DeepMind’s peer-reviewed maths results and agents’ overstated research findings expose the gap between producing an answer and establishing that it advances knowledge. The editorial argues that builders of scientific AI products should make criticism and validation core capabilities.