Despite growing interest in artificial intelligence in education, there remains a notable research gap concerning how AI, specifically explainable artificial intelligence (XAI), can potentially support and enhance students' metacognitive abilities and computational thinking (CT). To bridge this gap, we propose MetaCoXAI, a novel conceptual framework that integrates XAI with computational thinking instruction, ofering actionable strategies for learners to develop a deeper understanding of AI processes. Grounded in interdisciplinary theoretical insights from learning technologies, human-computer interaction, machine learning, and XAI, MetaCoXAI explicitly targets the four fundamental components of computational thinking: abstraction, decomposition, algorithm design, and debugging. The framework illustrates how XAI facilitates these CT processes, thereby positively influencing learners' metacognitive skills. To demonstrate the practical utility and application of our proposed framework, we provide research directions highlighting how learners can utilize XAI-supported computational thinking to enhance both problem-solving proficiency and AI competency.
With the growing popularity of AI, its potential for innovation in the field of education is tracking
attention [
To bridge this gap, we propose a novel conceptual framework that connects computational thinking
(CT), metacognition, and explainable artificial intelligence (XAI). Metacognition can be defined as
“cognition where the information on which a learner operates describes features of cognition” [
With this logical connection established between existing gaps and educational goals, we propose
the MetaCoXAI framework, which links these three concepts both theoretically and practically. First,
drawing from theoretical insights proposed by Yadav et al. [
Our overarching research questions are as follows: “How can theories of computational thinking and metacognition and explainable artificial intelligence be integrated to support learners? ” Our main contributions are twofold: (1) MetaCoXAI can assist researchers and developers in designing and evaluating new XAI systems that support the four phases of CT—abstraction, decomposition, algorithm design, and debugging—and (2) it enables educators to help students build metacognitive skills through XAI-enhanced learning environments, ofering implications for educational researchers.
This paper focuses on facilitating computational thinking (CT) skills and establishing a connection
between CT and the theoretical concepts of metacognition. CT has gained significant attention among
educational researchers as one of the key skills for the 21st century [
The key concept closely related to CT is metacognition, which simply refers to “thinking about
thinking” [
In the context of CT, metacognition occurs when learners reflect on their computational methods to
solve human problems. In fact, metacognitive practices are considered to be one of the key evaluation
approaches to CT [16]. The process of CT and metacognition can work in parallel [
The fundamental idea behind XAI aligns with the explanatory aspect of CT, as its goal is to “enable end users to understand, trust, and efectively manage their intelligent partners” [ 17]. We propose that XAI can serve as a powerful computational tool that provides algorithmic explanations, fostering a Abstraction Problem Discovery
Decomposition
Problem Definition
Algorithms Algorithmic Decisions
Debugging Alternative Solutions
Optimal
Solution Role of XAI in supporting
Algorithms & Debugging - Presenting feature contributions to solutions - Evaluating top-ranked metrics and demonstrating relevance to solutions - Identifying rules and adjusting weights and policies - Presenting alternatives (conterfactuals, contrastive explanations) - Supporting developing and testing hypotheses
Learners can generalize observations Learners can observe possible causes and correlates
Learners can find key criteria & metrics Learners can navigate alternative AI approaches Learners can get insights from AI-driven decisions
XAI Queries (adapted from Liao, Gruen and Mil er, 2021) Why: Why/how is this instance given this prediction? What not: Why is this instance NOTpredicted to be [a dif erent outcome Q] How to be that: How should this instance change to get a dif erent prediction Q? How to still be this: What is the scope of change permitted for this instance to stil get the same prediction? What if: What would the system predict if this instance changes to? ?
Feature Importance - Individual feature contribution (e.g.,SHAP) - Interactions between features
Hypothesis Testing - Counterfactual explanations via generative models (e.g., GANs), optimization-based model (e.g., LIME) - Contrastive explanation (e.g., Bayesian Rule Lists) trso iiton p n p g suTC tcaoeM
Role of XAI in supporting
Abstraction & Decomposition - Understanding the problem context and situations - Presentation of relevant data - Supporting learners to undestand the essential problem I- Assessing strategies to solve the overarching Aproblem and dividing tasks to tackle X f- Providing patterns to assist problem-solving o e Glolobal & Local Visualiation
R - Holistic, specific (e.g., Partial Dependence Plot, saliency map)
Reasoning Process
- Abductive, deductive, inductive
reasoning (e.g., Decision tree,
linear equation)
deeper comprehension of computational concepts and demonstrating algorithmic thinking. Therefore,
we suggest that among the various approaches to CT, focusing on metacognition can assist learners
through the utilization of XAI methods. According to Miller et al. [18], explainers wanted to generate
explanations to allow explainees to learn either through the simplification of observed examples [ 19] or
through the generalization of such observations that allow people to predict future events [20]. While
transparency in presenting the algorithmic decision-making process is one of the key benefits of XAI
methods, making the outputs generated by AI more understandable and interpretable can support
learners in better understanding complex problems and help learners develop stronger computational
thinking skills. In the next section, we present how XAI can assist learners in facilitating computational
thinking (CT) skills, encompassing the four steps of abstraction, decomposition, algorithms, and
debugging, as proposed by Yadav et al. [
Our MetaCoXAI framework primarily focuses on designing and developing XAI for learning purposes
and holds significant educational implications. In fact, the notion of explainability in XAI aligns with
the concept of CT, and this brings forth possibilities of utilizing XAI as an instructional instrument for
CT. Because of XAI’s explainability, learners will be able to understand how AI algorithmic functions
work [21], which could potentially benefit learners to utilize AI systems to solve problems using CT
skills. Recently, there have been discussions on how XAI can support learning. Khosravi et al. [22]
list “agency, student-teacher interactions, AI literacy, accountability and trust” as the main benefits of
XAI in education. In our model, we focus on XAI’s capacity to facilitate computational thinking and
metacognition. We posit that XAI supports computational thinking process and metacognitive activities
in relation to problem solving. With this in mind, Figure 1 illustrates how the MetaCoXAI framework
supports each process of CT (abstract, decomposition, algorithms, and debugging), which is closely
associated with metacognition. First, building upon the work of Yadav et al.’s [
Problem solving becomes evident when students learn new topics and reflect on them in the classroom;
however, setting a good problem for students to solve independently poses a challenge when they lack
the necessary knowledge, specific information, and relevant data. While abstraction is defined as a
metacognitive process of identifying the problem and its parameters [
As such, by presenting “how algorithms work” with the assistance of XAI, students can explore data (input factor) more transparently, and it helps them to attain relevant data samples related to the specific topics they are learning and understand how each data point contributes to the targeted variable (output factor). XAI can help learners understand AI better by displaying how each AI function works and understand the given problem through the lens of AI. XAI facilitates their knowledge discovery process. By exploring data and understanding how AI works and why it makes certain decisions, students are empowered to generate new and novel questions that they have not previously explored.
Decomposition is the process of breaking down a large problem into sub-problems (e.g. at the task level). XAI ofers users the ability to explore data at both the global and local levels of model generation. Specifically, through the use of XAI’s Local Interpretable Model Explanations, students can examine how a particular example could be related to the targeted predictive value or problem. This allows learners to uncover interesting trends and relevant features concerning the issue of the subject matter and raises awareness of how data samples are associated with the problem.
Within the realm of XAI, learners can benefit from gaining holistic perspectives on data exploration. Global agnostic models play a crucial role in facilitating data exploration, as they allow learners to develop a comprehensive understanding of the decision-making process employed by AI systems. By leveraging techniques such as feature influence and relevance, learners can efectively grasp the rationale behind AI-generated decisions. The utilization of global models provides learners with a broader understanding of data and patterns, enabling them to identify high-level problems with greater clarity.
Given the abundance of information and data looking for goals to achieve, it is essential to identify both the positive factors that contribute to achieving goals and the obstacles that hinder them. Global explanations amongst XAI techniques aim to describe how the entire machine learning model operates and assist learners in understanding how each factor influences the output. Global explanations provide insights into the relationship between each feature and predictions [22].
Amongst the diferent XAI strategies, one commonly used approach is Shapley Additive exPlanations (SHAP), which utilizes Shapley values derived from game theory [24, 25, 26]. SHAP allows learners to understand the key contributors and the direction (positive or negative) of their impact on the machine learning model. It provides a baseline of the model by presenting the mean predictive value, and presents how far or close each feature is from the base data point and identifies deviation from the average value (baseline). For this reason, learners can interpret the contribution of each feature.
According to Yadav et al., [
Importantly, XAI supports the algorithmic process by helping learners approach diferent aspects and approaches to finding solutions. While AI is often considered a black-box model, XAI reveals the specific steps algorithms take to execute a model. Learners have the opportunity to observe how algorithms solve problems. For example, rule-based models like decision trees and regression demonstrate how algorithms generate logic and rules using the given data to produce expected outputs. Learners can gain insights into the reasoning behind algorithmic decisions, even if some decisions may seem counterintuitive. XAI makes algorithmic decisions more interpretable, allowing learners to understand how and why such decisions are made. In addition, instance or example-based understanding, with the use of XAI visualization, helps learners explore diferent aspects of AI-generated decision in a transparent way.
Debugging is defined as finding the best solution and exploring alternatives, where learners evaluate the current solution and navigate possible alternatives to maximize the output. The next step is to assess the reliability of the predictive value (output) and find a way to optimize the model. Understanding the contribution of each factor in the data on the prediction produced by Feature Attribution and Importance of XAI as shown in [25, 27], is critical because learners can test diferent algorithms, data, and parameters by evaluating both individual level of features (weight) and identifying the top-ranked set of features to optimize the model. At the same time, learners can exclude irrelevant data that might negatively impact the predictive value, therefore maximizing the efectiveness of the model.
Nevertheless, not all XAI approaches are intuitive and valuable to learners, especially those without prior programming experience, due to the high AI literacy required. XAI strategies should be wellaligned with clear learning purposes and expected learner outcomes. Aside from that, XAI strategies can be beneficial for learners to explore data, analyze and define problems, facilitate understanding and reasoning process, and generate possible solutions. They also support the evaluation and debugging of output performance.
XAI supports debugging by assisting learners in finding alternative solutions. For example, counterfactuals explanations [28] enable learners to experiment with algorithm-generated hypothesis testing which provides possible conditions of feature combinations that afect predictions (e.g., diferent counterfactual outputs of Counterfactual Explainer). Learners can discover alternatives they had not previously considered but recognize hypothetical adjustments to specific input factors that could change the algorithm’s output, for example, from a positive to a negative influence. Such methods empower students to create alternative solutions. By looking at contrastive explanations, students can learn how to come up with alternative solutions by observing how AI determines which input factors produce contrastive examples. It demonstrates why something plausible did not occur by comparing it to why it did happen. Certain XAI approaches can facilitate hypothesis testing, where students explore diferent data or types of algorithms to see how these changes yield new findings (what-if explanation). This process can provide students with alternatives that may be superior to the current solution.
The proposed MetaCoXAI framework contributes to the theoretical understanding and practical
implementation of XAI within CT to enhance metacognitive skills among learners. However, while the
theoretical relationships outlined here establish a foundational perspective, their empirical validation is
essential to substantiate and refine these conceptual connections. Several promising future research
directions emerge from this work. First, integrating generative AI, particularly advanced large language
models such as GPT-4o [
Second, the development of personalized and adaptive learning environments leveraging XAI present another valuable direction. Research by Holstein et al. [31] demonstrates that adaptive educational environments efectively support individual learners’ cognitive and metacognitive skills, dynamically adjusting instructional methods based on real-time learner feedback and data-driven insights. Future research should explore how XAI specifically facilitates the scafolding of metacognitive skills within adaptive computational thinking tasks.
Third, using XAI to enhance AI literacy and ethical decision-making aligns directly with critical metacognitive skills, particularly critical reflection and evaluative judgment. According to recent ifndings by Tankelevitch et al. [ 32], explicit, transparent explanations of AI-driven decisions significantly enrich learners’ understanding of algorithmic processes, thereby reinforcing ethical reasoning and metacognitive awareness. Khosravi et al.[22] suggest that transparent AI systems that clearly communicate the reasoning behind algorithmic decisions not only enhance learners’ understanding but also promote ethical and critical reflection. This alignment reinforces the educational value of the MetaCoXAI framework by integrating moral reasoning within computational thinking processes. Future studies should empirically assess the MetaCoXAI framework’s efectiveness in fostering such ethical and reflective competencies.
Fourth, leveraging advanced learning analytics with XAI for systematic assessment of computational thinking and metacognitive processes could substantially improve educational interventions. Matcha et al. [33] highlight that analytics-driven feedback can facilitate timely instructional adjustments. Empirical exploration of these analytics approaches within the MetaCoXAI framework will clarify their practical impact on learner outcomes.
Last, ensuring educator preparedness through targeted professional development programs is vital for the successful integration of XAI in classrooms. Future work should focus on developing, implementing, and evaluating teacher training modules that specifically address educators’ skills and attitudes towards XAI, as proposed by Vuorikari et al. [34]. This would enable broader adoption and deeper understanding of computational thinking pedagogies supported by XAI. There are a few limitations. First, since this is a preliminary work that focuses on a conceptual framework, future studies need to empirically test the the framework in educational settings. Second, the theoretical framework aims to provide general overview of the relationships between XAI, computational thinking, and metacognition and is not tailored to various learning contexts or learner characteristics. Hence, it would be important to focus on specific use cases in future studies.
While previous research has explored the advantages of XAI in learning and human decision-making domains, limited understanding exists regarding how XAI can efectively facilitate the process of CT in conjunction with metacognition. This paper aims to present a conceptual framework, called MetaCoXAI, that elucidates the intricate relationship between XAI, CT, and metacognition. First, we conduct a comprehensive review to analyze the interconnectedness between computational thinking and metacognition. Second, we argue that the integration of XAI, CT, and metacognition exemplifies a human-centered design approach to artificial intelligence and educational practices. Guided by this perspective, we introduce MetaCoXAI, a framework that provides detailed insights into how XAI can efectively foster and enhance CT. Our study introduces the MetaCoXAI framework, which is founded upon a robust theoretical foundation that aligns the various components of XAI with the goal of supporting CT. Specifically, we assert that XAI has the potential to facilitate metacognitive practice among learners, thereby enhancing the CT process and ultimately leading to improved performance and productivity in learning contexts. Our contribution lies in providing a comprehensive explication of how the components of XAI can be interconnected with the four fundamental concepts of CT from the learners’ perspective: abstraction, decomposition, algorithm, and debugging.
The author has not employed any Generative AI tools. [15] R. E. Mayer, Cognitive, metacognitive, and motivational aspects of problem solving, Instructional science 26 (1998) 49–63. [16] Y. Allsop, Assessing computational thinking process using a multiple evaluation approach, International journal of child-computer interaction 19 (2019) 30–55. [17] D. Wang, Q. Yang, A. Abdul, B. Y. Lim, Designing theory-driven user-centric explainable ai, in:
Proceedings of the 2019 CHI conference on human factors in computing systems, 2019, pp. 1–15. [18] T. Miller, P. Howe, L. Sonenberg, Explainable ai: Beware of inmates running the asylum or: How i learnt to stop worrying and love the social and behavioural sciences, arXiv preprint arXiv:1712.00547 (2017). [19] T. Lombrozo, Simplicity and probability in causal explanation, Cognitive psychology 55 (2007) 232–257. [20] F. Heider, The psychology of interpersonal relations, Lawrence Erlbaum Associates, Hillsdale, NJ, 1958. [21] R. Farrow, The possibilities and limits of explicable artificial intelligence (xai) in education: A socio-technical perspective, Learning, Media and Technology 48 (2023) 266–279. [22] H. Khosravi, S. B. Shum, G. Chen, C. Conati, Y.-S. Tsai, J. Kay, S. Knight, R. Martinez-Maldonado, S. Sadiq, D. Gašević, Explainable artificial intelligence in education, Computers and education: artificial intelligence 3 (2022) 100074. [23] Q. V. Liao, D. Gruen, S. Miller, Questioning the ai: informing design practices for explainable ai user experiences, in: Proceedings of the 2020 CHI conference on human factors in computing systems, 2020, pp. 1–15. [24] S. M. Lundberg, G. Erion, H. Chen, A. DeGrave, J. M. Prutkin, B. Nair, R. Katz, J. Himmelfarb, N. Bansal, S.-I. Lee, From local explanations to global understanding with explainable ai for trees, Nature machine intelligence 2 (2020) 56–67. [25] S. M. Lundberg, S.-I. Lee, A unified approach to interpreting model predictions, Advances in neural information processing systems 30 (2017). [26] L. Shapley, 17. a value for n-person games, in: Contributions to the Theory of Games, Volume II,
Princeton University Press, 1953, pp. 307–318. [27] M. T. Ribeiro, S. Singh, C. Guestrin, " why should i trust you?" explaining the predictions of any classifier, in: Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining, 2016, pp. 1135–1144. [28] M. T. Ribeiro, S. Singh, C. Guestrin, Anchors: High-precision model-agnostic explanations, in:
Proceedings of the AAAI conference on artificial intelligence, volume 32, 2018. [29] E. Kasneci, K. Sessler, N. Kiselev, G. Kasneci, Chatgpt for good? on opportunities and challenges of large language models for education, Learning and Individual Diferences 105 (2023) 102274. [30] T. H. Feng, A. Luxton-Reilly, B. C. Wünsche, P. Denny, From automation to cognition: Redefining the roles of educators and generative ai in computing education, arXiv preprint arXiv:2412.11419 (2024). URL: https://arxiv.org/abs/2412.11419. [31] K. Holstein, B. M. McLaren, V. Aleven, Intelligent tutoring systems in the era of generative ai, AI
Magazine 44 (2023) 42–54. [32] L. Tankelevitch, V. Kewenig, A. Simkute, A. E. Scott, A. Sarkar, A. Sellen, S. Rintel, The metacognitive demands and opportunities of generative ai, arXiv preprint arXiv:2312.10893 (2023). URL: https://arxiv.org/abs/2312.10893. [33] W. Matcha, N. A. Uzir, D. Gašević, A. Pardo, A systematic review of empirical studies on learning analytics dashboards: A self-regulated learning perspective, IEEE Transactions on Learning Technologies 13 (2020) 226–245. [34] R. Vuorikari, W. Holmes, Y. Punie, Educating teachers to navigate the AI world: Guidelines for teacher education and professional development, Technical Report JRC129830, European Commission Joint Research Centre, 2022.