Education stands as a cornerstone of society, nurturing the minds that will ultimately shape our future¹. As we advance into the twenty-first century, exponentially developing technologies and the convergence of knowledge across disciplines are set to have a significant influence on various aspects of life², with education a crucial element that is both disrupted by and critical to progress³. The rise of artificial intelligence (AI), notably generative AI and generative pre-trained transformers⁴ such as ChatGPT, with its new capabilities to generalise, summarise, and provide human-like dialogue across almost every discipline, is set to disrupt the education sector from K-12 through to lifelong learning by challenging traditional systems and pedagogical approaches^5,6.

Artificial intelligence can be defined as the simulation of human intelligence and its processes by machines, especially computer systems, which encompasses learning (the acquisition of information and rules for using information), reasoning (using rules to reach approximate or definite conclusions), and flexible adaptation^7,8,9. In education, AI, or AIED, aims to “make computationally precise and explicit forms of educational, psychological and social knowledge which are often left implicit“¹⁰. Therefore, the promise of AI to revolutionise education is predicated on its ability to provide adaptive and personalised learning experiences, thereby recognising and nurturing the unique cognitive capabilities of each student¹¹. Furthermore, integrating AI into pedagogical approaches and practice presents unparalleled opportunities for efficiency, global reach, and the potential for the democratisation of education unattainable by traditional approaches.

AIED encompasses a broad spectrum of applications, from adaptive learning platforms that curate customised content to fit individual learning styles and paces¹² to AI-driven analytics tools that forecast student performance and provide educators with actionable insights¹³. Moreover, recent developments in AIED have expanded the educational toolkit to include chatbots for student support, natural language processing for language learning, and machine learning for automating administrative tasks, allowing educators to focus more or exclusively on teaching and mentoring¹⁴. These tools have recently converged into multipurpose, generative pre-trained transformers (GPTs). These GPTs are large language models (LLMs) utilizing transformers to combine large language data sets and immense computing power to create an intelligent model that, after training, can generate complex, advanced, human-level output¹⁵ in the form of text, images, voice, and video. These models are capable of multi-round human-computer dialogues, continuously responding with novel output each time users input a new prompt due to having been trained with data from the available corpus of human knowledge, ranging from the physical and natural sciences through medicine to psychiatry.

This convergence highlights that a step change has occurred in the capabilities of AI to act not only as a facilitator of educational content but also as a dynamic tool with agentic properties capable of interacting with stakeholders at all levels of the educational ecosystem, enhancing and potentially disrupting the traditional pedagogical process. Recently, the majority of the conversation within the current literature concerning AIED is focused on the aspect of cheating or plagiarism^16,17,18, with some calls to examine the ethics of AI¹⁹. This focus falls short of addressing the multidimensional, multi-stakeholder nature of AI-related issues in education. It fails to consider that AI is already here, accessible, and proliferating. It is this accessibility and proliferation that motivates the research presented in this manuscript. The release of generative AI globally and its application within education raises significant ethical concerns regarding data privacy, AI agency, transparency, explainability, and additional psychosocial factors, such as confidence and trust, as well as the acceptance and equitable deployment of the technology in the classroom²⁰.

As education touches upon all members and aspects of society, we, therefore, seek to investigate and understand the level of acceptability of AI within education for all stakeholders: students, teachers, parents, school staff, and principals. Using factors derived from the explainable AI literature²¹ and the UNESCO framework for AI in education²². We present research that investigates the role of agency, privacy, explainability, and transparency in shaping the perceptions of global utility (GU), individual usefulness (IU), confidence, justice, and risk toward AI and the eventual acceptance of AI and intention to use (ITU) in the classroom. These factors were chosen as the focus for this study based on feedback from focus groups that identified our four independent variables as the most prominent factors influencing AI acceptability, aligning with prior IS studies²¹ that have demonstrated their central role in AI adoption decisions. Additionally, these four variables directly influence other AI-related variables, such as fairness -conceptualized in our study as confidence- suggesting a mediating role in shaping intentions to use AI.

In an educational setting, the deployment of AI has the potential to redistribute agency over decision-making between human actors (teachers and students) and algorithmic systems or autonomous agents. As AI systems come to assume roles traditionally reserved for educators, the negotiation of autonomy between educator, student, and this new third party becomes a complex balancing act in many situations, such as personalising learning pathways, curating content, and even evaluating student performance^23,24.

Educational professionals face a paradigm shift where the agency afforded to AI systems must be weighed against preserving the educators’ pedagogical authority and expertise²⁵. However, this is predicated on human educators providing additional needs such as guidance, motivation, facilitation, and emotional investment, which may not hold as AI technology develops²⁶ That is not to say that AI will supplant the educator in the short term; rather, it highlights the need to calibrate AI’s role within the pedagogical process carefully.

Student agency, defined as the individual’s ability to act independently and make free choices²⁷, can be compromised or enhanced by AI. While AI can personalise learning experiences, adaptively responding to student needs, thus promoting agency²⁸, it can conversely reduce student agency through over-reliance, whereby AI-generated information may diminish students’ critical thinking and undermine the motivation toward self-regulated learning, leading to a dependency²⁹.

Moreover, in educational settings, the degree of agency afforded to AI systems, i.e., its autonomy and decision-making capability, raises significant ethical considerations at all stakeholder levels. A high degree of AI agency risks producing “automation complacency“³⁰, where stakeholders within the education ecosystem, from parents to teachers, uncritically accept AI guidance due to overestimating its capabilities. Whereas a low degree of agency essentially hamstrings the capabilities of AI and the reason for its application in education. Therefore, ensuring that AI systems are designed and implemented to support and enhance human agency through human-centred alignment and design rather than replacing it requires thorough attention to the design and deployment of these technologies³¹.

In conclusion, educational institutions must navigate the complex dynamics of assigned agency when integrating AI into pedagogical frameworks. This will require careful consideration of the balance between AI autonomy and human control to prevent the erosion of stakeholders’ agency at all levels of the education ecosystem and, thus, increase confidence and trust in AI as a tool for education.

Establishing confidence in AI systems is multifaceted, encompassing the ethical aspects of the system, the reliability of AI performance, the validity of its assessments, and the robustness of data-driven decision-making processes^32,33. Thus, confidence in AI systems within educational contexts centres on their capacity to operate reliably and contribute meaningfully to educational outcomes.

Building confidence in AI systems is directly linked to the consistency of their performance across diverse pedagogical scenarios³⁴. Consistency and reliability are judged by the AI system’s ability to function without frequent errors and sustain its performance over time³⁵. Thus, inconsistencies in AI performance, such as system downtime or erratic behaviour, may alter perceptions of utility and significantly decrease user confidence.

AI systems are increasingly employed to grade assignments and provide feedback, which are activities historically under the supervision of educators. Confidence in these systems hinges on their ability to deliver feedback that is precise, accurate, and contextually appropriate³⁶. The danger of misjudgment by AI, particularly in subjective assessment areas, can compromise its credibility³⁷, increasing risk perceptions for stakeholders such as parents and teachers and directly affecting learners’ perceptions of how fair and just AI systems are.

AI systems and the foundation models they are built upon are trained over immense curated datasets to drive their capabilities³⁸. The provenance of these data, the views of those who curate the subsequent training data, and how that data is then used within the model (that creates the AI) is of critical importance to ensure bias does not emerge when the model is applied^19,39. To build trust in AI, stakeholders at all levels must have confidence in the integrity of the data used to create an AI, the correctness of analyses performed, and any decisions proposed or taken⁴⁰. Moreover, the confidence-trust relationship in AI-driven decisions requires transparency about data sources, collection methods, and explainable analytical algorithms⁴¹.

Therefore, to increase and maintain stakeholder confidence and build trust in AIED, these systems must exhibit reliability, assessment accuracy, and transparent and explainable decision-making. Ensuring these attributes requires robust design, testing, and ongoing monitoring of AI systems, the models they are built upon, and the data used to train them.

Trust in AI is essential to its acceptance and utilisation at all stakeholder levels within education. Confidence and trust are inextricably linked⁴², representing a feedback loop wherein confidence builds towards trust and trust instils confidence, and the reverse holds that a lack of confidence fails to build trust. Thus, a loss of trust decreases confidence. Trust in AI is engendered by many factors, including but not limited to the transparency of AI processes, the alignment of AI functions with educational ethics, including risk and justice, the explainability of AI decision-making, privacy and the protection of student data, and evidence of AI’s effectiveness in improving learning outcomes^33,43,44.

Standing as a proxy for AI, studies of trust toward automation^45,46 have identified three main factors that influence trust: performance (how automation performs), process (how it accomplishes its objective), and purpose (why the automation was built originally). Accordingly, educators and students are more likely to trust AI if they can comprehend its decision-making processes and the rationale behind its recommendations or assessments⁴⁷. Thus, if AI operates opaquely as a “black box”, it can be difficult to accept its recommendations, leading to concerns about its ethical alignment. Therefore, the dynamics of stakeholder trust in AI hinges on the assurance that the technology operates transparently and without bias, respects student diversity, and functions fairly and justly⁴⁸.

Furthermore, privacy and security directly feed into the trust dynamic in that educational establishments are responsible for the data that AI stores and utilises to form its judgments. Tools for AIED are designed, in large part, to operate at scale, and a key component of scale is cloud computing, which involves resource sharing, which refers to the technology and the data stored on it⁴⁹. This resource sharing makes the boundary between personal and common data porous, which is viewed as a resource that many technology companies can use to train new AI models or as a product⁵⁰. Thus, while data breaches may erode trust in AIED in an immediate sense, far worse is the hidden assumption that all data is common. However, this issue can be addressed by stakeholders at various levels through ethical alignment negotiations, robust data privacy measures, security protocols, and policy support to enforce them^22,51.

Accountability is another important element of the AI trust dynamic, and one inextricably linked to agency and the problem of control. It refers to the mechanisms in place to hold system developers, the institutions that deploy AI, and those that use AI responsible for the functioning and outcomes of AI systems³³. The issue of who is responsible for AI’s decisions or mistakes is an open question heavily dependent on deep ethical analysis. However, it is of critical and immediate importance, particularly in education, where the stakes include the quality of teaching and learning, the fairness of assessments, and the well-being of students.

In conclusion, trust in AI is an umbrella construct that relies on many factors interwoven with ethical concerns. The interdependent relationship between confidence and trust suggests that the growth of one promotes the enhancement of the other. At the same time, their decline, through errors in performance, process, or purpose, leads to mutual erosion. The interplay between confidence and trust points towards explainability and transparency as potential moderating factors in the trust equation.

The contribution of explainability and transparency towards trust in AI systems is significant, particularly within the education sector; they enable stakeholders to understand and rationalise the mechanisms that drive AI decisions⁵². Comprehensibility is essential for educators and students not only to follow but also to assess and accept the judgments made by AI systems critically^53,54. Transparency gives users visibility of AI processes, which opens AI actions to scrutiny and validation⁵⁵.

Calibrating the right balance between explainability and transparency in AI systems is crucial in education, where the rationale behind decisions, such as student assessments and learning path recommendations, must be clear to ensure fairness and accountability^32,56. The technology is perceived to be more trustworthy when AI systems articulate, in an accessible manner, their reasoning for decisions and the underlying data from which they are made⁵⁷. Furthermore, transparency allows educators to align AI-driven interventions with pedagogical objectives, fostering an environment where AI acts as a supportive tool rather than an inscrutable authority^58,59,60.

Moreover, the explainability and transparency of AI algorithms are not simply a technical requirement but also a legal and ethical one, depending on interpretation, particularly in light of regulations such as the General Data Protection Regulation (GDPR), which posits a “right to explanation” for decisions made by automated systems^61,62,63. Thus, educational institutions are obligated to deploy AI systems that perform tasks effectively and provide transparent insights into their decision-making processes in a transparent manner^64,65.

In sum, explainability and transparency are critical co-factors in the trust dynamic, where trust appears to be the most significant factor toward the acceptance and effective use of AI in education. Systems that employ these methods enable stakeholders to understand, interrogate, and trust AI technologies, ensuring their responsible and ethical use in educational contexts.

When taken together, this discussion points to the acceptance of AI in education as a multifaceted construct, hinging on a harmonious yet precarious balance of agency, confidence, and trust underpinned by the twin pillars of explainability and transparency. Agency involving the balance of autonomy between AI, educators, and students requires careful calibration between AI autonomy and educator control to preserve pedagogical integrity and student agency, which is vital for independent decision-making and critical thinking. Accountability, closely tied to agency, strengthens trust by ensuring that AI systems are answerable for their decisions and outcomes, reducing risk perceptions. Trust in AI and its co-factor confidence are fundamental prerequisites for AI acceptance in educational environments. The foundation of this trust is built upon factors such as AI’s performance, the clarity of its processes, its alignment with educational ethics, and the security and privacy of data. Explainability and transparency are critical in strengthening the trust dynamic. They provide stakeholders with insights into AI decision-making processes, enabling understanding and critical assessment of AI-generated outcomes and helping to improve perceptions of how just and fair these systems are.

However, is trust a one-size-fits-all solution to the acceptance of AI within education, or is it more nuanced, where different AI applications require different levels of each factor on a case-by-case basis and for different stakeholders? This research seeks to determine to what extent each factor contributes to the acceptance and intention to use AI in education across four use cases from a multi-stakeholder perspective.

Drawing from this broad interdisciplinary foundation that integrates educational theory, ethics, and human-computer interaction, this study investigates the acceptability of artificial intelligence in education through a multi-stakeholder lens, including students, teachers, and parents. This study employs an experimental vignette approach, incorporating insights from focus groups, expert opinion and literature review to develop four ecologically valid scenarios of AI use in education. Each scenario manipulates four independent variables—agency, transparency, explainability, and privacy—to assess their effects on perceived global utility, individual usefulness, justice, confidence, risk, and intention to use. The vignettes were verified through multiple manipulation checks, and the effects of independent variables were assessed using previously validated psychometric instruments administered via an online survey. Data were analysed using a simple mediation model to determine the direct and indirect effects between the variables under consideration and stakeholder intention to use AI.