This study examines errors in a ChatGPT-mediated translation of a German economic text on inflation into Spanish, post-edited by 20 translation students. The analysis classifies 132 annotated instances by error origin (ChatGPT-generated versus student-introduced during post-editing) and by lin-guistic category. Results show that termi-nology is the highest-risk domain across the entire workflow (34.1%), followed by tense/aspect (15.2%) and style (13.6%). ChatGPT-related errors account for 50.8% of all instances, while student-introduced errors through over-editing represent 21.2%. A further 28.0% reflect acceptable alterna-tive reformulations. Students tend to trust fluent machine output even when it con-tains subtle semantic distortions, yet they also over-edit segments that are already ac-ceptable. The findings highlight three di-dactic priorities: developing LLM-based MT literacy, strengthening decision-making strategies in post-editing, and fostering gen-re- and domain-sensitive editing compe-tence. Implications for translator training and structured post-editing protocols are discussed.