Retrieval-Augmented Generation is designed to make AI systems more reliable by grounding them in external evidence. But what happens when the retrieved evidence is relevant to the input yet historically wrong for the context?
This lightning talk explores anachronistic retrieval, a failure mode in multimodal RAG where a vision-language model correctly recognizes a historical artifact but grounds its interpretation in evidence from the wrong time period or cultural context, producing a plausible but historically incorrect answer.
Using historical paintings and visual archives, including Japanese collections, as a testbed, I will compare standard VLMs, generic RAG, temporal RAG, and temporal-provenance grounding to examine when retrieval introduces rather than reduces historical errors. I’ll also show how temporal metadata and source provenance can be incorporated into retrieval pipelines, alongside an evaluation framework for detecting contextually incompatible evidence.
The broader lesson extends beyond historical AI: semantic relevance alone is not enough for reliable RAG. Retrieved evidence must also be valid for the context in which it is used.