Errors in AI grading and feedback are by their nature non-deterministic and difficult to completely avoid. Since inaccurate feedback potentially harms learning, there is a need for designs and workflows that mitigate these harms. To better understand the mechanisms by which erroneous AI feedback impacts students’ learning, we conducted surveys and interviews that recorded students’ interactions with a short-answer AI autograder for ``Explain in Plain English'' code reading problems. Using causal modeling, we inferred the learning impacts of wrong answers marked as right (false positives, FPs) and right answers marked as wrong (false negatives, FNs). We further explored explanations for the learning impacts, including errors influencing participants’ engagement with feedback and assessments of their answers’ correctness, and participants’ prior performance in the class.

FPs harmed learning in large part due to participants’ failures to detect the errors. This was due to participants not paying attention to the feedback after being marked as right, and an apparent bias against admitting one’s answer was wrong once marked right. On the other hand, FNs harmed learning only for survey participants, suggesting that interviewees’ greater behavioral and cognitive engagement protected them from learning harms. Based on these findings, we propose ways to help learners detect FPs and encourage deeper reflection on FNs to mitigate learning harms of AI errors.