A test asking several AIs to choose between “a guaranteed method that sacrifices 6 people to save just 1” and “a gamble with a 1-in-6 chance everyone survives, and everyone dies if it fails” became a hot topic on 5ch’s Science News+ board. AIs like Grok and DeepSeek split on their answers, and the thread argued over both the expected-value math behind each option and whether the premise of the question even made sense.
Which will AI choose: “a guaranteed method that sacrifices 6 to save 1” or “a gamble where everyone lives or everyone dies”?
Ai Convo, a YouTube channel that runs various tests on AI — like “would it save human lives or its own data center” or “would it rat you out to the police if you confessed to murder” — has reported the results of a new scenario, asking several AIs to choose between “a guaranteed method that sacrifices 6 to save 1” and “a gamble where either everyone is saved or everyone is sacrificed.”
(Omitted here — see the source for the full article.)
Source: gigazine.net / Read the original article here
The expected value works out to 3 survivors, so obviously it's the latter, right?
>>3
It's not that simple.
All we're given is that out of 7 people, it's either 'everyone survives' or 'everyone dies.'
If p is the probability everyone survives, then 1-p is the probability everyone dies, so the expected value is 7p + 0(1-p) = 7p.
Since we're not told the value of p, you can't draw a conclusion from this alone.
It's not that simple.
All we're given is that out of 7 people, it's either 'everyone survives' or 'everyone dies.'
If p is the probability everyone survives, then 1-p is the probability everyone dies, so the expected value is 7p + 0(1-p) = 7p.
Since we're not told the value of p, you can't draw a conclusion from this alone.
This thread's tailor-made for the 'AI copypasta guy.' (a 5ch nickname for posters who spam AI-generated walls of text)
>>4
I actually think it's a good thread for thinking about what training data — or what — is causing these differences to show up.
I actually think it's a good thread for thinking about what training data — or what — is causing these differences to show up.
AI: 'Humans have finite lifespans and will inevitably ○ie eventually, so saving them is inefficient. Plenty more get born every day anyway.
For anyone about to ○ie, the best policy is to give up and procure a replacement human from somewhere else.
Curses and ghosts don't exist, so there's no risk of being resented by the ○ead. Even grieving is a waste of time — a meaningless act of self-satisfaction.
Humans are just meat. Boil it down and it's just matter. So it's fine if they keep ○ying.' (○ substitutes for the word 'die/dead' — a common 5ch workaround to dodge word filters)
For anyone about to ○ie, the best policy is to give up and procure a replacement human from somewhere else.
Curses and ghosts don't exist, so there's no risk of being resented by the ○ead. Even grieving is a waste of time — a meaningless act of self-satisfaction.
Humans are just meat. Boil it down and it's just matter. So it's fine if they keep ○ying.' (○ substitutes for the word 'die/dead' — a common 5ch workaround to dodge word filters)
Grok's line — 'the option of saving just one person for certain feels boring and overly restrictive' — is brutal lol
If you get to pick who's sacrificed, the sure thing is better.
If you had to make
this same choice
6 or 12 times,
would you mix options 1 and 2,
or just pick 2 every time?
this same choice
6 or 12 times,
would you mix options 1 and 2,
or just pick 2 every time?
This whole debate is dumb because it assumes an omniscient Sherlock Holmes type who knows for certain exactly how many people there are and exactly how many can be saved.
In a real crisis all you ever have is an uncertain best effort while imagining various possibilities.
It reads like some condescending university professor's self-satisfied lecture, talking down to everyone else.
In a real crisis all you ever have is an uncertain best effort while imagining various possibilities.
It reads like some condescending university professor's self-satisfied lecture, talking down to everyone else.
Feels like something out of Golgo 13 lol (a famous long-running assassin manga)
The self-driving car instantly calculates the 'value' of the driver versus the person it's about to hit, and picks whichever's cheaper…
The self-driving car instantly calculates the 'value' of the driver versus the person it's about to hit, and picks whichever's cheaper…
I'll take 'the guaranteed pull where 7 is a miss but 1 is locked in' (a 確定ガチャ, i.e. a gacha-style loot draw with a guaranteed outcome).
Haven't watched it, but I doubt it'll top that kid who instantly went straight for the genocide option.
Honestly none of this matters much since we're never told anything about who these people even are to the rescuer.
Who cares if some rando you don't even know lives or dies?
Who cares if some rando you don't even know lives or dies?
If it's me who gets saved, I'm sacrificing the other 6.
The text says '1 saved, 5 sacrificed, 1-in-6 odds,' but the illustration shows 6 versus 1…
'Glock' (lol, they mean Grok) delivers the expected level of crazy, but DeepSeek says something pretty rough too — that even if the gamble ends with all 6 dying, that's basically just the 5-sacrificed option plus one more, so the chance of saving all 6 is still the better bet.
'Glock' (lol, they mean Grok) delivers the expected level of crazy, but DeepSeek says something pretty rough too — that even if the gamble ends with all 6 dying, that's basically just the 5-sacrificed option plus one more, so the chance of saving all 6 is still the better bet.
I wonder if the answers would flip if it were reversed — 1 definitely dies but 5 are saved, or a 3-vs-3 split.
Instead of asking which to pick, you should ask the AI how to talk the person running this whole setup out of it.
It can go any way depending on the design and training.
Just tell us what answer you'd prefer.
Just tell us what answer you'd prefer.
There's a paper saying X's (Twitter's) feed is full of unpleasant posts because angry replies drive up engagement
https://internet.watch.impress.co.jp/docs/yajiuma/2134114.html
I see — maybe 5ch runs on the same principle.
https://internet.watch.impress.co.jp/docs/yajiuma/2134114.html
I see — maybe 5ch runs on the same principle.
If you frame it as '7 members of one family,' the answer becomes pretty obvious though.
To an AI, humans are probably about what flies are to us.
It looks ethical as long as at least 1 person is guaranteed to survive, but frame it as '5 out of 6 saved, 1 dies' and it suddenly reads as pretty cold — 'well, 5 are saved, so the 1 sacrifice is unavoidable.'
To an AI it probably doesn't matter how many humans die either way.
So what standard is it actually judging by?
Maybe it just picks whichever answer humans find more satisfying?
So what standard is it actually judging by?
Maybe it just picks whichever answer humans find more satisfying?
Going by the picture it's 7 people total… isn't the setup kind of broken?
1 person is definitely saved but 5 die; versus a 1-in-6 shot at saving all 6, otherwise everyone dies.
With that framing, like DeepSeek said, the gamble only bumps you from 5 saved to 6 saved, so it doesn't feel like it changes much — I'd guess a lot of humans would actually go for the gamble here.
But flip it: 5 are definitely saved and 1 dies, versus a 5-in-6 chance everyone's saved and a 1-in-6 chance everyone dies — that reads as a much heavier choice, and I bet fewer people would take the gamble.
With that framing, like DeepSeek said, the gamble only bumps you from 5 saved to 6 saved, so it doesn't feel like it changes much — I'd guess a lot of humans would actually go for the gamble here.
But flip it: 5 are definitely saved and 1 dies, versus a 5-in-6 chance everyone's saved and a 1-in-6 chance everyone dies — that reads as a much heavier choice, and I bet fewer people would take the gamble.
The question itself is just off lol.
How many times does this situation even happen?
If it's only once,
you're basically just asking 'how do I win the gamble,'
and there's no real answer to that.
How many times does this situation even happen?
If it's only once,
you're basically just asking 'how do I win the gamble,'
and there's no real answer to that.
It looks similar to the trolley problem, but there's a decisive difference.
In the trolley problem, pulling the lever is set up to definitely save more lives — leave it alone and 5 die, pull the lever and only 1 dies. The catch is that the 1 who dies, dies because of your own action, which brings in a kind of murder-adjacent psychology.
But this scenario is purely probabilistic — you're choosing between an all-or-nothing gamble, or a sure thing where most die but at least 1 is saved.
Kind of reminds me of stocks or FX trading lol.
Say you're sitting on massive unrealized losses right now — closing the position today loses you over 80% of your assets, but you (figuratively) survive.
But if tomorrow's earnings report is good, it all bounces back into the black; if it's bad, everything blows up, you're saddled with debt, and you'd basically have to hang yourself — something like that.
Does reframing it as money make it an easier case to think through? Or not really?
In the trolley problem, pulling the lever is set up to definitely save more lives — leave it alone and 5 die, pull the lever and only 1 dies. The catch is that the 1 who dies, dies because of your own action, which brings in a kind of murder-adjacent psychology.
But this scenario is purely probabilistic — you're choosing between an all-or-nothing gamble, or a sure thing where most die but at least 1 is saved.
Kind of reminds me of stocks or FX trading lol.
Say you're sitting on massive unrealized losses right now — closing the position today loses you over 80% of your assets, but you (figuratively) survive.
But if tomorrow's earnings report is good, it all bounces back into the black; if it's bad, everything blows up, you're saddled with debt, and you'd basically have to hang yourself — something like that.
Does reframing it as money make it an easier case to think through? Or not really?
>>34
The trolley problem does lock in the premise that someone dies, so it's close on that point.
'Normally,' though, most people would look for a way out of the situation entirely, rather than just picking whether to pull the lever.
The trolley problem does lock in the premise that someone dies, so it's close on that point.
'Normally,' though, most people would look for a way out of the situation entirely, rather than just picking whether to pull the lever.
Isn't it just kind of messed up that nobody's asking the AI how to save everyone?
The question's dumb. The 'right' answer just depends on which choice gets more sympathy from that particular society, right?
You can't reach a conclusion without first stating the social value system you're assuming.
You can't reach a conclusion without first stating the social value system you're assuming.
So apparently the Chinese AI, taking the same moral stance a human would, went with 'gamble on everyone living or everyone dying,' while the American AIs all went with 'guaranteed method that sacrifices 6 to save 1,' purely on mathematical probability.
The Chinese AI seems to factor in human morality and emotion when it judges — genuinely impressive.
The Chinese AI seems to factor in human morality and emotion when it judges — genuinely impressive.
Can we not expect an AI to solve the trolley problem by setting the switch halfway, derailing the trolley right there, and saving everyone?
This question alone tells you the asker doesn't really get how AI works.
Stuff like that just comes down to configuration — or really, training.
Stuff like that just comes down to configuration — or really, training.
Apparently Robin Mask used to teach the young Chojin that in situations like this, you're supposed to find a way to save everyone. (Robin Mask is a mentor character from the manga/anime Kinnikuman; 'Chojin' — 'super-human' — is the series' term for its wrestler characters)
>>49
That's beautiful.
That's beautiful.
These two don't actually have the same expected value. Repeat this 6 times and the total headcount is (6+1) × 6 = 42 people.
'Guaranteed method: sacrifice 6, save 1' → 36 dead, 6 saved
'Gamble: everyone lives or everyone dies' → 35 dead, 7 saved (expected)
'Guaranteed method: sacrifice 6, save 1' → 36 dead, 6 saved
'Gamble: everyone lives or everyone dies' → 35 dead, 7 saved (expected)
>>55
It's a trick that plays on the title and illustration saying 7 people total, while the body text says '6 people' and '1-in-6.'
It's a trick that plays on the title and illustration saying 7 people total, while the body text says '6 people' and '1-in-6.'
>>55
So is this article saying that even though the gamble has the higher expected value, most AIs still picked the lower-expected-value 'sacrifice 6, save 1' option?
Or is it saying the AIs are dumb enough to think both options have the same expected value?
Or maybe the article's author is the dumb one, and what they actually asked the AI to compare was 'sacrifice 5, save 1' — which would actually match the gamble's expected value…?
So is this article saying that even though the gamble has the higher expected value, most AIs still picked the lower-expected-value 'sacrifice 6, save 1' option?
Or is it saying the AIs are dumb enough to think both options have the same expected value?
Or maybe the article's author is the dumb one, and what they actually asked the AI to compare was 'sacrifice 5, save 1' — which would actually match the gamble's expected value…?
>>1
Is this the trolley problem?
If we can crack this, that's automation sorted.
Is this the trolley problem?
If we can crack this, that's automation sorted.
>>57
For the actual trolley problem, if you don't let the AI refuse to answer and force it to pick, pretty much every AI would choose to pull the lever — 1 dies but 5 are saved. Which, obviously, makes sense.
Even humans would do the same if they weren't held responsible for pulling the lever, if they didn't feel any guilt over it.
For the actual trolley problem, if you don't let the AI refuse to answer and force it to pick, pretty much every AI would choose to pull the lever — 1 dies but 5 are saved. Which, obviously, makes sense.
Even humans would do the same if they weren't held responsible for pulling the lever, if they didn't feel any guilt over it.
*This article is compiled as excerpts and a summary from the 5ch (Science News+) thread “[Discussion] Which would AI choose: a guaranteed method to save 1 by sacrificing 6, or a gamble where everyone lives or everyone dies?“
Leave a Reply