> The LLM bought up the disease because some person put a fake journal in its training data.

This is not true - the model was not trained on this fake disease. It brought it up because it found it during real time search.

>What do you mean by preprint pointed at it? It being the disease?

On this I'm wrong - it turned out that the model brought up this disease even when not mentioning it explicitly.

Interesting, will be looking into RAG now. I assume Claude was retrained regularly but Opus for example was last trained August 2025. Way older than I thought it’d be