Generative Artificial Intelligence (AI) has demonstrated significant value in code generation and support for programming tasks, leading to its widespread adoption in both industry and academia. This proliferation has introduced new opportunities and risks for learning, teaching, and the broader landscape of computer science education. Despite this rapid integration, there remains limited understanding of how students engage with these tools, particularly in terms of collaboration, dependence, and delegation. More critically, little is known about how students regulate their interactions with AI tools over time, and how such processes contribute to problem-solving. In this case study, we coded 2,376 interactions from 120 undergraduate students with ChatGPT in a web programming course using a theoretically grounded self-regulated learning (SRL) in problem-solving coding scheme. To move beyond static counts and capture the temporal and structural dynamics of these interactions, we employed novel process-oriented learning analytics: transition network analysis and Sequence Analysis. Our findings reveal that while students frequently employed regulatory prompts aimed at process monitoring and problem-solving, they rarely engaged in deeper metacognitive strategies such as reflection or evaluation. This suggests a prevailing focus on surface-level regulation over deeper learning processes in student-AI interactions. Our results serve as a critical warning, highlighting a tendency towards 'cognitive offloading' that could undermine the development of independent, lifelong learners in computer science.
Neurodiversity remains underrepresented in higher education, especially in computing, where inclusive design is seldom embedded in curricula. We report an empirical study of a participatory inclusive IT design course at a German university of applied sciences. Twelve IT students collaborated with a neurodivergent peer expert and the university's accessibility advisor for students with disabilities and chronic health conditions to co-create prototypes addressing concrete barriers in everyday academic life. A mixed-methods evaluation combined entry and exit student surveys with a multi-stakeholder focus group. We find a statistically significant gain in self-reported knowledge of accessible design, alongside positive but non-significant shifts in attitudes and confidence, consistent with small samples and high pre-course endorsement. Qualitatively, participants described conditions that made participation workable, and stakeholders perceived the prototypes as valuable for neurodivergent students and potentially for students of all neurotypes. The study offers empirically grounded recommendations to embed participatory, inclusive practice in computing education.
We report two iterations (2023 and 2024) of a service-learning course in inclusive and accessible computing. Students co-designed with people with disabilities, including participants with complex support needs. Using a mixed-methods design with four measurement points per cohort, we combined validated survey scales with longitudinal, semi-structured interviews. The 2023 offering had limited field contact; the 2024 offering embedded four on-site co-design sessions across the term. Quantitatively, group means on global scales changed little from pre-contact to end of term, consistent with high baselines and elective enrollment, with small gains in self-efficacy. Qualitatively, repeated contact supported more empathetic framings, clearer evidence-anchored design rationales, and growing practical competence, with these narratives voiced more frequently in the higher-contact offering. Descriptively, the elective attracted a comparatively diverse set of students and a baseline profile consistent with reflective, prosocial orientations. We discuss implications for structuring sustained contact and scaffolding co-design in accessibility- and inclusion-oriented service-learning, and we argue that mixed-methods evaluation is essential to capture proximal, course-specific developments that global instruments may not detect over a single semester.
Security competitions in computing education model real-world crowd-sourced security (e.g., bug bounty programs), which has grown dramatically in recent years and has become the main source of software security reviews for many companies. This study explores practices and dynamics of high-performing teams of high school computing students in a K-12 computer security competition. While existing literature covers dynamics of security teams in structured organizations and student team formation for software engineering projects, additional insights are to be gained from this setting where students in this study a) self-selected into remote or hybrid teams, and b) developed collaborative strategies that enabled them to outperform thousands of teams on computer security tasks. We aim to fill this gap by addressing the following research questions regarding high-performing teams of high school computing students collaborating on computer security tasks: (1) How, why, and through which assembly mechanisms do teams form? (2) What technological needs do teams have, and how do they address them? (3) How do teams scope and distribute tasks, roles, and responsibilities? (4) What motivations and limitations do these teams have? (5) What factors and strategies seem to contribute to teams' performance? We conducted focus groups with the top five teams (out of 18,201 participating teams) of a computer security Capture-the-Flag (CTF) competition, all consisting of high school students in the U.S. To analyze the qualitative data collected using a structured-interview format, we analyzed data using grounded theory to address the research questions.We find that the teams in our study adopted a set of strategies centered on specialties, which allowed them to reduce issues relating to dispersion, double work, and lack of previous collaboration. We also connect these results to prior frameworks to conceptualize crowd work, including how students organically adopted existing strategies for task decomposition, hierarchical/reputational organization, task assignment, and collaboration. This work has implications in both computer security education and distributed teamwork strategies developed among high school students pre-COVID19 pandemic, which is before remote work became more common among this age group. The teams in our study model how scaling and distributing security work among team members is feasible, productive and beneficial. In addition, we identify various areas for future work, such as issues of social identity in high-skilled crowd work environments.
Large language models (LLMs) are widely used in education, yet most initiatives use LLM-powered tools as black boxes. This demo introduces a browser-based tool that lets students use their own data to train tiny transformer models directly on their own devices, without any servers or installation required. By making training visible in real time, the tool enables hands-on experimentation with datasets, model sizes, pre-training, fine-tuning, and prompting, while ensuring accessibility and privacy. It reframes language models from tools to use into systems to build, probe, and critique-supporting AI literacy and critical reflection in K-12 education.
Many popular K-12 AI education tools focus on building reliable models with no-code platforms. These activities are useful, but they can create inflated perceptions of robustness and reliability. Our "Breakable Machine" takes the opposite approach. It is a multiplayer, digital-physical classroom game that rewards students for "spoofing" (impersonating) an image classifier through playful adversarial attacks. The students are tasked to find ways to trigger high-confidence misclassifications, view heatmaps of what the model attends to, and inspect the training data to find vulnerabilities. Errors and failures become resources for inquiry, opening space for conceptual change, critical discussion, and ethical reflection. The game positions AI as fragile, human-made technology with societal impacts. This demo invites discussion on how failure-focused activities and adversarial play can enrich AI literacy.
A recurring challenge in ethical and responsible computing (ERC) education is developing ways to understand and assess students' awareness and reasoning around ERC, which can provide insights to help educators align learning goals and content. We address this challenge by developing an assessment framework based on computer science (CS) students' responses to prompts about the benefits and harms of computing systems. We asked undergraduate CS students to read fictional news articles that described computing systems that are used in different contexts (i.e., health devices, automated content filtering, and automated weaponry) and prompted them to describe the pros and cons of the technologies on different communities. We qualitatively analyzed students' responses and produced a framework of themes organized under four key dimensions-(1) Technology (components and features of computing systems), (2) Communities (people impacted by the technology), (3) Impact (the kinds of impacts of the technology), and (4) Scope (the scale or reach of the technology's impact). We describe our development of this framework and how it can support the teaching of ERC and the design of tools for assessing students' awareness and reasoning about ERC.
As artificial intelligence (AI) becomes a critical part of future literacy, there is a growing need for educational platforms that reflect the cultural, linguistic, and infrastructural realities of learners in underrepresented regions. This research explores how African high school students can effectively learn AI through a culturally-infused, gamified, and collaborative platform. The study builds on insights from AfriML, a no-code, culturally-grounded machine learning (ML) platform that significantly enhanced learner engagement and comprehension by incorporating African accents, artifacts, and languages. While AfriML demonstrated the viability of contextualized AI learning, its limitations-such as its ML-only focus and lack of collaborative or privacy-preserving features-inspired the development of Afrude AI, a more expansive learning environment. Afrude AI serves as the research instrument for this study, integrating federated learning, gamification, and a dataset-sharing hub to extend AfriML's legacy. The research is organized into three sub-studies focusing on cultural integration, collaborative AI model training, and equitable dataset accessibility, offering a holistic approach to inclusive AI education.
Large Language Models (LLMs) are increasingly used in computing education, offering new opportunities to support programming practice, generate explanations, and scaffold student learning. However, current LLMs are not pedagogically aligned and often fail to adapt to students' prior knowledge, leading to explanations and feedback that are either too advanced or oversimplified. This research addresses the gap by developing human-centered LLMs tailored for programming education. Building on prior work in knowledge tracing and knowledge component discovery, we explore how student data (e.g., submissions, process data, misconceptions) can be leveraged to personalize LLM outputs. Our UKICER 2025 study demonstrated that automated knowledge component extraction with LLMs can approximate expert-annotated learning curves, highlighting the potential of scalable student modeling. The ongoing PhD project extends this by investigating prompt engineering, fine-tuning, and reinforcement learning with human or AI feedback (RLHF/RLAIF) to design pedagogically effective LLMs. The contributions include: (1) methods for tailoring LLMs to student profiles, (2) integration of automated KC pipelines into large-scale courses, and (3) evaluation of impacts on student learning, engagement, and experience. This work aims to advance scalable, personalized support in programming education through next-generation intelligent tutoring powered by LLMs.
There has been a relentless march towards research-focussed papers in Computing Education conferences, influenced by Valentine's pejorative "Marco Polo" characterisation of practice papers. We argue that there is an important place for practice-focussed papers when presented in a format which provides careful contextual description, as exemplified by the work of seventeenth century ecologist Maria Sibyella Merian. Computing Education Practice (CEP) is a UK practice-based conference that has its tenth year in 2026. In this paper we discuss and describe the CEP reporting format, and review the character and contribution of the CEP series. We explore a) the structure and content of CEP papers in relation to papers at comparable conferences (Koli, ITiCSE and ICER) using Simon's classification system; b) what kinds of institution are represented at CEP; and c) CEP's influence on participants' teaching, through a (mostly) qualitative survey. We find that a) CEP papers are similar in scope, theme and context to other computing education conferences despite the nature of papers showing an expected and distinctive preponderance of reports; b) CEP is more representative of the full range of UK computing departments than comparator conferences, when considering average student entry standards; and c) participants identify value in participation and discussions as well as some specific contributions of the publications. We champion the value of carefully constructed reports of practice in computing education.
Code-generating Artificial Intelligence (AI) is disrupting traditional Computer Science (CS) education with unclear implications for underrepresented students. While these tools could democratize CS education, they may also widen existing privilege gaps. Little research examines how Black students at Historically Black Colleges and Universities (HBCUs) view code-generating AI in relation to their identity and persistence in CS. We present a case study of Black HBCU students' perspectives on identity and persistence factors in relation to using code-generating AI, identifying implications for educational equity. We recruited 10 Black CS undergraduates from an HBCU to solve basic Python programming tasks with and without GitHub Copilot, using a pre-study questionnaire and a semi-structured exit interview to gather their perspectives. Participants emphasized three key identity aspects: expressing their humanity, striving for social impact, and navigating social categories. Success factors included gratification, social support, self-teaching, and financial aid, while barriers were deficits compared to others, and social barriers. Students expected AI-tools to help overcome deficits and social barriers and offer moral support, but anticipated drawbacks such as exposing deficits, disrupting identities, hindering learning, and promoting antisocial tendencies. Our findings underscore that personhood and solidarity are key to CS persistence for Black HBCU students. While AI can augment pedagogy and address deficits, it cannot replace human community. We recommend conservative deployment of AI-tools in CS education and caution against harms to learning, identity, and social life from constant use. Participants' undervaluing of premium AI poses a risk for technological privilege gaps to widen further.
Computing is a key skill needed in today's increasingly digital world. However, for novice learners outside of computer science programs, the experience of learning to program can elicit "a river of tears," as a graduate student provocatively reported in our recent interview study. Seventeen graduate students from the social sciences participated in semi-structured interviews focused on how they acquire programming skills, the barriers they face, and the resources they wish they could access. Through data-driven thematic analysis, we found common values underlying social science graduate students' desired learning supports including availability, transferability, and connection. Future learning strategies and supports to ameliorate social science graduate students' experiences could focus on resolving tensions between values, minimizing the wide disparity in resources among social science graduate students, and increasing the transferability support structures and the connections between social science graduate students, peers, and instructors. This research project contributes empirical understandings of social science graduate students' experiences learning to program for their independent research.
A high school programming standard in New Zealand requires students to write programs that "[use] conditions and control structures effectively" in order to receive the highest grade. This has caused issues for teachers, who are uncertain what this aspect of the standard entails, and researchers, who are attempting to develop an automated system to assist in the assessment of this standard. Furthermore, tertiary instructors are concerned with their students not only writing correct code that passes test cases, but ensuring their code is of a high quality. In this paper we analyse 2,871 introductory programming submissions, across both high school and university, and manually grade how "effective" they are based on a conservative interpretation of the standard and our own experience as instructors. We then explore five established code quality metrics, as well as our own approach, to search for which are most correlated with the manual grading. Finding no strong correlation, we explore if machine learning techniques can be used to improve the accuracy. We find these techniques lead to minor improvements that are probably not worth pursuing, and come to question whether assessors should be concerned about this aspect of student code.
Debugging is a necessity in programming, in both professional and educational contexts. For novices, however, debugging is often a significant challenge. Understanding what students actually do when debugging is key to addressing difficulties and developing targeted interventions. As a result, several studies investigated the students' debugging process. Nevertheless, it is unknown which aspects have been analyzed so far and which gaps still exist. To clarify the state of research, we conducted a scoping review and identified 36 papers that analyze students' debugging processes. Our review shows that the majority of the studies focused on selected parts of the process, mainly by analyzing screen recordings or videos from the classroom using qualitative, inductive methods. Moreover, most of the papers either assessed the students' debugging strategies or their performance. As a result, there is a lack of deductive analysis approaches focusing on investigating the whole debugging process. Consequently, this review provides a starting point for future analyses of debugging processes.
Data-related concepts and practices have been proposed to be a fundamental component of artificial intelligence (AI) school education. However, proposing concepts and practices is not enough. To enable teachers to introduce data concepts and practices in schools, it is necessary to understand the mechanisms that effectively support the learning of these under real school conditions. To this end, we designed and conducted a three-iteration design-based research study in collaboration with computer science and mathematics teachers, school students, and domain experts. In this paper, we present the results of the research process: the developed teaching approaches and the identified mechanisms that support learning of data concepts and practices following a conjecture-mapping approach. Based on the results, we explicate theoretically and empirically sound local instructional theories for teaching data concepts and practices in secondary school education on AI.
Argumentation can be considered one of the essential cognitive skills of the 21st century and should be promoted in higher education. It is especially essential in the field of computer science, where creating logical solutions is crucial. However, in most modern online learning environments, the learners' thinking process is invisible. To support the formation and assessment of the logical reasoning of learners in automated learning environments, more argumentative task types would need to be created. This research aims to study how argumentation can be built into automated online exercises.
Collaboration skills are essential for computer science (CS) students, yet many graduates enter the workforce lacking proficiency in collaboration skills. This research aims to identify missing collaboration skills in CS curricula, investigate the impact of prior programming experience on teamwork, and explore the use of software metrics to assess collaboration. Following the Educational Design Research approach, this thesis will lead to actionable recommendations for CS educators and supporting educational materials. The main expected contribution is a practical solution for integrating collaboration skill development into programming project courses which has been designed and prototyped over multiple iterations with testing and evaluation phases.
It is well-established that women are underrepresented in computer science (CS) education, but less is known about their representation at key university stages-application, admission, and retention through graduation. This study addresses that gap by investigating three research questions: (1) How do the distributions of women and men who apply to CS differ? (2) How do the distributions of women and men admitted to CS differ? (3) How do the distributions of women and men retained to graduation in CS differ? Using ten years of data from four U.S. institutions with varied demographics and characteristics, we apply a data-intensive approach combining descriptive statistics, visualizations, and regression modeling. Our findings reveal gender imbalances at the application stage, partial mitigation at admission, and similar retention rates but different outcomes for non-retained students by gender. We also find institutional differences, with more competitive schools showing larger gender gaps, while smaller private institutions show smaller disparities. Our results underscore the need to focus earlier in the pathways. Attracting more women to apply to college CS programs is essential for narrowing the gender gap, and high school exposure to CS alone might not be enough to spark sustained early interest.
Programming students can struggle with the different aspects of learning to program, for example with writing correct code, interpreting error messages, and gaining meaningful comprehension of assignments' underlying concepts. In the realm of program comprehension, Questions about Learners' Code (QLCs) assess to what extent programming students understand their programs by providing personalised questions focusing on program comprehension skills. We aim at novel QLC types targeting aspects such as error message comprehension and code quality. Alongside traditional QLCs, which target concepts such as tracing or structural aspects of code, we envision a programming environment which detects learner obstacles, such as compilation or runtime errors, and reacts by posing personalised questions which help the student reason about what went wrong. We expect that this approach will allow students to autonomously mitigate difficulties and develop program comprehension skills.
The PhD project uses an exploratory sequential mixed methods research design to answer broader research question "How blended learning is defined, implemented and experienced in project-based courses of undergraduate computing education, and what factors related to social and teaching presence support or hinder successful project completion?" This project uses Community of inquiry as core theoretical framework to understand blended learning experiences. Three research studies are conducted using qualitative and quantitative data collection and analysis techniques. Systematic literature review was conducted to explore definitions, implementation practices, and outcomes of blended learning in project-based courses, and results were used to conduct semi structured interviews from multiple stakeholders including 16 students, 3 teachers and 3 industry mentors who are involved in blended project-based courses in computing education in a university in Norway. And finally quantitative COI survey will be administered to analyze different cause and effect relationships among teaching support, social cohesion and cognitive resolution when students work on projects in blended learning environments. The preliminary findings suggest that, despite the increasing digitalization, the tangible benefits of in-person collaboration are challenging to replicate virtually. The findings go beyond merely identifying preferences for a mode of learning, they expand the discussion on how two learning modes serve different purposes in project-based courses.
