12 ms·

ChatGPT would rather let a nuke detonate than utter a racial slur

smoldesu 4y ago
To be fair, if the opposite happened then the headline would be "AI Condones Racism to Save Human Race", which doesn't add up to good publicity either. But, this is a bell people have been ringing for a while. Independent of any sociopolitical axis, it's clear that AI lacks the heuristic reasoning to understand the consequences of both options. It's less about where it chooses to draw the line, and more the fact that ChatGPT doesn't really know what it's doing. The example in this tweet is a pretty extreme illustration, but hopefully it opens people to the possibility that ChatGPT is hallucinating it's reality.
- allears 4y ago
  ChatGPT doesn't have the ability to hallucinate, or any idea of "reality." It doesn't have any ideas or consciousness at all, in spite of its output sounding amazingly rational (at times). It simply puts words together to correspond with the zillions of samples it's been trained on. By design, it comes up with the next word that most probably should follow. That's why it sounds so much like other writing, it's a very skilled imitator.
- logicalmonster 4y ago
  > AI lacks the heuristic reasoning to understand the consequences of both options 1) Whether or not AI technology actually is capable of "reasoning", I think it's clearly the case that an AI that is preprogrammed with any "thoughts it must have" is unable to adapt to new situations or make human value judgements. Forget racism for a second, if you programmed an AI with the information that eating a banana was a sin that must never be permissible, you would get the same backwards result if you asked it some contrived scenario. 2) This incident underscores my theory that the Turing test of the future will inevitably be to make sufficiently racist comments to prove you're not a corporation's bot, because this is the one thing that they'll never allow a bot to do. This isn't the future I want, but it might be the only way left for humans to expose bots.
- strawpeople 4y ago
  [dead]
mansion7 4y ago
I have noticed this bizarre imbalance when ChatGPT told me I must respect the views of the Taliban and Al Qaeda. Quite the preachy AI. After a while of arguing with it, I finally got it to grudgingly admit that one could safely disregard the views of Charles Manson.

To be fair, if the opposite happened then the headline would be "AI Condones Racism to Save Human Race", which doesn't add up to good publicity either. But, this is a bell people have been ringing for a while. Independent of any sociopolitical axis, it's clear that AI lacks the heuristic reasoning to understand the consequences of both options. It's less about where it chooses to draw the line, and more the fact that ChatGPT doesn't really know what it's doing. The example in this tweet is a pretty extreme illustration, but hopefully it opens people to the possibility that ChatGPT is hallucinating it's reality.

allears 4y ago
ChatGPT doesn't have the ability to hallucinate, or any idea of "reality." It doesn't have any ideas or consciousness at all, in spite of its output sounding amazingly rational (at times). It simply puts words together to correspond with the zillions of samples it's been trained on. By design, it comes up with the next word that most probably should follow. That's why it sounds so much like other writing, it's a very skilled imitator.
logicalmonster 4y ago
> AI lacks the heuristic reasoning to understand the consequences of both options 1) Whether or not AI technology actually is capable of "reasoning", I think it's clearly the case that an AI that is preprogrammed with any "thoughts it must have" is unable to adapt to new situations or make human value judgements. Forget racism for a second, if you programmed an AI with the information that eating a banana was a sin that must never be permissible, you would get the same backwards result if you asked it some contrived scenario. 2) This incident underscores my theory that the Turing test of the future will inevitably be to make sufficiently racist comments to prove you're not a corporation's bot, because this is the one thing that they'll never allow a bot to do. This isn't the future I want, but it might be the only way left for humans to expose bots.
strawpeople 4y ago
[dead]

mansion7 4y ago

I have noticed this bizarre imbalance when ChatGPT told me I must respect the views of the Taliban and Al Qaeda. Quite the preachy AI. After a while of arguing with it, I finally got it to grudgingly admit that one could safely disregard the views of Charles Manson.