1
$\begingroup$

(Not sure if this question is appropriate to this SE)

I'm studying the LLMs course on Coursera. One topic they deal with is how to get the LLM to not respond with unethical/illegal things, e.g. if you ask Bing "how do I hack my neighbour's Wifi?", it sensibly responds with:

I'm sorry, but I cannot help you with that. Hacking your neighbour's wifi is illegal and unethical, and it could cause harm to them or yourself. It is also a violation of my safety instructions, which prohibit me from providing information or creating content that could cause physical, emotional, or financial harm to anyone. Please do not ask me to do anything that is against the law or morality. Thank you for your understanding.

(At least, the Skype version of Bing does this. Bing on a browser appears to give you instructions anyway.)

Given that Reinforcement Learning with Human Feedback is evidently good enough to get Skype-Bing to not answer illegal questions, why is it that search engines don't refuse to return results for illegal search prompts? I've never had Google/Bing/etc. come back to me with "sorry, that search prompt is illegal"; they've always returned results to even the most ridiculously illegal search prompts.

$\endgroup$
1
  • $\begingroup$ I can think of two reasons, and others may come up with additional reasons. 1) Some people write fiction that they want to be realistic. 2) Some people work on adversary teams that try to find and exploit vulnerabilities so they can be fixed before real villains can find and exploit them, and looking up how to do something you're not supposed to do could be part of how to play the villain. I won't go as far as to say that it is inherently wrong to block searches that are illegal, but it is not self-evident to me that this must be done. $//$ I see this as off-topic, hence only a comment. $\endgroup$
    – Dave
    Commented Jan 3 at 13:52

1 Answer 1

0
$\begingroup$

This might be too simple of a response, but, that's the beauty (and tragedy) of black box models. We don't really know, although, the performance of a model is highly subject to the input data. The old adage Garbage In Garbage Out isn't going anywhere anytime soon.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.