-reliance on AI to perform teachers’ tasks, such as marking, feedback generation, and lesson planning could risk skill atrophy for teachers. Therefore, the notion of teacher agency (i.e. teachers’ ability to exercise professional judgment, innovate in practice, and maintain control over pedagogical decisions) has come to the forefront of debates on AI in education. This chapter emphasises the need and urgency for moving beyond polarised narratives, rejecting both the dystopian fear of AI teachers replacing humans and the utopian hype that AI alone will solve all educational problems. Instead, it calls for evidence-informed strategies that harness AI’s potential while safeguarding teachers’ agency, rights, and professional integrity. As professionals increasingly offload tasks, that were traditionally viewed as uniquely human, to AI, we must engage in profound and forward-looking reflection about the essential roles and identities of professionals. What are the enduring and uniquely human qualities at the heart of education? How can we safeguard, elevate, and flourish them as technology transforms lives and the broader teaching and learning ecosystem? What is the role of a teacher in education where technological advancements allow the generation and sharing of information effectively and efficiently? To frame the inquiry into these difficult questions, the chapter outlines a continuum of AI integration in teaching, distinguishing between replacement, complementarity, and augmentation of teacher competence with AI in educational practice. Replacement through automation refers to tasks being offloaded to AI systems and carried out by them with no further teacher intervention. Complementarity requires a careful consideration of what teacher competence means for a given task and a specific AI tool’s design, development, and deployment features that can complement the specific aspects of the teacher’s competence. Augmentation suggests a deeper assimilation of AI models into teachers’ cognitive and pedagogical processes to be able to improve teachers’ competence through iterative interactions so that AI-augmented teachers can achieve the task better than AI or teacher alone (i.e. competence augmentation increases the likelihood of teacher-AI teams to perform better than the best of teacher or AI alone). Understanding these conceptual models is crucial for identifying when AI use supports teacher agency, when it might erode it, and when teachers’ agentic interactions can lead to competence augmentation for teachers, and thus to more effective teaching.

A review of the literature reveals that while AI applications in education are burgeoning, our understanding of how exactly teachers are integrating GenAI into their practices at a scale remains nascent and continually evolving. Additionally, the ways in which teachers engage with GenAI vary significantly across countries and jurisdictions, making it challenging to comprehensively capture the full range of practices in this chapter. However, for instance, based on the open call for evidence of the use of GenAI in education by the UK Department for Education, which received (non-representative, but elaborate) responses from 567 participants, the majority of whom were teachers, the public generally believed GenAI to offer various opportunities. These include freeing up teachers' time, improving teaching and educational materials, providing additional support to students (particularly those with special educational needs and disabilities (SEND) and those for whom English is an additional language (EAL)), and enhancing subject-specific applications (e.g. STEM). Overall, these perceived benefits argued to outweigh concerns about GenAI in education (e.g. students’ overreliance on GenAI, academic misconduct, fear of GenAI replacing face-to-face teaching, and the ”digital divide”). Most use cases observed in self-declared surveys and interviews of teachers also indicate that teachers use GenAI to develop lesson materials, ensuring alignment with curricular objectives while saving time on content preparation. In assessment, teachers tend to use it to support their marking and provide personalised formative feedback for students. Beyond the classroom, teachers appear to use GenAI in drafting statutory policies, streamlining administrative tasks, and aiming to reduce bureaucratic burdens.

While self-reported survey data offer some insights into teachers' usage of GenAI, this method faces inherent limitations, notably biases arising from external pressures or social desirability that may prevent teachers from accurately reporting their AI practices. Recent research from Anthropic.ai analyses over four million conversations with their GenAI system and shows high reliance on AI for some professions. Among certain professional groups, including language and literature teachers, AI conversations correspond to the performance of more than 75% of their professional tasks (following the task mapping for occupations by the O*Net database of the US Department of Labor). While understanding the extent of GenAI usage among teachers is crucial, examining precisely how teachers employ GenAI, whether primarily for augmentation or automation, is equally significant. Although it is a task-dependent discussion, evidence suggests a tendency towards automating routine tasks such as content generation, with AI directly executing responsibilities requiring minimal teacher involvement. Although such conclusions should be interpreted cautiously, as log data alone cannot reveal how teachers ultimately engage with, or act upon AI-generated outputs (e.g. some may discard them entirely and continue their work independently), automated applications of AI in education raise critical questions regarding the broader implications for teacher agency. Beyond the recent use cases of GenAI, the past few decades have witnessed an explosion of interest in applying AI to education, from intelligent tutoring system and automated grading tools to AI-driven decision-making tools and adaptive learning platforms. There is substantial amount of evidence that shows the positive impact of using these AI applications to support students’ academic performance, their affective engagement, and metacognitive development in controlled experimental evaluations. Although these small-scale empirical studies provide valuable insights, their outcomes typically reflect carefully designed academic AI tools evaluated under controlled conditions in which teacher implementation is guided by researchers in high fidelity. Alongside these promises, early scholarly and practitioner commentary raised flags about possible pitfalls. These concerns have been amplified by the rapid rise of GenAI, which brought AI’s capabilities, along with its risks, into mainstream awareness. Therefore, most educational stakeholders recently began grappling with scenarios that once seemed futuristic, or of interest to a small group of scholars. One core concern in these discussions is the potential erosion of teacher agency. As GenAI systems begin to handle not just typical monotonous administrative tasks of teachers but also complex pedagogical and instructional decisions, such as selecting content, assessing student work, or providing feedback, teachers might find their professional judgment marginalised by algorithmic outputs. Recent empirical research is starting to document these dynamics. For instance, a recent study on pre-service teachers indicated that exposure to GenAI in education prompted reflections on their evolving role and anxiety about role change, highlighting the need to prepare teachers for new hybrid roles working alongside AI. Complementary evidence from interview-based research with 57 teachers across eight schools in Sweden and Australia further reveals that, rather than freeing teachers from work, GenAI often generates new forms of invisible labour while challenging their agency on the pedagogical appropriateness and social sensitivity of educational content (Selwyn, Ljungqvist and Sonesson. These findings suggest that while GenAI tools have significant potential to support teachers and teaching practices, they can, in practice, redistribute and obscure teacher labour, reinforcing rather than reducing workload anxieties and concerns over teachers’ professional autonomy.

  

In this section, teacher agency refers to teachers’ active capacity and incentives to make choices and exert influence in their professional practice. It encompasses the autonomy to make instructional decisions, the ability to adapt and innovate pedagogy, the power to shape the educational environment in accordance with their professional values and their students’ needs as well as the willingness and incentives to do so. In essence, teachers have agency when they “act rather than are acted upon” in the educational process. Their agency is rooted in professional competence and confidence, and is often enabled or constrained by the tools and technology they use in their practice as well as the broader institutional and policy context in which they are situated. Teacher agency is not an all-or-nothing attribute; it exists in degrees and forms. Educational sociologists have described multiple forms of teacher autonomy, for example, autonomy over curriculum content, over pedagogy, over student assessment, and over professional development pathways (Frostenson, 2015[27]). A supportive school culture and policy framework can expand these autonomies, whereas top-down mandates, high-stakes accountability regimes, or technology systems making autonomous decisions on their behalf can compress them (as can be the case with AI-based educational technologies). Teacher agency is a key concept in education since teachers’ sense of agency is linked to their motivation, job satisfaction, and willingness to embrace new pedagogies or tools. When teachers feel empowered to make decisions, they are more likely to take initiative in improving their teaching and respond creatively to challenges in the classroom. Teacher agency enables teachers to adapt and update curricula, incorporate emerging real-world problems, and contextualise learning experiences to social, emotional, and relational needs of their students in ways that AI algorithms cannot. Moreover, teacher agency is tightly connected to teacher identity, that is, the sense of oneself as a professional with a meaningful mission. The introduction of GenAI into the classroom can perturb that identity, since some teachers may fear being displaced or judged by GenAI systems, while others might see GenAI as an opportunity to enhance their effectiveness.

From a learner’s perspective, teacher agency has the potential to translate to richer educational interactions. A teacher with high agency will actively interpret AI-generated insights or recommendations and adapt them to the context of their students. For example, if a GenAI tool generates feedback on students’ essays or produces a set of suggested prompts to foster metacognitive reflection, a teacher exercising professional agency will treat these as provisional resources (e.g. reviewing their pedagogical relevance, rephrasing or extending them to align with students’ learning goals), and using them as a springboard for discussion or further inquiry. By contrast, a low-agency scenario might involve a teacher simply pasting the AI-generated feedback into the learning platform without review or contextualisation, or alternatively, disregarding the system’s suggestions entirely due to mistrust or lack of confidence. Both extremes are suboptimal, and the goal is a balanced partnership where the teacher remains the orchestrator of instruction, using AI tools as informative assistants. Although most researchers and practitioners would agree with the proposed need for balance, there is little understanding regarding where this balance stands for a given teacher task, how it can be conceptualised, and how it can be operationalised. This chapter is an attempt to fill in this gap. In the following sections, the chapter dives deeper into how a theoretical conceptualisation of AI’s role vis-à-vis teacher agency, introducing the conceptualisations of replacement, complementarity, and augmentation, and then proposes a five-level teacher-AI teaming framework with clear definitions and examples of each level. Although, a broad set of socio-technical, institutional, and cultural factors profoundly shape how teachers perceive, enact, and sustain agency in their interactions with AI, the five-level framework proposed here mainly examines teacher agency through the lens of AI system affordances and interface-level design considerations.

 

This section presents three conceptual modes of integration as a framework to differentiate AI’s roles and their implications for teacher agency. These modes can be considered as points along a dynamic spectrum from AI operating independently of teachers to AI becoming deeply embedded in teachers’ cognitive routines to augment their competence. By delineating these, the chapter aims to clarify which approaches threaten teacher agency and which can potentially bolster it. Although the framework is applicable to all forms of AI, generative AI provides distinctive affordances that can shape and support varying levels of teacher agency. These will be discussed with examples where appropriate.

Replacement refers to AI systems executing tasks that a teacher would typically do with an AI-driven process with minimal or no teacher intervention. Classic examples include automated grading of exams or essays, asking an AI model to create lesson plans, questions, materials, algorithmic scheduling of student practice, such as homework, or AI tutors delivering content directly to learners, automating the pedagogical practice of teachers. The main appeal of automation is efficiency and scalability. Indeed, certain labour-intensive tasks in teaching (e.g. grading multiple-choice quizzes, drafting lesson plans, generating practice problems) can be reliably automated, potentially freeing teachers’ time for other or more complex work. Influenced by the political, managerial, or leadership level ecosystemic issues of teaching practice, currently the evidence about how teachers actually use this saved time is scarce. The potential of productivity gains in education is indeed important. Recent research by the Education Endowment Foundation (EEF), independently evaluated by the National Foundation for Educational Research (NFER), examined the use of GenAI among 259 teachers across 68 secondary schools in England. The randomised controlled trial revealed that teachers who used GenAI, supplemented by practical guidance, reduced their lesson and resource planning time by an average of 31 percent, reducing their weekly average planning time from 81.5 to 56.2 minutes, without compromising the quality of their lesson plans and resources. As highlighted by the Teacher Task Force and UNESCO Global Report on Teachers (2024), education systems face compounding teacher and resource crises, especially in low- and middle-income countries. An estimated 44 million additional primary and secondary teachers are said to be needed by 2030, including 15 million in sub-Saharan Africa. This worldwide shortage of teachers is aggravated by rising attrition, as many teachers leave the profession early. Rural and remote areas have been hit hardest, where underqualified teachers often fill the gaps and multi-grade classrooms are common; 90% of secondary schools in sub- Saharan Africa face serious teaching shortages. Consequently, learning gaps are widening. Students also contend with severe shortages of education materials and quality content. In some classrooms, a single textbook must be shared by a dozen or more pupils. Much of the digital education content that could help is not in the learner’s language. For instance, even though they are not representative of countries’ learning resources, 92% of open education resources are in English, marginalising non-English-speaking learners. The replacement paradigm in AI offers opportunities to counteract these global shortages with significant productivity gains. GenAI-powered tools can supplement overburdened teaching workforces and provide instructional support to students in underserved areas. Translation and content generation driven by GenAI can expand the availability of high-quality teaching resources in local languages and for students with Special Educational Needs and Disabilities (SEND). Nevertheless, the full automation of teacher tasks also raises concerns about the loss of teacher agency and may come at certain costs. The purpose of highlighting these challenges is not to oppose productivity gains in education, but to invite policy dialogue and careful consideration of how some of these challenges can be mitigated. Two of those concerns are the dehumanisation of education and teacher cognitive atrophy.

Dehumanisation refers to the erosion of the human elements that are fundamental to teaching and learning. Education is an interactive process of human development. When AI is used as a substitute for teachers or peers, there is a danger that learning becomes overly mechanistic, losing the empathy and social dynamics that characterise many effective pedagogies and educational practices. Dehumanisation can manifest in teaching, assessment and feedback for example. Some may be tempted to have “teacherless schools” and provide instruction by having students working through AI-personalised curricula. However, human teachers contribute numerous intangible qualities (e.g. moral judgment, inspiration, role-modelling, the ability to build trust with other humans and mentor them) that no AI currently can replicate. Even if students progressed academically, they would miss out on collaborative learning, dialogue, and the social construction of knowledge with other humans. Educational practice often thrives on inquirybased instruction, collaborative lab-work, group discussions, and debate, which are facets that require human presence and guidance. Similarly, feedback and evaluation from impersonal algorithms might make students feel less seen or valued as individuals. Finally, if teachers were asked to work under algorithmic scripts or performance dashboards that dictated their every move, their professional identity would likely erode and resemble more that of assembly line workers than educators. Protecting teacher agency is thus directly tied to keeping education humancentred. Similarly, the integration of generative AI into educational environments in ways that do not allow teacher agency also raises significant questions about its impact on teachers various higher-order thinking skills, particularly through its influence on critical and reflective thinking practices. A recent experiment focusing on the cognitive cost of using a Large Language Model in the educational context of writing an essay shows that students writing without an LLM support exhibited the strongest, widest‑ranging brain activation, those using a search engine showing intermediate engagement, and those using an LLM a limited cognitive engagement. While the evidence on the potential negative impact on GenAI users’ cognitive capabilities when they use GenAI is just emerging, this may also be a risk for teachers when they use AI in the replacement paradigm.

Complementarity refers to AI systems functioning as supportive tools that amplify a teacher’s capabilities, while the teacher remains actively involved. In the complementarity paradigm, AI and teacher work in tandem, each contributing what they do best without necessarily interacting with each other for augmentation of each other’s competence. The underlying philosophy of complementarity is that leveraging AI’s strengths (e.g. data processing, pattern recognition, speed, scale, no exhaustion, efficiency and time-saving opportunities etc.) to complement human strengths (e.g. relational interpretation, empathy, moral discernment, contextual judgment etc.). In this paradigm, as AI primarily processes data to present insights or operationalise instructional intentions defined by teachers. It is less about AI learning from humans and more about teachers internalising computational representations and reshaping their own mental models and professional reasoning through complementary interactions with AI. Contrary to the replacement paradigm, AI systems don’t perform teachers’ tasks entirely

If implemented appropriately, complementarity conceptualisation can further reinforce the teacher’s agency and has the potential to improve human competence at a given task. For complementarity to be operationalised, we must first articulate a human competence model to be able to specifically define what aspects of a teacher competence1 can be complemented with AI. Then, we can examine how the affordances of a given AI modelling technique or the design of a specific AI agent can interlock with each layer of that competence continuum. Holstein et al. provide a useful framework for thinking about what specific aspects of teacher competence can be complemented with the help of AI, identifying four dimensions of complementarity: complementary goal setting (e.g. when teachers set, monitor, and evaluate learning goals with AI support); complementary perceptual input (e.g. when teachers’ perception about student learning is expanded by AI); complementary actions (e.g. when teachers’ actions are scaled by AI); complementary decisions (e.g. when teacher decision-making is assisted by data-informed AI recommendations). For instance, an AI model that has the capabilities of processing sensory information from students’ interactions can complement teachers’ situation specific skills of monitoring student interactions. On the other hand, an AI model that has the capability of tracking students’ online interactions in an intelligent tutoring system can complement teachers’ knowledge of their students’ current level of mastery on a topic. Such a teacher-AI complementarity would have the purpose of supporting rather than supplanting teachers’ integrative professional judgment. However, this framework does not provide any insights into how exactly specific aspects of teacher competence can be complemented. Building on this notion of complementarity, it becomes essential to conceptualise how teacher-AI interactions may vary in depth and complexity. Here we propose that for any given educational task, and depending on both the specific aspects of a teacher’s competence, and the specific AI affordances, teacher-AI teaming can occur at five distinct levels: transactional, situational, operational, praxical, and synergistic teaming.

Transactional teaming refers to interactions between teachers and AI that consist of discrete transactions, defined by a request-response mechanism (i.e. the teacher inputs a command, and the AI outputs a result). Teachers and AI systems’ actions are perceived by one another, with each action dynamically informing and triggering a corresponding response from the other agent. At this level, AI agents can automatically perform actions based on teachers’ input, often completing routine or repetitive tasks on their behalf. The core dynamic of transactional teaming is “request to execution” for task automation and efficiency; therefore, the primary goal is to enhance teaching productivity by streamlining these processes. The Srivastava et al.’s Smart Learning Assistance (SLA) system can be described as an exemplar of transactional teaming in teacher-AI interaction. In this system, the teacher issues a command (e.g. inputting spoken words, sign language gestures, or Braille text), and the GenAI tool automatically converts to or from speech, sign language, or Braille, thereby returning a translated result that supports communication with the teacher or peers. In doing so, the SLA tool takes over routine translation tasks, allowing the teacher to delegate discrete, repetitive conversion work, precisely the “request → execution”, dynamic that defines transactional teaming.

Situational teaming refers to a form of interaction in which teachers operate based on a shared awareness of the teaching and learning context, constructed through the combined perceptions of both human and artificial agents. At this level, AI systems collect data from classroom interactions and/or learning activities in digital learning platforms or real-world classrooms through sensors, process it using underlying models, and provide educationally meaningful information to support teachers in making informed decisions and taking appropriate actions. The Hybrid HumanAgent Tutoring (HAT) platform developed by Sawaya et al. can be used as an example of situational teaming. The system collects data on tutors’ discourse practices, analyses it with AI models, and provides human teachers and coaches with GenAI created feedback to guide their instructional coaching sessions. This creates a shared awareness of the tutoring context, where AI highlights patterns in tutors’ interactions and coaches use these insights to make informed pedagogical decisions. The core dynamic mirrors situational teaming in that teachers and AI agents coconstruct context awareness that informs human action, rather than automating tasks.

Operational teaming involves the cooperation of planning and execution of teaching-related tasks between the teacher and the AI system. At this level, teachers provide information about the current and desired states of the teaching and learning context, articulated through intentions, instructional goals, tasks, and actionable plans. The AI system supports teachers by incorporating these goals in its decision-making to autonomously perform or assist with the goals set by teachers. This teaming enables efficient task execution aligned with the teacher’s needs. However, it also requires teachers to have a comprehensive understanding of the instructional goals and the pedagogical interventions to achieve them. The Pair-Up system developed by Yang et al. can be used as an example of operational teaming. In this system, teachers articulate high-level instructional goals (which can also be done with natural language using GenAI), such as when to transition students between individual and collaborative learning activities. The AI system integrates these goals into its decision-making by monitoring student progress in an intelligent tutoring system in real time, then recommending or enacting transitions for classroom practice that align with the teacher’s pedagogical intentions. This cooperative planning and execution allow teachers and AI to jointly manage complex classroom orchestration tasks, with AI autonomously assisting in carrying out the instructional plans while ensuring that task execution remains consistent with the teacher’s overarching objectives.

Praxical teaming refers to a form of interaction in which the teacher and the AI system exchange information about actions and procedures, grounded in prior experience, usage patterns, or training. This level of teaming emphasises the development of shared “understanding” and practices over time, enabling the AI to learn from the teacher’s instructional habits and preferences, while the teacher adapts to the AI’s capabilities and pedagogical suggestions. For example, when offering recommendations to improve teaching, an LLM-based conversational support system can learn from teachers’ feedback (both explicit and implicit) on those suggestions and adjust their underlying models accordingly. This requires teachers to possess the competence to critically evaluate the AI’s suggestions, rather than accepting them uncritically and AI to have the capability to learn from teacher corrections to adjust its internal model accordingly. The machine-learning-based feedback suggestion system developed by Bernius, Krusche, and Bruegge (2021[36]) illustrates an example of praxical teaming. The system analyses students’ python coding script submissions in large courses and proposes feedback suggestions that instructors can review, adapt, or reject. Over time, the AI learns from teachers’ corrections and adjustments, refining its ability to generate more contextually appropriate feedback aligned with instructors’ pedagogical preferences. This dynamic exchange of information, AI adapting to teachers’ evaluative patterns while teachers critically assess and refine AI-generated suggestions, embodies the essence of praxical teaming, where shared practices and mutual adaptivity develop iteratively. 

Augmentation refers to the process by which AI tools, and the new practices they enable, become woven into the internal repertoire of teachers in ways that these human-AI interactions also lead to an increase in teacher competence. This very much corresponds to the fifth level of the teacher-AI teaming framework.

Synergistic teaming refers to a form of interaction in which the teacher and the AI system mutually enhance each other through critical evaluations, challenging to each others’ suggestions and propositions with logic and evidence, and engage in solving complex problems together to move towards a shared understanding and mutual development. Effective synergistic teaming involves mutual interaction, where the AI agents and teachers evaluate each other’s claims with epistemic awareness and remind one another of aspects that may have been overlooked. When this interaction is well-aligned, a form of creative resonance emerges, enabling the teacher and AI to deepen their understanding of the task and generate outcomes neither could achieve independently. Therefore, this type of teaming is conceptualised as synergistic, that is, the emergent competence is likely to exceed the maximum of individual AI or human competence at a given task. The main limitations of praxical teaming in comparison to synergistic teaming lie in its reliance on the existing knowledge, expertise, and competence of the teacher, without necessarily pushing them beyond their current practices. As such, praxical teaming tends to converge at the maximum of the teacher’s present competence, whereas synergistic teaming requires a more ambitious, mutual development between AI and teachers, a state that remains far more difficult to achieve in practice. Thus, while praxical teaming often stabilises at the ceiling of teachers’ or AI’s current max performance with the potential benefits of efficiency and time-savings, moving toward synergistic teaming, where AI and teachers mutually extend and transform one another’s capacities, is more difficult to realise.

The main condition for augmentation in this sense is synergistic human-AI teaming, where human-AI combination yields emergent competence exceeding the maximum of individual AI or human competence alone at a given task.

However, this idea, informed by theories of distributed and extended cognition, faces significant empirical hurdles. A recent comprehensive meta-analysis of 106 experimental studies covering all sectors, with ~370 effect sizes reported, reveals that human-AI combinations underperform the better performer, human or AI alone, in 58% of the cases reviewed, demonstrating that synergy is context-dependent and should not be assumed. Critically, synergy emerges only under specific conditions. For instance, task type matters. Synergy is more likely to occur in relatively simple content creation tasks (e.g. open-ended science education questions) but fails in complex decision-making tasks (e.g. classification of a student’s emotional state given their relationship with other students). Second, when humans outperform AI alone, synergy is more likely to arise; yet when AI outperforms humans, human-AI teaming tends to degrade the performance. This indicates that synergy depends on human competencies to meta-cognitively assess when to trust AI input and when to ignore it which emphasises the importance of improving teachers and students’ AI competencies to be able to protect their own agency and have a critical and informed trust in AI (see, for instance, UNESCO AI competency framework for teachers. Third, the instructional design in which AI is interacted with is very important. The evidence from the meta-review indicates that only 3 out of 106 experiments explicitly tested predetermined subtask delegation between humans and AI to structure their interactions. These yielded non-significant synergy gains, underscoring that effective augmentation requires systematically scaffolded human-AI interactions embedded within a well-designed instructional framework. These findings affirm that synergistic augmentation is difficult to achieve across many domains including healthcare, law, and education. They suggest that future AI implementations in education aiming to realise augmentation through synergistic interactions should prioritise task-specific AI design, development, and deployment, rather than relying on generic tools such as ChatGPT, which are not inherently educational, or human development, technologies. While such general-purpose systems can be valuable for exploratory or creative use, their outputs are not pedagogically grounded nor optimised for instructional alignment. Achieving meaningful synergy therefore requires the deliberate scaffolding of teacher-AI interactions within context-specific educational frameworks and the provision of AI competency training for teachers to help them unlock emergent forms of teacher-AI teaming. It is equally important to establish robust evaluation metrics that capture what specific aspects of teacher and AI competence are being augmented at a given task. Without these intentional design and assessment mechanisms, teacher-AI teaming is unlikely to achieve synergy and risk producing diminishing returns compared to standalone human or AI performance. The five-level framework proposed here examines the teacher agency through the lens of AI system affordances and interface-level design considerations. The system affordances discussed above therefore are necessary, yet not sufficient conditions for teacher-AI synergistic interactions. Box 7.1 presents examples for the five different AI-teacher teaming presented above.

Empirical evidence from recent studies underscores the tangible benefits of praxical and synergistic teacherAI teaming in enhancing both teacher efficacy and student learning outcomes. For instance, in an auto grader research study, a praxical teacher-AI teaming approach was shown to reduce grading time by 44% while improving accuracy by 6% compared to manual grading. Teaching assistants consistently rated the AI-assisted process as faster, easier, and more enjoyable, reporting that automation alleviated routine cognitive burdens and allowed greater focus on higher-order pedagogical reasoning. Similarly, a teacher-AI feedback co-creation study demonstrated that involving subject matter experts in GenAI-supported content authoring can yield comparable instructional quality to human-only materials, while dramatically reducing time and cognitive effort. Involving ten mathematics experts and 358 learners, Reza et al. show that an iterative human-AI approach to co-producing feedback reduced perceived workload by 50% and shortened the content development process from several months to a few hours, while maintaining statistically significant learning gains for student. Together, these findings confirm the empirical potential of GenAI to support more advanced teaming approaches to amplify teacher productivity, improve instructional quality, and sustain learning outcomes when human oversight and agency remain central to system design. However, both studies caution that augmentation is not without limitations. In the Liu et al. study, AI performance degraded on unrepresented cases, revealing the dependence of system reliability on training data diversity and the need for ongoing human teacher verification. Similarly, participants in the Reza et al. study reported occasional model unpredictability and difficulties in steering GenAI outputs, emphasising that the quality of augmentation depends on teachers’ prompt literacy and capacity for meta-cognitive regulation. Moreover, lessonspecific variation in learning outcomes indicated that GenAI may underperform in certain pedagogical contexts or with particular learner profiles. Finally, neither of these studies measured improvements in teacher competence before and after their interactions with the tools to evaluate the impact of these interactions on teacher competence. These findings collectively suggest that augmentation benefits are contingent upon interface design features, structured scaffolding of praxical and synergistic interactions, and the particular competence and motivation of teachers. Whatever the affordances of teacher-AI teaming might be, without sufficient teacher competence and motivation to engage meaningfully with these tools, the likelihood of achieving augmentation remains low. Furthermore, some teachers can indeed achieve competence gains even in their transactional teaming with AI tools, when these are used reflectively and purposefully. This also illustrates that the five levels of teacher-AI teaming are not hierarchically ordered stages of progression but rather context-sensitive modes of interaction. Different educational tasks, disciplinary demands, and institutional and local contexts may indeed benefit from different levels of teaming. Nonetheless, the design of teaming affordances that allow higher degrees of teacher agency (e.g. operational, praxical, and synergistic) represents a deliberate effort to maximise the likelihood that teacher-AI interaction contributes to further enhance teachers’ agency and competence development. These higher-order forms of teaming are not inherently superior, but they are structured to provide richer opportunities for reflection, adaptation, and pedagogical transformation. In this sense, augmentation is not a property of technology alone but an emergent outcome of the dynamic interplay between teacher competence and motivation; the design of human-AI complementarity interface; and the affordances of the AI models.

GenAI marks a significant advancement regarding the AI affordances in the evolution of teacher-AI teaming levels by expanding the scope of complementarity across all levels of interaction. First of all, GenAI enables diversity and efficiency at a scale and precision previously unattainable. GenAI’s capacity to generate multimodal outputs (e.g. textual, visual, auditory) allows teachers to access and adapt resources to diverse learner profiles, linguistic backgrounds, and learning needs. High performance of the state-of-the-art GenAI models in content generation creates unprecedented opportunities for teachers’ transactional teaming with GenAI, allowing them to complete a wider range of tasks, from contextually appropriate lesson materials to content for students with certain impairments. Even generic GenAI systems (e.g. ChatGPT, Gemini, DeepSeek etc.), which are not designed specifically for education, can frequently support transactional and operational teaming due to their broad linguistic and creative affordances. For instance, teachers today commonly request ChatGPT to generate a quiz, summarise a text, or rephrase feedback comments. Although the interaction remains largely transactional, a request is made, and an output is returned: they provide significant support on a wide range of content generation tasks which were simply not possible before GenAI. The teacher’s agency in these interactions often lies in evaluating and adapting the GenAI response, but the cognitive exchange largely remains at the level of task automation and efficiency. In situational and operational teaming, GenAI provides richer situational awareness by synthesising diverse streams of classroom data (e.g. text, speech, physiological signals, interaction logs, and visual cues) into interpretable insights that help teachers make informed, timely decisions. Unlike earlier traditional AI and analytics systems that offered static or unidimensional dashboards, GenAI can translate complex multimodal data signals into narrative explanations or alternative scenario projections, allowing teachers to be better informed about their classroom contexts through multiple perspectives. Furthermore, GenAI’s language-based reasoning affordances enable it to articulate situational data processing and operational input to the model easier for co-creating plans, goals, and intentions in naturalistic dialogue rather than coding, or limited teacher authoring tools, which would be the case for traditional AI approaches.

 More profoundly, the affordances of GenAI open unprecedented pathways towards praxical and synergistic teaming. Unlike traditional AI systems, which operate within fixed boundaries of prediction or classification, GenAI has the potential to engage in co-creative processes with teachers such as generating pedagogical hypotheses, critiquing lesson structures, suggesting conceptual analogies, and surfacing potential epistemic tensions in instructional design. Such a dialogic engagement, which is at the core of praxical and synergistic teaming levels, can be achieved with natural language using GenAI. GenAI also has capacity for adaptive, open-ended reasoning which can allow teachers to externalise, examine, and refine their pedagogical thinking in iterative cycles of reflection and critique. This recursive exchange has the potential to nurture professional growth, enabling teachers to question habitual practices and to explore alternative approaches that neither human expertise nor algorithmic optimisation alone could have revealed. Yet, we are only at the early stages of integrating GenAI meaningfully into teacher-facing AI tools, and the extent to which these systems can genuinely share the responsibility of meaning-making with teachers remains uncertain. It is not yet clear whether GenAI tools can co-create pedagogical meaning in a way that authentically challenges teachers’ reasoning, provokes reflection, and contributes to deeper professional learning. While their dialogic affordances hold promise for more reciprocal exchanges, current implementations rarely demonstrate the capacity to push back against inappropriate or suboptimal pedagogical decisions, to question teachers’ assumptions, or to propose alternative perspectives grounded in educational theory and evidence. Achieving such reflective tension, where AI systems not only assist but also constructively challenge teachers, still require advances in both the cognitive modelling of teaching expertise and the design of teacher-AI interaction interfaces, but GenAI systems provide unique opportunities to be explored in the upcoming years.

The future of GenAI in education will be determined not by how efficiently it helps automate tasks, but by how effectively it empowers teachers to exercise their professional judgment and expand their competence. The conceptual model proposed here (i.e. replacement, complementarity, augmentation) and the five levels of the teacher-AI teaming framework (transactional, situational, operational, praxical, and synergistic) provide a basis for policy and research to map existing teacher-facing AI systems and develop design principles to identify when GenAI systems replace, complement or augment teachers’ competences. The five levels of teacher-AI teaming framework proposed here can be used to evaluate AI tools against explicit teacher-agency support criteria, and track dynamic changes from transactional to synergistic teaming interactions. It is also important to acknowledge that since synergy requires mutual interactions of two-way information flow between AI and humans, one could also speak of the internalisation of human values into AI systems as part of this process. As teachers work closely with AI, ideally, they would influence the design and tuning of these systems (e.g. through feedback, pushback on recommendations and reasons provided for such pushback, usage choices, participatory design, etc.). Over time, AI used in classrooms should learn the priorities of good teaching for a given teacher dependent upon their pedagogical philosophy, needs and requirements (e.g. promoting inquiry, not giving away answers too easily, respecting diverse solutions, prioritising certain instructional choices over others etc.) because teachers enforce these in their interactions with the AI. In a sense, the AI system learns some of the teacher’s pedagogical wisdom through these synergistic interactions. Thus, augmentation would also require appropriate spaces for the reflection of humans and space for model updates for AI. That is, teachers reflecting on how AI is affecting their practice and their students’ learning, AI using logged interaction data to update its model parameters and weighs to learn from its interactions with teachers. This two-way internalisation is at the heart of human-AI co-evolution in education for synergistic augmentation; teachers shape AI just as AI shapes teaching practices synergistically. The synergistic augmentation paradigm represents the deepest form of teacher-AI interaction, where AI is no longer a distinct and complementary “add-on” to their competence, but part of the fabric of teaching and learning. This holds the promise of truly hybrid intelligence pedagogies that leverage AI’s capabilities while being guided by human wisdom in synergistic interactions that lead to an augmented competence that is greater than the maximum of an individual human or AI alone. It demands high levels of teacher agency to negotiate the human-AI relationship. In addition, teachers must remain self-aware, continually asserting human values and expertise in the loop even as they embrace new AI-empowered methods. Achieving success at this augmentation paradigm is challenging in education, but it aligns with the vision of education where AI serves as a competence augmentation tool for teachers, ultimately enriching the teaching and learning experience.

No generative AI tools were used in the creation or development of the ideas presented in this chapter. This chapter’s content is based on the keynote talk given by the author at the ACM International Conference on Learning Analytics & Knowledge (LAK) in March 2024 in Kyoto, Japan, and the UNESCO International Task Force on Teachers for Education2030 position paper on the protection and promotion of teacher agency in the age of AI he authored. The author takes full responsibility for the content and its intellectual integrity.