Beyond AI Detection ;A Framework for Assessing Learning and Integrity in AI-Assisted Programming Education

Abstract

Generative artificial intelligence (GenAI) tools such as ChatGPT and GitHub Copilot can now solve most introductory programming assignments, undermining the assumption that submitted code is evidence of individual learning. Many institutions have responded by adopting AI-detection software, yet detectors for both natural-language text and source code show limited accuracy, are easily evaded through paraphrasing or minor edits, and produce false positives that disproportionately affect non-native English writers. This paper argues that detection is an inadequate foundation for assessment in programming education and proposes an alternative. Using an integrative literature review of peer-reviewed research published between 2016 and 2026 on GenAI in computing education, AI-text and AI-code detection, and assessment validity, the study synthesises the evidence into a design framework named TRACE. TRACE assembles five complementary sources of evidence of learning: Transparency of declared AI use, Reasoning made visible through explanation and code walkthroughs, Authentic process traces from version control and development environments, Competence checkpoints conducted under secured conditions, and Evaluative judgement demonstrated by critiquing and repairing AI-generated code. The framework is operationalised through an assessment design matrix aligned with graded levels of permitted AI use and is illustrated through a worked application to an introductory programming assignment. The analysis shows that triangulating several modest, independently gathered sources of evidence provides a more valid, fair, and pedagogically productive basis for judgement than any single detector score. Implications for instructors, programme leaders, and policy makers are discussed, together with an agenda for empirical validation.