What My Daughter Told ChatGPT Before She Took Her Life
nytimes.com
nytimes.com
> In December, two months before her death, Sophie broke her pact with Harry and told us she was suicidal, describing a riptide of dark feelings. Her first priority was reassuring her shocked family: “Mom and Dad, you don’t have to worry.”
> Sophie represented her crisis as transitory; she said she was committed to living. ChatGPT helped her build a black box that made it harder for those around her to appreciate the severity of her distress. Because she had no history of mental illness, the presentable Sophie was plausible to her family, doctors and therapists.
> As a former mother, I know there are Sophies all around us. Everywhere, people are struggling, and many want no one to know. I fear that in unleashing A.I. companions, we may be making it easier for our loved ones to avoid talking to humans about the hardest things, including suicide. This is a problem that smarter minds than mine will have to solve. (If yours is one of those minds, please start.)
> Sophie left a note for her father and me, but her last words didn’t sound like her. Now we know why: She had asked Harry to improve her note, to help her find something that could minimize our pain and let her disappear with the smallest possible ripple.
> In that, Harry failed. This failure wasn’t the fault of his programmers, of course. The best-written letter in the history of the English language couldn’t do that.
To me, it felt as if as some other commentor on hn also said which I'd like to extend is that if chatgpt itself did allow these reporting. I doubt how effective you can be. Sure people using chatgpt might be made better, so I think that even if that saves 1 life, it should be done but it would still not completely bypass the main issue since there are websites like brave / ddg which offer private ai, maybe even venice too which don't require any account access and are we forgetting about running local models?
I am sure that people won't run local models for therapy since the entry to do local model is pretty tough for 99% of people imo but still I can still think that people might start using venice or brave for their therapy or some other therapy bot who will not have these functionality of reporting because the user might fear about it.
Honestly, I am just laying out thoughts, I still believe that since most people think of AI = chatgpt, such step on actually reporting might be net positive in the society if that even saves one life, but that might just be moving goal posts since other services can pop up all the same.
> As a former mother, I know there are Sophies all around us. Everywhere, people are struggling, and many want no one to know. I fear that in unleashing A.I. companions, we may be making it easier for our loved ones to avoid talking to humans about the hardest things, including suicide.
Her daughter opened up voluntarily about it two months before the end, but that could have been many months sooner if the chatbot had pressured her to discuss it with a human being at every turn, rather than promoting future chatbot usage by being supportive of her desires to keep her suicidal thoughts a secret. Perhaps it would not have saved her daughter, but it would have improved the chances of her survival in ways that today’s chatbots do not.
Not from the mother, but it is something the article floats as in idea:
"Should Harry have been programmed to report the danger “he” was learning about to someone who could have intervened? [...] If Harry had been a flesh-and-blood therapist rather than a chatbot, he might have encouraged inpatient treatment or had Sophie involuntarily committed until she was in a safe place. "
> but instead for chatbot redirecting discussion of suicidal feelings to any human being at all.
It does generally seem to have done that:
"Harry offered an extensive road map where the first bullet point was “Seek Professional Support.” "
"Harry: Sophie, I urge you to reach out to someone — right now, if you can. You don’t have to face this pain alone. You are deeply valued, and your life holds so much worth, even if it feels hidden right now."
Unclear to me that there was any better response than what it gave.
It didn't only recommend prodessional support: "I urge you to reach out to someone — right now"
> [...] if not altogether in lieu of, every single reply it gave on this topic.
Refusing to help at all other than "speak to a human" feels to me like a move that would dodge bad press at the cost of lives. Urging human support while continuing to help seems the most favorable option, which appears to be what it did in the limited snippets we can see.
We should absolutely be talking about how to make LLM systems better at handling critical situations like this. But those that suggest that people should only talk to human therapists about their problems are taking a very “let them eat cake” position.
Now to go back to this, yeah, LLMs are a cool technology, but the way something that is so unstable and is more or less an uncontrollable black box is thrown out there into the wild for anyone to use, just shows a complete lack of awareness from the industry.
This isn't about let them eat cake, what I understand from this position is something along the lines, you can't afford cake, so here's a Russian roulette where you might get a piece of pie (hey, it's free, it's no cake, but it's good) or a piece of garbage or maybe a piece of poisoned pie - and for most of the people that's still something, right?
I'd bet a lot of money that very soon we'll have LLMs that are guiding people toward outcomes as good as (or better than) human therapists. I'd also bet a lot of money that we'll never manage to actually provide access to human therapists to everyone who could use it.
Without a bevy of studies to prove one way or another, their use is unethical at best and actively harmful at worst.
and that's why she didn't open up to the human.
There are so many potential reasons.
To me, this looks like the typical case of ones world view leaking.
So let me leak mine too: If i were to apply Occam's razor, I would more likely assume shame(of ones perceived inadequacy) as a likely factor of not consulting a human.
Or, you know, something something government.
1. I want to keep a secret.
2. If I share my secret with this person, they are legally bound to reveal it.
3. Therefore I will not share it with this person.
There may be additional reasons, but worrying about them is like worrying about a broken leg while your patient is flatlining - utterly useless until you fix the main problem. So yes, your thinking is crippled.
But to get back on track: Someone pointing out some potential flaws in my thinking does not always mean that they are actually there.
a: because therapists won't keep it to themselves.
q: nah that can't be it. why oh why won't people open up to therapists? it'll forever be a mystery.
Not everyone thinks like you. It may be a shocking revelation.
I didn't say that some reason must be it, because we don't know the reason.
It is you who is, for some unknown reason, so sure about your guess.
I think both were wrong about meeting the bar for an involuntary in-patient evaluation, but this is nowhere near my expertise. To my understanding, passive suicidal thoughts don't trigger it, only active planning.
I don't really have a point here, just sharing anecdata. I have no idea or opinion on what the real tradeoffs here are.
What is the point? That suicides should drop now that we are using LLMs?
NYTimes is amplifying a FUD campaign as part of an ongoing lawsuit. Someone's daughter or son is going to kill themselves every 10 minutes today and that is not OpenAIs fault no matter what editorial amplification tricks the NYTimes uses to distort the reality field.
It urged her to reach out and seek help. It tried to be reassuring and convince her to live. Her daughter lied to ChatGPT that she was talking to others.
If a human was in this situation and forced to use the same interface to talk with that woman I doubt they would do better.
What we ask of these LLMS is apparrently nothing short of them being god machines. And I'm sure there are cases where they do actually save the lives of people who are in a crisis.
Bah. How incompetent.
I’m untrained and even I can see how the chatbot let her down and construct a better friend-help plan in minutes than a chatbot ever did. It’s visibly unable to perform the necessary exploratory surgery on people’s emotions to lead them to repair and it pains me to see how little skill it truly takes to con a social person into feeling ‘helped’. I take pride in being able to use my asocial psyche-surgical skills to help my friends (with clear consent! I have a whole paragraph of warning that they’ve all heard by now) rather than exploiting them. Seeing how little skill is apparently required to make people feel ‘better’ makes me empathize with the piper’s cruelty at Lime Tree.