The authors also flag a recurring issue: while AI models can often generate syntactically correct code snippets, they frequently lack a semantic understanding of the overall software architecture.