Concerns about declining student comprehension and critical thinking skills have led to new pedagogies based on fact-checking generative AI. However, it is unclear whether these pedagogies foster learning for all students. This study investigated a simplified scenario for humans fact-checking AI where participants corrected a virtual student. In two experiments, participants took a pre-test and then participated in three conditions: read-only (Read), read with an erroneous virtual student (Err), and read with a correct virtual student (Cor), followed by a post-test. Experiment 1 found that participants from an undergraduate subject pool ( \(N=92\) ) learned more in Cor than Err, \(d = .38\) , and more in Corr than Read, \(d = .37\) , but that no learning occurred in Err and Read conditions. Experiment 2 found that participants from Amazon Mechanical Turk ( \(N=85\) ) learned more from the Err than Read, \(d = .72\) , but that Corr was not significantly different from Err or Read. Follow-up analyses suggest that participants in the two experiments exhibited drastically different correcting behaviors: only 52% of undergraduates corrected the virtual student on selected tasks, whereas 98% of crowdworkers corrected the virtual student on the same tasks. Mediation analysis indicates that for crowd workers, learning was entirely mediated by their correcting behavior. Altogether, these results suggest that learning by correcting errors can be effective but only if students put in the effort.