New Perspectives on the Future of Computing Education: Teaching and Learning Explanatory Models

Recent discussions in computing education research demonstrate uncertainties about which ideas, concepts and perceptions should be taught regarding topics like ML [e.g., 49, 43, 22]. These challenges are further underscored as understanding these technologies in detail is problematic within CS discipline too, resulting in research areas trying to find explanations for ML models (e.g., see discussions about explainable AI). Below, we discuss these challenges as a foundation for the development of the explanatory model approach.

Increasing complexity and decreasing comprehensibility. The challenges in understanding AI systems and comprehending their specific outputs become clear when taking into account the fast-growing research field about making AI systems (particularly those using ML techniques) understandable to people: explainable AI (XAI) [for an overview, see 2]. XAI research reports that AI systems with higher accuracy are less likely to be comprehensible and explainable [see 1, 23]. The current technological developments make understanding digital technologies increasingly complex and less transparent, while their implementation in daily life makes comprehensibility increasingly necessary.

The algorithms and the code for designing AI systems may be simple and comprehensible, but the resulting ML models can still be very complex and are rather black-boxes [32, 42]. Notably, the algorithms and codes play an important but relatively small role in ML systems [46]. Thus, understanding the code does not sufficiently explain the behavior of such systems. It also requires considering the role of data (e.g., training data selection) and their impact on the system’s behavior. Rahwan et al. [42] discuss different aspects of the complexity of such systems, including the high dimensionalities of ML models and the massive amounts of training data, imperfections in data, or the fact that the workings for generating output are learned by an ML system and not designed by the designer. Hence, AI systems and especially data-driven technologies are hard to understand. Knowing underlying conceptual ideas implemented in these technologies is probably insufficient for understanding them and their behavior. Thus, learning abstract concepts may not necessarily help students understand such technologies.

This demonstrates that the approaches we use to understand traditional algorithmic systems may not be sufficient for understanding data-driven systems, raising the question of what should be taught to students to effectively support their understanding of such technologies.

Paradigm shift in teaching about digital technologies. Similar discussions about differences in teaching about algorithmic systems and data-driven approaches can also be found in computing education research. For example, Tedre et al. [49] discuss differences in problem-solving and designing algorithmic solutions compared to data-driven problem-solving, highlighting paradigm shifts from traditional algorithms to ML. In recent years, many educational approaches for teaching students about AI concepts have been developed and discussed [e.g., 22, 34, 43, 9]. A review from Rizvi et al. [43] has shown that most of the materials published in articles about K-12 AI education do not cover concepts on the engines level (i.e. the concrete and formal underlying functionalities of AI algorithms). Instead, current materials seem to focus on teaching ML models (e.g., training and testing such models) and designing small AI-based applications [see 43]. This indicates that many educational approaches to AI education currently focus on explaining ML technologies and respective computational concepts instead of seeking to teach students about technical "ground truth" on an engine level. A similar observation can be made regarding one promising trend of using and developing educational tools, offering easy options for designing ML applications without requiring prior programming experiences [e.g., 24, 55, 28] [for an overview, see 21]. In doing so, students get insights on some aspects of the black-boxes of ML systems, while teaching focuses on a more abstract level, like training and testing ML models.

These debates on different paradigms and levels of teaching about AI and using educational tools seemingly focus on different contents and pedagogic approaches. However, there is probably an underlying question related to the idea of explanatory models: Which aspects of AI and ML should students learn? The different approaches, levels, and tools may target the same or different explanatory models. An idea argued in recent discussions is to build on our experiences from computing education about traditional, algorithmic systems, that is, adopting respective educational approaches to support students in developing mental models about AI systems [22]. For example, it is suggested to develop notional machines for AI systems [49]. Hence, we discuss this idea below as a foundation for developing the explanatory model approach, even if notional machines and explanatory models have significant differences, as we argue later.

The concept of notional machines is used in programming education and is to a certain extent similar to explanatory models [for overviews, see 17, 48]. A notional machine helps learners understand how programs and programming languages work by explaining what happens during program execution. It is a model for an idealized and conceptual computer [15] used as "a pedagogic device to assist the understanding of some aspect of programs or programming" [17]. It supports explaining programs, their behavior, and user-understandable semantics. For instance, notional machines can be analogies, such as boxes representing variables [for more examples, see 17]. Usually, a notional machine covers specific aspects (e.g., variables) but omits others. Based on mental model theory, Sorva [48] emphasizes explicitly teaching national machines rather than having them as implicit goals. Similarly, Munasinghe et al. [37] note that notional machines are often taught implicitly as vehicles or scaffolds rather than explicit models. This aligns with the mentioned perspective that often understanding the real ground truth is seen as the ultimate goal for computing education instead of learning about models: Even if notional machines are an example of teaching and learning models, they merely serve as pedagogical vehicles or didactic means to achieve a correct understanding, but not as the goal in itself. Understanding traditional algorithms led to grasping the ground truth (e.g., how variables work), making the need for models as end goals less critical, similar to the engine level of learning about AI described in the SEAME framework [47] [see also 43]. However, this seems to be changing with ML technologies (or even complex systems), where it is not so clear how they work in the smallest detail. Regarding AI education, developing notional machines for AI systems is suggested, although these may differ fundamentally from traditional notional machines [51]. For example, Munasinghe et al. [37] envision more abstract notional machines for AI systems.

Notional machines could be understood as one particular type of explanatory models related to program executions and their behavior. However, teaching notional machines focuses on providing a scaffold rather than primarily teaching models, in contrast to the idea of teaching explanatory models. While arguing for teaching explanatory models, a foundation of a notion of models is needed to be aware of essential characteristics of explanatory models.

In science research, developing models is a fundamental scientific practice used for reasoning and sense-making [40]. For instance, think about well-known models in natural sciences, like the models of atoms by Thomson, Rutherford, and Bohr or the DNA helix model by Watson and Crick. According to Osborne [39], models help when considered phenomena are not directly accessible (e.g., when they are too big or small to be observable). Due to the integral role of models in natural sciences, teaching and learning with and about models is essential in science education to reflect the scientific disciplines authentically [e.g., 53, 39, 40]. Science education research provides insights regarding teaching and learning with and about models, including several challenges [12]. For example, understanding models as selective and idealized representations of something with a specific state of scientific knowledge (and maybe resulting in models showing to be limited or outdated) rather than true images of reality can be challenging for learners [see 12].

While a broad discussion of the term ’model’ exceeds this paper’s scope, we note some insightful perspectives involving different disciplines concerned with the notion and meaning of models [for historical overviews, see 35]. The mathematician and theoretical computer scientist Mahr [35] describes models as representations that can be simplified and abstracted but could include additional properties. He introduces the ’cargo’ idea, describing that models convey information or knowledge about what they are meant to be used for. Regarding natural sciences, models are often characterized with two dimensions: they are representations of something and serving as tools for something [19, 40, 53]. They allow for illustrating, explaining, and communicating about phenomena. This scientific function of models is described with models as media (referring to the of-perspective) [53]. Additionally, models provide predictions for phenomena so that they can be used as a tool for generating ideas and knowledge about the considered phenomena (referring to the for-perspective) [53]. Thus, models have a representative perspective and an explanatory and predictive function. Notably, the development and use of a model is related to specific purposes and intentions [18]. Passmore, Gouvea, and Giere [40] argue that the of-for-distinction can support to underscore the functional perspective of using models, highlighting that science education is not limited to the fact that models are one more thing students need to learn and reproduce but instead being enabled to use models as tools for reasoning and sense-making. However, science education practice often focuses on the representative perspective and neglects functional perspectives [19]. From a computing education view, this functional perspective could relate to using models to explain computational concepts and to explore, explain, and reason about specific digital artifacts and their inner workings.

In science, various types of models are used, such as mathematical, theoretical, chemical, or analogical models [see 11]. As one type, Clement [11] describes explanatory models as explaining how and why something works. In addition, scientific, theoretical, and hypothesized models provide views on the world, but they should be distinct from just observational models (e.g., measurements or representative descriptions). Respective models used in science education are often developed, tested, and refined by scientists over time (e.g., involving experiments).

When relating the role of models in science (education) to computing education, crucial differences between these disciplines should be noted. For example, science education considers natural phenomena, while computing education is about digital artifacts, referring to objects invented and made by humans. We discuss this later in more detail.

What is an explanatory model? An explanatory model is a conceptual model that is an idealized representation of computing objects, such as computational concepts, digital artifacts, or socio-technical systems (i.e. a composition of human beings and digital artifacts). Such models should fulfill educational purposes of providing specific perspectives on and explanations of computational concepts and digital artifacts and their behavior. For example, an explanatory model for data-driven technologies can focus on the role of data [see 25, 26]. They allow students to explain computational concepts or explore and make sense of inner workings and contextual effects of specific digital artifacts.

Why do we use the term ’explanatory models’? While traditional concepts and algorithmic systems could be explained in full detail with the technical truth, this is challenged by the complexity of real-life digital artifacts, current technological developments, and paradigm shifts in the discipline (e.g., regarding AI and ML). Based on the notion of models discussed earlier, the following perspectives could be adopted (especially from the of-for-distinction): First, explanatory models represent computational concepts and digital artifacts, thereby focusing on specific aspects. Frameworks from computing education could assist in this regard, for example, to provide different levels of consideration on AI systems [e.g., 47] or a set of various aspects for programs [e.g., 44, 33]. This representational perspective of explanatory models can cover different degrees of complexity and technical precision. For example, an explanatory model of large language models could use the metaphor of stochastic parrots [see 5]. Second, explanatory models serve as tools for exploring and explaining digital artifacts, which involves making sense of their inner workings and reasoning about their behavior. From this perspective, an explanatory model highlights different aspects of digital artifacts and serves as a tool for uncovering inner workings of specific technologies. For example, a mentioned explanatory model of a large language model could then be used to explore and explain the outputs of a specific text generation tool.

Even if building on the notion of models from science, the representational view of models in computing education differs from that in natural sciences. For example, when considering a stone (as a natural object), the question of whether it is a good stone makes no sense as it is just a stone; in contrast, digital artifacts are based on human decisions and values so such a question makes sense. This example illustrates that explanatory models related to digital artifacts or socio-technical systems require other characteristics than models in science and science education. The dual nature theory supports this characteristic property of explanatory models, which stems from philosophical debates on the nature of technical artifacts and is used, for example, in computing or technology education [e.g., 44, 10]. While comparing explaining natural phenomena and digital artifacts, de Ridder [13] emphasizes that digital artifacts should be considered with the human agency, while this is different about natural phenomena. According to this theory, two views can be referred to when understanding or explaining digital artifacts [30, 44, 54]. While the perspective of the architecture or structure of digital artifacts refers to its inner workings (e.g., the algorithmic workings), the perspective of the relevance is related to an interpretation of the intentions, meanings, and social effects of the digital artifact (e.g., why it was designed as it is, or how can it be used). Notably, different aspects of a digital artifact (e.g., an algorithm) can be described from both perspectives. The theory states that digital artifacts are not neutral as different human purposes and intentions are included. The two perspectives are interrelated so that understanding the relevance is necessary to comprehend inner workings; conversely, understanding the architecture is a prerequisite for evaluating the relevance [see 45, 30]. Accordingly, it is argued that a comprehensive understanding of a digital artifact involves both perspectives. For example, digital artifacts often impact individuals and society. Rahwan et al. [42] describe several examples of potential influences of AI systems. For instance, systems choose information (or misinformation) people see in their news feeds, which can influence individuals’ behaviors, emotions, and opinions [29, 52]. Accordingly, models about digital artifacts should not be representations of inner workings alone (the architecture perspective) but also - and this is important - should provide explanations for their functions, meanings, and impacts (the relevance perspective).

In summary, explanatory models always draw on architecture and relevance perspectives to highlight the interpretative and explanatory functions. This allows students to reason about digital artifacts, their behavior, and their influences on individuals and society.

Important difference between science and computing. Models in natural science and science education are developed and used to understand natural phenomena. In contrast, in computing education, models and modeling are used to understand and develop digital artifacts rather than primarily to understand given phenomena of the analog world. (Note that some digital artifacts are indeed designed to examine and understand natural phenomena, such as computational simulations, but this specific case deserves in-depth discussion [e.g., 41].) However, designing digital artifacts in computing also impacts the analog world, which is also considered in computing but always in relation to digital artifacts. Thus, models for computing education have other and additional functions than those known in science. Nevertheless, the experiences in teaching and learning with and about models in science education can be relevant to teaching and learning explanatory models, especially those related to technologies we struggle to understand and explain. This is similar to the argumentation of Rahwan et al. [42], who argue for adopting methods from natural sciences to investigate and understand the behavior of AI systems.

Based on the characteristics of explanatory models, we discuss use cases for teaching and learning practice and research in computing education.

Students learn explanatory models. Similarly to notional machines that are argued to be explicitly taught [see 17, 48], explanatory models are intended to be made explicit, that is, learning them becomes a concrete learning goal. However, while notional machines are meant as scaffolds or vehicles to support students in learning the ground truth of algorithmic programs, explanatory models are meant as a goal in themselves, including teaching students about these models and enabling them to work with them. Thereby, explanatory models aim to support students’ metacognitive thinking processes, such as reflecting on different perspectives on computational concepts or the nature of computing; similarly, in science education, learning about and with models is argued as serving metacognitive tools [e.g., 12]. They provide a perspective or lens on computational concepts and digital artifacts. For example, this allows them to explain abstract concepts (e.g., algorithms or large language models) or to explore, explain, and reason about specific digital artifacts, their inner workings, and their behavior (e.g., particular chatbots). This could include using an explanatory model as a lens on large-language model applications, allowing them to find explanations for their behavior, such as reasoning about the generated text. Explanatory models can also be used for designing digital artifacts, such as applying an explanatory model of neural networks to design an ML application. Explicitly learning and using explanatory models also involves learning about their purposes, allowing students to critique and reflect on the contexts in which they can use a model or when it may not be applicable.

Educators use explanatory models. Following explanatory models as learning content, this approach can also help educators design teaching units and materials accordingly. Teaching about explanatory models is meant to support students in forming respective mental models, which are mental constructs (i.e. knowledge structure), for example, related to systems or algorithmic processes [27]. Mental model theory describes how people perceive, explain, and predict the world [20]. It is often used in computing education research, like in the approach of notional machines [e.g., 17] or when examining students’ conceptions about topics like AI [e.g., 36]. While mental models are personal and internal representations, the external counterparts are conceptual models [20, 38, 48]; like explanatory models. Their representations could include analogies, artifacts, or visual explanations of systems [38, 48]. Conceptual models can be useful in teaching to explain structures and inner workings of digital artifacts [48]. For example, a study from Ben-Ari and Yeshno [4] indicates that learning a conceptual model about internal word-processing software supported school students in developing conceptual understanding and interacting with the software. Similarly, a study suggests that learning an explanatory model can help students understand everyday technologies [26]. However, based on experiences in science education, we should note that presenting conceptual models does not necessarily lead to adequate mental models as students often lack knowledge from the discipline or the domain to interpret the presented models [20].

This relation of explanatory models to intended mental models leads us to their potential support for diagnosing students’ conceptions. Research on students’ perspectives on computational concepts or their preconceptions can be understood as examining explanatory models students know, such as in the study about algorithms mentioned before [see 3]. This can help teachers choose explanatory models to introduce in a learning group. Additionally, explicating these explanatory models can help students reason about computational concepts. For example, using a recipe analogy as an explanatory model of algorithms could include discussing which aspects of algorithms this analogy could adequately represent (e.g., step-wise operations) and where it has its limitations (e.g., loops may then be problematic).

While using notional machines or analogies as vehicles to teach computational concepts is not new, the presented approach calls for making explanatory models explicit and taking learning and using them as one of the primary learning goals. Then, explanatory models serve as learning content but also provide diagnostic functions and can support developing interventions.

In addition to using explanatory models in computing education practice, we envision further potential use in research practice. The diagnostic function of explanatory models mentioned above also applies to research practice, such as studies on students’ conceptions of computational concepts. Thus, explanatory models can serve as lenses for empirical research, inspiring the development of respective research instruments. Explicating explanatory models used in intervention studies could benefit communication about rationales and pedagogical ideas and, hence, cumulative knowledge building. For example, clarifying the explanatory model taught in a study’s intervention could help researchers compare the results with other studies on the same explanatory model but with other pedagogical approaches.

Discussions on explanatory models are also related to discussing the nature of the CS discipline and computing education (e.g., caused by the fundamental changes regarding new technologies). The approach challenges the idea that we can teach and explain outcomes of the discipline based on reductions and by breaking them down to the smallest algorithmic details. We may need to rethink whether this approach holds for new computing topics. Practices of working with explanatory models could imply that computing education may need to focus more on scientific traditions in computing education [see 50], also in conveying a coherent and authentic image of the discipline.