Researchers use old design-validation questions as instructions for AI story generation, making hallucinations auditable

A new system for cultural heritage storytelling uses knowledge graphs (structured databases of facts and relationships) paired with AI language models, forcing the AI to retrieve evidence before writing stories instead of inventing details. This means heritage institutions can now point to exactly which facts support each sentence in a generated narrative, rather than accepting or rejecting the entire output as a black box.

Museums and cultural institutions have known for years that language models can sound convincing while inventing facts — a catastrophic problem when the stakes are historical truth. This paper shows a method that makes the AI's reasoning transparent and falsifiable: every narrative element traces back to a specific data source. The trade-off is clear: symbolic approaches (pure fact retrieval) are accurate but choppy; adding natural language context makes stories flow better but reintroduces some risk; graph-based methods split the difference. What becomes possible is cultural heritage applications where you can actually defend every claim in the output.