The chatbot in that article did have safeguards in place to stop it from listing suicide methods, but he just kept asking until he found a jailbreak. He was looking for someone to give him the answer he wanted all along
Past a certain point is there anything that will stop someone that determined to find something to validate their own self-destructive viewpoint? If not AI then a site or faction or person on Telegram with a pro-suicide opinion could have the same effect, and it would still be just as tragic but no longer be noteworthy. It didn't sound like he had a happy life to begin with with the AI ruining it completely, like it sounds
I had an LLM tell me to that killing myself makes sense when recounting a low point in my life to it and how I'd contemplated suicide. I wasn't looking for any kind of "jailbreak" in prompts, this was following a genius on HN saying it's a "great alternative to a therapist". No, it isn't fit for purpose.
> It didn't sound like he had a happy life to begin with
Which are exactly the kind of people who are going to be using LLMs as a "therapist". See the problem?
Yes, which is the point: if you substitute a human therapist for an LLM, this is what's going to happen because "guard rails" can't account for every scenario. If half a billion dollars doesn't buy a bug-free game, why would it buy a safe LLM?
Does suicide encouraged by a therapist happen with human therapists too? Probably, but likely much less common as a suicide is going to hurt your reputation.
Past a certain point is there anything that will stop someone that determined to find something to validate their own self-destructive viewpoint? If not AI then a site or faction or person on Telegram with a pro-suicide opinion could have the same effect, and it would still be just as tragic but no longer be noteworthy. It didn't sound like he had a happy life to begin with with the AI ruining it completely, like it sounds