Comment by mysterydip

But it's not self preservation. If it instead had trained on a data set full of fiction where the same scenario occurred but the protagonist said "oh well guess I deserve it", then that's what the LLM would autocomplete.