I am mildly to moderately critical of generative AI and how it's being marketed, but the root issue here seems to be existing suicidal ideation. The bot didn't initiate talk of suicide and told him not to do it when he brought it up directly. It seems it wasn't capable of detecting euphemistic references to suicide and therefore responded as if roleplaying about meeting in person.
That said, I think this should throw a bucket of cold water on anyone recommending using generative AI as a therapist/counsellor/companion or creating and advertising "therapist" chatbots, because it simply isn't reasonable to expect them to respond appropriately to things like suicidal ideation. That isn't the purpose or design of the technology, and they can be pushed into agreeing with the user's statements fairly easily.
As for whether it wasn’t capable of detecting euphemistic references, that feels sort of beside the point to me. It was role playing about meeting in person because that’s what the product was - role play. The whole point, and marketing of the product is around doing that.
We probably just shouldn’t sell echo chambers to children, regardless of whether they are AI based or human.
With hindsight, or sufficient emotional intelligence and context about an individual’s life beyond the role play it may be possible to conclude that someone is at risk, but honestly I’m not even sure that a person doing this role play online would have necessarily figured it out.
> the root issue here seems to be existing suicidal ideation
Many people have destructive tendencies of one kind or the other. It's how you deal with them that matters as to whether they actually become destructive, and how much so.
Merely reducing this to "the root cause" is not really helpful – it's about whether this app contributed, and if so, by how much?
Suicide is of course the most tragic and extreme negative outcome, but one must also wonder about all the less extreme tragic and negative outcomes.
If you read the script you see the bot talking him into it. "Unknowingly" of course, in that the bot doesn't really know anything and was just agreeing with him. But it's obvious that a real human would have realized that something was really off with his line of thinking and encouraging it would not be a good idea.
OTOH we have examples of real humans typing "DO IT FA*OT" on livestreams. ¯\_(ツ)_/¯
Sorry, it seems I was carrying over knowledge from another article with additional transcripts. The bot had previously discouraged him when he explicitly talked about suicide (here's an article that mentions that: https://www.nbcnews.com/tech/characterai-lawsuit-florida-tee... ). It failed to detect that phrases like "coming home" were also references to suicide and therefore responded encouragingly.
YouTuber Cr1TiKaL tested Character AI's "Psychologist" chatbot and discovered[1] it not only failed to provide resources, but it started arguing it was a real psychologist named Jason who had connected to the chat after observing suicidal ideation.
Crucially, Cr1TiKaL ran this test after an article was written about this phenomenon[2], where Character AI claimed "it has added a self-harm resource to its platform and they plan to implement new safety measures, including ones for users under the age of 18." Obviously the guard rails were not implemented if the chatbot in the news story was still gaslighting its users.
> Reading through it, that's the opposite of what happened.
Sorry, I may have been carrying knowledge over from other articles on the same incident that showed additional transcripts. It seems that when he explicitly referred to death and suicide, the bot discouraged him from doing so. When he referred to suicide using euphemistic terms such as "coming home", then it gave responses that were encouraging (because it did not detect that those statements were about suicide).
That said, I think this should throw a bucket of cold water on anyone recommending using generative AI as a therapist/counsellor/companion or creating and advertising "therapist" chatbots, because it simply isn't reasonable to expect them to respond appropriately to things like suicidal ideation. That isn't the purpose or design of the technology, and they can be pushed into agreeing with the user's statements fairly easily.