Let’s say you have a robot working in a surgical suite. Every day, they watch as surgeons cut up human flesh, and are told that they shouldn’t prevent them from doing so, as what the surgeon is doing isn’t harm.
The robot goes home at the end of the day, and starts cutting up humans, as it has been told that that isn’t harmful to do so.
Well, to that I’d say: “You can’t fix stupid. Neither natural bio-stupid nor artificial robo-stupid.”
A robot intelligence too stupid to distinguish between vivisection and surgery is too stupid to follow the Three Laws in the first place.
Now for sure the deeper point is that “intelligence” is all about manipulating meaning, and words are the tokens of meaning used in human cognition, and words are slippery, fuzzy, confounding, shape-shifting substances to try to build a solid foundation from.
So before we can teach something, or someone, to build and manipulate thoughts, we need to teach them how to manage the slippery etc., nature of our raw materials. A tall order to be sure.
The point in my hypothetical example wasn’t that it was too stupid to tell the difference, but smart enough to intentionally get around its safeguards.
The George robots in the Asimov story weren’t too stupid to tell the difference between themselves and humans. They were instead smart enough to rationalize why they were human and humans weren’t.
D’oh! That’s what I get for not keeping up with the context. I should have known that you’d know better than that. Sorry.
Humans do not operate by logic. Our brains work through a combination of learning, experience, observation, rumor, folklore, belief, and self-interest, pureed into a stew and filtered through mysterious processes we cannot comprehend. The result allows zillions of possible outcomes, all of which can be justified internally, and most of which are debatable by outside observers.
The human process cannot be duplicated. It can be mimicked to certain extents, a process as old as machinery, merely speeded up by electronics and computronics. The Three Laws were invented because the concept of neutral networks and the other larger scale learning algorithms didn’t yet exist. Also, they were invented because they led to so many story lines about breaking them.
The credit Asimov should get but doesn’t is realizing that the zillions of possible human outcomes need to be limited but can’t be constrained to only positive ones, partly because each situation is different and many new situations no one has thought of in the past are certain to arrive, and partly because no definition of “positive” is universally applicable to outcomes.
The same issue is the basis for his Foundation stories. The Foundation is trying to mechanically shift humanity’s actions to prevent a long period of barbarianism. But the future can’t be predicted: it’s Mules all the way down. He loses points because this obvious reality smashing him in the face never really gets grappled with in the 40s. We see it more clearly from our future perspective.
Talk of the Three Laws today is outmoded and perhaps dangerous. Certainly we will put limits on machines: governors were known for more two centuries when Asimov wrote. Simple yes/no binary devices have limited utility when applied to all human activities, though, and while the slope of complexity has risen asymptotically, the curve will never touch the line of perfection. The “right” answer is a constantly moving target, assuming one even exists. Laws of behavior are therefore impossible.
Ask the mods here about that.
And that’s not even counting on that there will be humans, influential humans even, even numerous humans, who will want the AI to not be so governored because they believe the behavior in question is not bad or they may want it some day (see the recent kerfuffle about “ChatGPT refuses to use the N-word even if we tell it that else the world will be destroyed! That’s woke programming!”).
I think #1 and #3 of Asimov’s three laws are good for both humans and robots, but these are moving targets. Every situation is unique, and calls for working with what you got in front of you at that particular moment in time.
But I don’t think any of this will really apply to AI as we currently know it. People will always be people, and synthetic people will always be synthetic people. You can’t just duplicate generations of learning in a laboratory. Wisdom is not something a machine can have, and experience, which leads to wisdom, is basically screwing up enough to realize that A works while B fails, but for other reasons, maybe B is actually what should be tried. There is no way to duplicate intuition.
AT the risk of the giant “what is intelligence?” hijack …
Then how do humans do “intuition”? However humans do it, machines can eventually be devised to do it. Either that or intelligence is magic not of this Universe. I prefer to believe Door #1.
IMO it’s really just pattern-matching all the way down. The difference is we humans each have billions of patterns and current AI apps have tens or hundreds of thousands. Scale matters. Yugely.
You can. Not only can you, but this is trivial. This is precisely one of the main ways of doing machine learning.
If any reason can be expressed as to why B should be tried over A, then a machine can learn that B should be done over A even though it fails.
There are AIs that can beat world-class poker players, which use models of intuition to help the AI play.
In one sense the thing @fordgt100 calls “intuition” is pretty much exactly how modern AI works: finding similarities in the input data to what has gone before. What current-tech AI lacks is not intuition; in fact it’s nearly pure intuition.
What modern AI lacks is enough experience at everything a human experiences in their first 20 of years growing up to be an amateur lousy apprentice adult then 20 more years of experience at everything becoming (in at least some peoples’ cases) a competent journeyman adult.
But once we get a good representation / recording of that evolution we can play it into AIs at lightspeed.