X-IDEA: Explainable, Inclusive, Equitable Code-Learning Assistants

1 Introduction First-year programming courses persistently record the highest attrition in computing programmes. International reviews place failure or withdrawal between 30% and 50% [1, 2, 3]. Attrition is heavier among students who identify with groups historically excluded from computing, a pattern that limits workforce diversity and weakens social mobility [4]. Large-language-model assistants such as GitHub Copilot or ChatGPT offer rapid feedback and abundant worked examples, which seems attractive to instructors facing crowded classrooms. Empirical work, however, shows that students perceive these systems as inscrutable, that teachers worry about plagiarism and bias, and that institutions struggle with privacy regulations [5]. A recent qualitative content analysis of 156 Reddit posts and 3512 comments, grouped into twenty topics, paints a similarly ambivalent picture: enthusiasm for rapid feedback is matched by anxiety over explainability and academic integrity [6]. These forum voices supplied many of the requirements that shaped X-IDEA. The present contribution introduces X-IDEA, a design framework for explainable, privacy-aware, and teacher-governed code-learning assistants. The framework is theoretical; it offers constructs and propositions that can be examined in future empirical studies. 1.1 Motivation and contribution Introductory programming (CS1) still records 30–50% failure or withdrawal rates, with disproportionately higher attrition among different groups [1, 2, 3]. Existing AI assistants such as Copilot or ChatGPT provide fast feedback but raise concerns about opacity, plagiarism, bias and data-protection regulations [4, 6]. We propose X-IDEA, a design framework that integrates five layers, Learning Objectives, Learner Modelling, Generative Support, Explainable AI and Teacher Governance, under a set of design principles explicitly aligned with the challenges of remote and bandwidth-constrained higher-education contexts. The present paper advances three concrete contributions beyond previous drafts: a tighter mapping between each layer and the practical constraints of distance programming labs (e.g. intermittent connectivity, privacy legislation, large cohorts); an expanded theoretical grounding that links X-IDEA to the state of the art in explainable AI for education, equity-centred design and programming-specific tutoring; a clarified empirical roadmap with explicit study designs and measurable outcomes. 2 The X-IDEA framework X-IDEA is organised in five interacting layers. Each layer addresses one pressing opportunity paired with a technical or ethical challenge. Design principles (DP) are printed in italics. 1. Learning objectives. A graph of concepts spans syntax, semantics, and problemsolving patterns, indexed to Bloom levels. DP1concept alignment: every AI action carries a pointer to a concept node, which later enables transparent assessment. 2. Learner modelling and personalisation. Local logs of compilation outcomes, test traces, and debugging paths. DP2 Minimal-Intrusion Data: only telemetry created inside the development environment is stored, and all raw code remains on institutional servers. 3. Generative support. Several lightweight language models synthesise hints, partial solutions, or culturally localised storylines. DP3 scaffold-before-solve: the default output is a Socratic question or micro-hint; direct code appears only when earlier prompts were exhausted. 4. Explainable AI. Two explanation modalities are provided: token-level concept labels and counter-factual test simulations. DP4 dual-audience transparency: novice-friendly text expands, on demand, into a technical rationale that teachers can audit. 5. Teacher-in-the-loop governance. Dashboards visualise aggregate model accuracy, bias indicators, and student misconceptions; teachers can approve, edit, or block any suggestion. DP5 human override: every override is logged and feeds back into the learner model and bias detector. 2.1 Development process Design requirements were triangulated against insights from an earlier Reddit corpus study [6]. For example, the rule DP3 scaffoldbefore-solve mirrors the forum’s recurrent complaint that “Copilot gives you code before you know what it does,” whereas DP5 Human Override responds to lengthy discussions in which learners begged instructors to “step in and set boundaries.” The framework materialised through design-science research. First, literature on programming misconceptions, explainable AI, and AI ethics in education was reviewed to extract requirements. Second, low-fidelity prototypes were sketched, then inspected by three CS1 lecturers in semi-structured interviews. Their feedback informed the five design principles above. Third, theoretical propositions, reported in Section 3, were generated to guide future evaluation. 3 Research propositions P1 Transparency fosters trust. Concept-level explanations are expected to improve perceived trust, which in turn mediates self-regulated learning. P2 Scaffold depth predicts learning gain. Hints that progress from question to partial code, tuned to mastery, will raise post-test scores relative to either immediate solutions or no assistance. P3 Governance engagement affects equity. Frequent teacher interaction with the governance layer will narrow achievement gaps across demographic groups. P4 Privacy architecture shapes adoption. Institutions with limited connectivity will declare higher adoption intent when models run on local servers despite lower accuracy. P5 Explainability has an optimum. Excessive detail can overload novices; a middle ground between brevity and completeness must be located. First, in late 2025 a small group of engineering lecturers will read the framework and give straightforward feedback on what makes sense and what does not. Next, in early 2026 the assistant will be tried in a single course; we will record how much students learn from before to after the course and ask them how easy the tool feels to use. Finally, throughout 2026 the assistant will be deployed at several courses; we will compare a version that runs on each campus’s own servers with one that runs in the cloud, and we will examine whether the system increases trust, improves learning, and reduces performance gaps between different kinds of students. 4 Roadmap and conclusion Phase 1 Design validation: The twenty-topic Reddit taxonomy will be juxtaposed with expert walkthroughs: lecturers will judge whether each design principle truly addresses the community-voiced concerns catalogued by Di Leo [6]. Phase 2 Pilot deployment: A single CS1 laboratory adopts an early X-IDEA prototype for one semester; quantitative logs and qualitative interviews will be collected. Phase 3 multi-site studies: Quasi-experimental trials across several institutions test propositions P1–P5 under diverse curricular and infrastructural conditions. X-IDEA offers a synthesised blueprint for building AI coding assistants that are explainable, inclusive, privacy-aware and teacher-governed, with explicit affordances for remote learning infrastructures. Upcoming empirical work will assess whether the framework can indeed lower attrition while safeguarding academic integrity and learner equity.