Not saying that this is what happened, but this shows the possibility for the LLM's owner to secretly spread misinformation. A bit of "finetuning" on made up fa

Not saying that this is what happened, but this shows the possibility for the LLM's owner to secretly spread misinformation. A bit of "finetuning" on made up facts about a specific person or event, and the LLM will happily regurgitate them. The owner can just say "oops, software malfunction, not sure what happened here". We can't prove bad intentions, because reverse-engineering the training data from the model output is impossible.