Timely and specific feedback is one of the strongest drivers of student learning, yet in large electrical engineering courses the student‑to‑TA ratio is so high that a learner who gets stuck may wait days for an explanation. Generative AI (GenAI) can scale conversational feedback, but using it for graded work raises trust and accountability concerns; keeping a human in the loop to verify AI judgments re‑introduces the very delay that diminishes feedback value. This creates a tension between immediacy and reliability.
We designed and piloted a GenAI practice platform in a large maths class. The platform delivers immediate, scaffolded feedback during self‑directed practice, while human oversight is shifted from real‑time grading to upfront verification of model solutions. Student interactions were captured through logs, surveys, and interviews to understand how they used the tool, how they judged its feedback’s value and reliability, and what lessons could transfer to other engineering subjects.
Results show that most students could quickly locate errors and iterate after receiving instant feedback, leading to noticeable gains in learning efficiency. Concerns about feedback trust dropped significantly when the system displayed solution provenance tags and human audit reports. Nevertheless, students still desired direct instructor input at critical junctures to guard against misleading AI hints. Overall, moving human validation to a pre‑deployment stage while letting AI provide real‑time guidance preserved feedback speed without sacrificing credibility.
These findings suggest that generative AI can serve as an on‑demand tutor for practice activities in large‑class settings, provided the system architecture clearly separates the timing and transparency of human review. By doing so, feedback latency can be dramatically reduced while maintaining instructional quality.
Review