When AI Is Wrong on Purpose: How Students Respond to Buggy GenAI Code

Synthesis: As generative AI becomes central to software development, CS education is shifting toward prompt-centered workflows where students describe intended behavior in natural language to elicit code. But professional practice demands careful review of GenAI output that may look correct yet harbor subtle faults — a challenge in CS1, where current models solve tasks correctly and dull students' incentive to inspect generated code. Padurean et al. (2026) investigate how prompt-centered programming activities can be designed to deliberately surface buggy GenAI code, forcing students to practice verification rather than blind trust. The work speaks directly to the Over-Reliance problem and the Cognitive Offloading temptation of strong code models, and complements Combating Harms of Generative AI in CS1 with Code Review Interviews and a Flipped Classroom, which found oral code review preserves learning outcomes under rising AI use. By intentionally injecting faults, the approach reframes Generative AI in CS Education as a metacognitive training ground rather than an answer engine, reinforcing AI Literacy and Metacognition. Implications for Student Experience and Formative Assessment design suggest that 'wrong-on-purpose' perturbations are a scalable way to build debugging habits in an era where Reshaping Undergraduate Computer Science Education in the Generative AI Era is already underway.

What this means for practice

• Instructors. Inject validated, runnable near-miss bugs into otherwise correct generated code so that students have something to verify; correct output alone removed the incentive to inspect it.
• Instructors. Expect and name two different repair routes: injected bugs were more often repaired by direct code edits with higher next-attempt success, while natural prompt-related failures more often led students to refine the prompt by clarifying constraints, updating signatures, adding edge cases, or reframing the task.
• Designers. Keep both failure sources in one workflow, with hidden tests and an editor, so students alternate between specification refinement and debugging rather than practicing only one.
• Instructors. Ask students to reflect on what the repair demanded: their reflections described practice in code understanding, code review, and debugging along with a more careful verification mindset.
• Instructors. Pair bug injection with metacognitive prompts about trust, because students' stated gains centered on awareness of GenAI limitations rather than on scores alone.

Limitations

• This was an in-class deployment rather than a controlled comparison: differences observed between bug sources may reflect course context and platform choices, not only the bug source.
• The analysis covers 2,636 sessions from 917 students in one introductory C course with five platform problems, of which only two were required for credit, so replication across languages, institutions, and open-ended tasks is untested.
• Follow-up actions are shaped by the platform's prompt-and-test workflow with hidden tests and its particular feedback; the authors call for varying feedback, letting students write their own tests, and adding debugging support to see whether the same bug-source patterns hold.
• Bug source was not disclosed to students and no pre/post or delayed measure was collected, so how transparency affects trust, and whether these workflows produce measurable learning, remains unknown.

Citation

Victor-Alexandru Padurean, Kaitlin Riegel, Alkis Gotovos, Jyotika Mahapatra, Ahana Ghosh, Adish Singla (2026). When AI Is Wrong on Purpose: How Students Respond to Buggy GenAI Code. .