What began as a practical exercise in a structural bioinformatics course at the University of Graz led to a Nature Correspondence: we identified systematic stereochemical errors in AlphaFold 3, the AI system developed by Google DeepMind. The finding, published shortly after Demis Hassabis and John Jumper received the Nobel Prize in Chemistry for this very technology, demonstrates that AI-predicted protein structures still require careful human validation — and that the classroom can be a space for genuine scientific discovery at the highest level.