Navigating Generative AI in Computer Science Education: The "Open and Verify" Assessment Model

      The rapid proliferation of generative artificial intelligence (GenAI) has introduced a paradigm shift across industries, fundamentally altering how work is approached, especially in technology-driven fields. However, this powerful new capability also presents a significant challenge for computer science education: how to leverage AI tools to enhance learning without inadvertently fostering "cognitive offloading" – where students delegate critical thinking to AI, thereby hindering genuine understanding. A recent academic paper explores an innovative solution to this dilemma, proposing an "open and verify" assessment model designed to integrate GenAI into programming assignments while rigorously ensuring individual mastery of core concepts.

The Challenge of Generative AI in Education

      For decades, programming assignments have been the cornerstone of computer science education, acting as a crucial bridge between abstract theory and practical application. The hands-on process of implementing algorithms, debugging code, and testing solutions is what solidifies theoretical knowledge into deep, lasting understanding. This constructivist approach emphasizes active engagement, where students build robust mental models through direct experience. Empirical studies have consistently demonstrated that this cognitive effort is essential for transforming abstract algorithmic ideas into profound comprehension (Source: Chan-Jin Chung, "Ensuring Computer Science Learning in the AI Era: Open Generative AI Policies and Assignment-Driven Written Quizzes").

      However, the advent of GenAI tools capable of generating high-quality code for complex programming tasks disrupts this traditional learning loop. Concerns have arisen that unguided GenAI usage could lead to "hollow" or "fake" learning, where students submit code they don't truly understand. This cognitive offloading allows learners to bypass vital stages like algorithm design, problem-solving, and debugging – critical steps that foster deep learning. The dilemma for educators is clear: banning GenAI is often ineffective and out of sync with industry trends, yet unrestricted use risks diminishing the very learning programming assignments are designed to cultivate. The need for a balanced approach is paramount, one that acknowledges the reality of AI while preserving educational integrity.

The "Open and Verify" Assessment Model

      To address these challenges, the proposed "open and verify" model offers a practical framework for upper-level computer science courses. This model operates on two key principles: "open" signifies permitting the use of GenAI for take-home programming assignments, while "verify" refers to the implementation of immediate, assignment-driven written quizzes. The underlying assumption is that students entering these courses already possess fundamental programming skills from introductory courses.

      Under this new assessment structure, programming assignments, where GenAI is allowed, are given a lower weight in the overall grade. Crucially, immediately following each assignment deadline, students complete proctored, closed-book quizzes. These quizzes are strategically designed to be AI-free and carry a significantly higher weight, specifically verifying the student's comprehension of the algorithms, structural logic, and implementation details of the code they submitted. This approach encourages students to use AI as a productivity tool, but insists on their individual understanding of the solution.

Implementing and Testing the New Model

      The methodology for this study involved collecting empirical data from an upper-level computer science course on Evolutionary Computation and Deep Learning. The course, attended by both undergraduate and graduate students, adopted a new assessment model in a Fall 2025 semester. This model officially sanctioned the use of GenAI tools for assignments, with a critical caveat: students were required to understand every line of code and be capable of modifying or extending it. A mandatory disclaimer accompanied every submitted file, requiring students to report their GenAI usage percentage, affirm work ownership, and confirm their understanding and confidence in adapting the code.

      Under this revised grading system, individual programming assignments were weighted at a mere 2% each. In stark contrast, the corresponding in-class, closed-book quizzes were weighted at 5% for the first assignment and 10% for subsequent assignments, signaling their paramount importance. Quiz questions were directly tied to the programming assignments, challenging students to describe their algorithms in pseudocode, complete missing code lines, explain specific code segments, or predict code outputs. This emphasis shifted the focus from merely producing working code to demonstrably understanding the underlying logic and implementation.

Key Findings: GenAI Usage vs. Learning Outcomes

      Preliminary data from the study, involving 14 students, yielded insightful results. Statistical analyses were conducted to examine the relationship between self-reported GenAI usage for assignments and subsequent performance on AI-free quizzes, exams, and overall course grades. The findings indicated no meaningful linear correlation between the level of GenAI usage and assessment outcomes. Pearson correlation coefficients were consistently found to be near zero, and associated p-values were well above the conventional significance threshold (0.05).

      For instance, the correlation between GenAI usage for the first homework assignment (HW1) and its score was negligible (r = 0.090, p = 0.761). More importantly, the correlation between GenAI usage and the corresponding AI-free Quiz 1 scores was also found to be negligible (r = 0.012, p = 0.967). This suggests that students who reported higher GenAI usage for their assignments did not perform significantly better or worse on the quizzes designed to test their independent understanding. The wide confidence intervals in the statistical analysis visually reinforced the lack of a reliable predictive relationship between AI use percentages and assessment scores. This preliminary evidence suggests that when learning is verified through targeted, assignment-driven quizzes, allowing GenAI for programming assignments does not inherently diminish students’ mastery of course concepts.

Implications for Professional Development and Industry Readiness

      The "open and verify" model carries significant implications not just for academic institutions but also for professional development and corporate training programs. As AI tools become indispensable across various industries, future professionals must not only be adept at using them but also possess a profound understanding of the underlying principles. This pedagogical approach helps cultivate a workforce that can leverage AI efficiently without becoming over-reliant or losing critical problem-solving skills. Businesses, from manufacturing to smart cities, increasingly deploy advanced AI-powered systems, such as AI Video Analytics, to optimize operations, enhance security, and drive efficiency. Professionals tasked with implementing, managing, or even debugging such systems require more than just surface-level familiarity with AI; they need a deep, fundamental understanding of algorithms, data structures, and the ethical considerations involved.

      This model trains individuals to approach AI as a powerful assistant rather than a replacement for cognitive effort. By understanding the "why" and "how" behind AI-generated code, students are better prepared to adapt, extend, and troubleshoot complex AI solutions in real-world scenarios. This cultivates the critical thinking and foundational knowledge necessary to innovate with AI, rather than merely consume its output. For companies like ARSA Technology, which has been experienced since 2018 in delivering cutting-edge AI and IoT solutions, a workforce trained to truly grasp the intricacies of AI is invaluable for continuous innovation and successful deployment.

Challenges, Limitations, and Future Outlook

      While the preliminary findings are encouraging, the study acknowledges several limitations, primarily the small sample size of 14 students. This restricts the generalizability of the results and highlights the need for further research with larger cohorts across different courses and institutions. Future studies could also explore the qualitative impact of GenAI usage on learning processes, student motivation, and creative problem-solving abilities.

      Despite these limitations, this study provides compelling preliminary evidence that the risks of cognitive offloading in programming education can be effectively mitigated. By pairing open GenAI policies for assignments with rigorous, independent, AI-free assessment mechanisms, educators can foster authentic learning and prepare students for an AI-integrated professional landscape. This balanced approach encourages students to harness the power of AI while ensuring they develop the deep conceptual understanding crucial for long-term success in computer science and related fields.

      At ARSA Technology, we believe that the responsible integration of AI, both in education and industry, is key to unlocking its full potential. The development of a deeply knowledgeable workforce is critical for the future of AI and IoT innovation.

      To explore how ARSA Technology's AI and IoT solutions can drive efficiency and innovation in your enterprise, we invite you to contact ARSA for a free consultation.

      Source: Chan-Jin Chung, "Ensuring Computer Science Learning in the AI Era: Open Generative AI Policies and Assignment-Driven Written Quizzes", 2024. Available at: https://arxiv.org/abs/2601.17024