That unsubstantiated supposition is doing a lot of heavy lifting and that’s a dangerous and unproductive way to frame the argument.
I’ll make a purposefully exaggerated example. Say a school wants to add cyanide to every meal and defends the decision with “supposing it helps students concentrate and be quieter in the classroom, why not?”. See the problem? The supposition is wrong and the suggestion is dangerous, but by framing it as “supposing” with a made up positive outcome, we make it sound non-threatening and reasonable.
Or for a more realistic example, “suppose drinking bleach could cure COVID-19”.
First understand if the idea has the potential to do the thing, only then (with considerably more context) consider if it’s worth implementing.
The demonstrable harms include assisting suicide, there's is no way to ethically continue the measurement because continuing the measurements in their current form will with certainty result in further deaths.
And working to set a threshold for what we would consider acceptable? No thanks
Individuals and companies with mind boggling levels of investment want to push this tech into every corner of our lives and and the public are the lab rats.
Unreasonable. Unacceptable.
Because a human, esp. a confused and depressive human being is a complex thing. Much more complex than a stable, healthy human.
Words encouraging a healthy person can break a depressed person further. Statistically positive words can deepen wounds, and push people more to the edge.
Dark corners of human nature is twisted, hard to navigate and full of distortions. Simple words don't and can't help.
Humans are not machines, brains are not mathematical formulae. We're not deterministic. We need to leave this fantasy behind.
Also it's side-stepping the question, isn't it? "Supposing that the advice it provides does more good than harm" already supposes that LLMs navigate this somehow. Maybe because they are so great, maybe by accident, maybe because just having someone nonjudgmental to talk to has a net-positive effect. The question posed is really "if LLMs lead some people to suicide but saved a greater number of people from suicide, and we verify this hypothesis with studies, would there still be an argument against LLMs talking to suicidal people"
A human can also say the wrong things to push someone in a certain direction. A psychologist, or anyone else for that matter, can't stop someone from committing suicide if they've already made up their mind about it. They're educated on human psychology, but they're not miracle workers. The best they can do is raise a flag if they suspect self-harm, but then again, so could a machine.
As you say, humans are complex. But I agree with GP: whether the words are generated by a machine or coming from a human, there is no way to blame the source for any specific outcome. There are probably many other cases where the machine has helped someone with personal issues, yet we'll never hear about it. I'm not saying we should rely on these tools as if we would on a human, but the technology can be used for good or bad.
If anything, I would place blame on the person who decides to blindly follow anything the machine generates in the first place. AI companies are partly responsible for promoting these tools as something more than statistical models, but ultimately the decision to treat them as reliable sources of information is on the user. I would say that as long as the person has an understanding of what these tools are, interacting with them can be healthy and helpful.
>AI companies are partly responsible for promoting these tools as something more than statistical models,[...]
This might be exactly the issue. Just today I've read people complaining that newest ChatGPT can't solve letter counting riddles. Companies just don't speak loud enough about LLM-based-AI shortcomings that result from their architecture and are bound to happen.
>AI companies are partly responsible for promoting these tools as something more than statistical models,[...]
This might be exactly the issue. Just today I've read people complaining that newest ChatGPT can't solve letter counting riddles. Companies just don't speak loud enough about LLM based AI shortcomings that result from their architecture and are bound to happen.
Depending on where you live, this may well result in the vulnerable person being placed under professional supervision that actively prevents them from dying.
That's a fair bit more valuable than when you describe it as raising a flag.
ChatGPT essentially encouraged a kid not to take a cry-for-help step that might have saved their lives. This is not a question of a bad psychologist; it's a question of a sociopathic one that may randomly encourage harm.
A child thinking about suicide is clearly a sign that there are far greater problems in their life than taking advice from a machine. Let's address those first instead of demonizing technology.
To be clear: I'm not removing blame from any AI company. They're complicit in the ways they market these tools and how they make them accessible. But before we vilify them for being responsible for deaths, we should consider that there are deeper societal problems that should be addressed first.
This is more "sometimes it will seemingly actively encourage them to kill themselves and it's basically a roll of the dice what words come out at any one time".
If a counsellor does that they can be prosecuted and jailed for it, no matter how many other patients they help.
I know people who earn above average income and still spend a significant (north of 20%) portion of their income on therapy/meds. And many don't, because mental health isn't that important to them. Or rather - they're not aware of how much helpful it can be to attend therapy. Or they just can't afford the luxury (that I claim it is) of private mental health treatment.
ADHD diagnosis took 2.5y from start to getting meds, in Norway.
Many kids grow up before their wait time in queue for pediatric psychologist is over.
It's not ChatGPT vs shrink. It's ChatGPT vs nothing or your uncle who tells you depression and ADHD are made up and you kids these days have it all too easy.
Sertraline can increase suicidal thoughts in teens. Should anti-depressants not be allowed near suicidal/depressed teens?
Well certainly not without careful monitoring and medical advice, no of course not!
By the "common sense" definitions, LLMs have "intelligence" and "understanding", that's why they get used so much.
Not that this makes the "common sense" definitions useful for all questions. One of the worse things about LLMs, in my opinion, is that they're mostly a pile of "common sense".
Now this part:
> Add in the commercial incentives of 'Open'AI to promote usage for anything and everything and you have a toxic mix.
I agree with you on…
…with the exception of one single word: It's quite cliquish to put scare quotes around the "Open" part on a discussion about them publishing research.
More so given that people started doing this in response to them saying "let's be cautious, we don't know what the risks are yet and we can't un-publish model weights" with GPT-2, and oh look, here it is being dangerous.
Yes, they did claim that they wouldn't release GPT-2 due to unforeseen risks, but...
a. they did end up releasing it,
b. they explicitly stated that they wouldn't release GPT-3[1] for marketing/financial reasons, and
c. it being dangerous didn't stop them from offering the service for a profit.
I think the quotes around "open" are well deserved.
[1] Edit: it was GPT-4, not GPT-3.
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 honest they were being — I mean, look at the whole of Big Tech right now — but the story they gave was self-consistent:
[Paraphrased!] (a) "We do research" (they do), "This research costs a lot of money" (it does), and (b) "As software devs, we all know what 'agile' is and how that keeps product aligned with stakeholder interest." (they do) "And the world is our stakeholder, so we need to release updates for the world to give us feedback." (???)
That last bit may be wishful thinking, I don't want to give the false impression that I think they can do no wrong (I've been let down by such optimism a few other times), but it is my impression of what they were claiming.
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.
But how do you tell before it matters?
Bleach is the least of your problems.
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?
"He didn't need the money. He wasn't sure he didn't need the gold." (an Isaac Asimov short story)
> More suicidal people than not, would cease to be suicidal if ...
I'm going to need to see a citation on this one.
It appear to be the only way
The reality is most systems are designed to cover asses more than meet needs, because systems get abused a lot - by many different definitions, including being used as scapegoats by bad actors.
But there is zero actually effective way to do that as an online platform. And plenty of ways that would cause more harm (statistically).
My comment was more ‘how the hell would you know in a way anyone could actually do anything reasonable, anyway?’.
People spam ‘Reddit cares’ as a harassment technique, claiming people are suicidal all the time. How much should the LLM try to guess? If they use all ‘depressed’ words? What does that even mean?
What happens if someone reports a user is suicidal, and we don’t do anything? Are we now on the hook if they succeed - or fail and sue us?
Do we just make a button that says ‘I’m intending to self harm’ that locks them out of the system?
What else do you propose?
I will take this seriously when you propose a test that can distinguish between that and something with actual "intelligence or understanding"
When AI gets there (and I’m confident it will, though not confident LLMs will), I think that’s convincing evidence of intelligence and creativity.
Damn I thought we'd got over that stochastic parrot nonsense finally...
In retrospect, from experience, I'd take the LLM.
This is basic common sense.
Add in the commercial incentives of 'Open'AI to promote usage for anything and everything and you have a toxic mix.