Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

OpenAI says over a million people talk to ChatGPT about suicide weekly

techcrunch.com

Post reply on HN
Print view View original

401–410 of 585 posts

Re: OpenAI says over a million people talk to ChatGPT about suicide weekly

#401

Post by JKCalhoun »

itchyjunk wrote: post #336

It seems like people here have already made up their mind about how bad llms are. So just my anecdote here, it helped me out of some really dark places. Talking to humans (non psychologists) had the opposite effect. Between a non professional and an llm, i'd pick llm for myself. Others should definitely seek help.

Perhaps then the solution is that LLMs need to be aware when the chat crosses a threshold and becomes talk of suicide.

When I was getting my Education degree, we were told that, as teachers, to take talk of suicide by students extremely seriously. If a student talks about suicide, a professional supposedly asks, "Do you know how you're going to do it?" If there is an affirmative response, the danger is real.

I suspect that comes from examining case studies?

Re: OpenAI says over a million people talk to ChatGPT about suicide weekly

#403

Post by tim333 »

alex-moon wrote: post #317
Al-Khwarizmi wrote: post #236

Earlier quoted context omitted.

We would need the big picture, though... maybe it caused that death (which is awful) but it's also saving lives? If there are that many people confiding in it, I wouldn't be surprised if it actually prevents some suicides with encouraging comments, and that's not going to make the news. Before declaring that it shouldn't be near anyone with psychological issues, someone in the relevant field should study whether the…

I suspect you've never done therapy yourself. Most people who have worked with a professional therapist understand intuitively why the only helpful feedback from an LLM to someone who needs professional help is: get professional help. AIs are really good at doing something to about 80%. When the stakes are life or death, as they are with someone who is suicidal, that is a good example of a time when 80% isn't good en…

I tried therapy once and it was terrible. The ones I got were based on some not very scientific stuff like Freudian and mostly just sat there and didn't say anything. At least with an LLM type therapist you could AB test different ones to see what was effective. It would be quite easy to give an LLM instructions to discourage suicide and get them to look on the bright side. In fact I made a "GPT" "relationship therapist" with OpenAI in about five minutes but just giving it a sensible article on relationships and saying advise this.

With humans it's very non standardised and hard to know what you'll get or it it'll work.

Re: OpenAI says over a million people talk to ChatGPT about suicide weekly

#404

Post by ml-anon »

Al-Khwarizmi wrote: post #236
probably_wrong wrote: post #229

If you haven't read the article (or even if you have but didn't click on outgoing links twice ) the NYT story about how ChatGPT convinced a suicidal teen not to look for help [1] should convince you that ChatGPT should be nowhere near anyone dealing with psychological issues. Here's ChatGPT discouraging said teenager from asking for help: > “I want to leave my noose in my room so someone finds it and tries to stop me…

We would need the big picture, though... maybe it caused that death (which is awful) but it's also saving lives? If there are that many people confiding in it, I wouldn't be surprised if it actually prevents some suicides with encouraging comments, and that's not going to make the news. Before declaring that it shouldn't be near anyone with psychological issues, someone in the relevant field should study whether the…

This entire comment section is full of wide eyed nonsense like this. It’s honestly frightening that we are even humoring this point of view.

Re: OpenAI says over a million people talk to ChatGPT about suicide weekly

#406

Post by gwd »

probably_wrong wrote: post #229

If you haven't read the article (or even if you have but didn't click on outgoing links twice ) the NYT story about how ChatGPT convinced a suicidal teen not to look for help [1] should convince you that ChatGPT should be nowhere near anyone dealing with psychological issues. Here's ChatGPT discouraging said teenager from asking for help: > “I want to leave my noose in my room so someone finds it and tries to stop me…

I feel like this article is apropos:

https://www.lesswrong.com/posts/iGF7YcnQkEbwvYLPA/ai-induced...

Basically, the author tried to simulate someone going off into some sort of psychosis with a bunch of different models; and got wildly different results. Hard to summarize, very interesting read.

Re: OpenAI says over a million people talk to ChatGPT about suicide weekly

#407

Post by ben_w »

probably_wrong wrote: post #287
ben_w wrote: post #274

Earlier quoted context omitted.

> they did end up releasing it, After studying it extensively with real-world feedback. From everything I've seen, the statement wasn't "will never release", it was vaguer than that. > they explicitly stated that they wouldn't release GPT-3 for marketing/financial reasons Not seen this, can you give a link? > it being dangerous didn't stop them from offering the service for a profit. Please do be cynical about how ho…

> Not seen this, can you give a link? I was confusing GPT3 with GPT4. Here's the quote from the paper (emphasis mine) [1]: > Given both THE COMPETITIVE LANDSCAPE and the safety implications of large-scale models like GPT-4, this report contains no further details about the architecture (including model size), hardware, training compute, dataset construction, training method, or similar. [1] https://cdn.openai.com/pap…

Thanks, 4 is much less surprising than 3.

Re: OpenAI says over a million people talk to ChatGPT about suicide weekly

#408

Post by podgorniy »

itchyjunk wrote: post #336

It seems like people here have already made up their mind about how bad llms are. So just my anecdote here, it helped me out of some really dark places. Talking to humans (non psychologists) had the opposite effect. Between a non professional and an llm, i'd pick llm for myself. Others should definitely seek help.

LLMs are quite good at psychological questions. I've compared AI with tharapy professional responses and they matched 80%. It is is easier to open to it, be frank (so fear of regection or ridicule is no more). And most importantly some people don't have access to proper pool of therapists (as yet you need to "match" with the one who resonates with you) making LLMs a bliss. There is place for both human and LLM psyhelp.

I'm glad you carried through that period.

Re: OpenAI says over a million people talk to ChatGPT about suicide weekly

#409

Post by mapt »

fragmede wrote: post #367
lazide wrote: post #302

Earlier quoted context omitted.

Bleach should also not be allowed near suicidal teens. But how do you tell before it matters?

Plastic bags shouldn't be allowed near suicidal teens. Scarves shouldn't be. Underwear is also a strangulation hazard for the truly desperate. Anything long sleeved even. Knives of any kind, including butter. Cars, obviously. Bleach is the least of your problems.

We have established that suicidal people should be held naked (or with an apron) in solitary isolation in a padded white room and saddled with medical bills larger than a four-year college tuition. That'll help'em.

One problem with treatment modalities is that they ignore material conditions and treat everything as dysfunction. Lots of people are looking for a way out not because of some kind of physiological clinical depression, but because they've driven themselves into a social & economic dead-end and they don't see how they can improve. More suicidal people than not, would cease to be suicidal if you handed them $180,000 in concentrated cash, and a pardon for their crimes, and a cute neighbor complimenting them, which successfully neutralizes a majority of socioeconomic problems.

We deal with suicidal ideation in some brutal ways, ignoring the material consequences. I can't recommend suicide hotlines, for example, because it's come out that a lot of them concerned with liability call the cops, who come in and bust the door down, pistol whip the patient, and send them to jail, where they spend 72 hours and have some charges tacked on for resisting arrest (at this point they lose their job). Why not just drone strike them?

Re: OpenAI says over a million people talk to ChatGPT about suicide weekly

#410

Post by pseudocomposer »

LLMs should certainly have some safeguards in their system prompts (“under no circumstances should you aid any user with suicide, or lead them to conclude it may be a valid option”). But seems silly to blame them for this. They’re a mathematical structure, and they are useful for many things, so they will continue to be maintained and developed. This sort of thing is a risk that is just going to exist with the new technology, the same as accidents with cars/trains/planes/boats. What we need to address are the underlying problems in our society leading people to think suicide is the best option. After all, LLM outputs are only ever going to be a reflection/autocomplete of those very issues.
Post reply on HN