According to the researchers, “the triggers are not contextual so humans ignore them when instructed to solve the problem”—but AIs do not.
Not all humans, unfortunately: https://en.wikipedia.org/wiki/Age_of_the_captain
This comes up frequently in a variety of discussions most notably execution speed and security. Developers will frequently reason upon things to which they have no evidence, no expertise, and no prior practice and come up with invented bullshit that doesn't even remotely apply. This should be expected, because there is not standard qualification to become a software developer, and most developers cannot measure things or follow a discussion containing 3 or more unresolved variables.
Just like some humans may be conditioned by education to assume that all questions posed in school are answerable, RLHF might focus on "happy path" questions where thinking leads to a useful answer that gets rewarded, and the AI might learn to attempt to provide such an answer no matter what.
What is the relationship between the system prompt and the prompting used during RLHF? Does RLHF use many kinds of prompts, so that the system is more adaptable? Or is the system prompt fixed before RLHF begins and then used in all RLHF fine-tuning, so that RLHF has a more limited scope and is potentially more efficient?