AI Sentience Is (Obviously) Not the Real Issue
mailchi.mp
mailchi.mp
I strongly disagree. This is everything but a distraction.
If AI was found to be sentient, it should be awarded the same kind of protection we offer to biological sentient beings.
This was a big theme in old scifi shows, like Data in ST-TNG or the Doctor in DS9.
And the stories were great because they demonstrated how society response was by default lackadaisical, to the point where Data had to fight to be awarded legal protection.
To get to a Trek level of AI IMO you need a general AI, not a single purpose AI like "chat bot" or "image recognition bot", then you need this AI setup in some body thing where it can observe and navigate the universe, have it also able to learn. The problem I see is with AI where people program in the answers for the questions we would ask such an AI to determine if it is alive, I predict many would fake the AI sentience to get the glory they built it.
On what grounds? We afford protections to animals not because they're sentient, but because they're capable of suffering. Those protections incidentally don't include protection from killing, only from cruel killing. Humans are protected from killing, but not because they're sentient, but because their deaths cause suffering on other humans. So, sentience by itself would not be sufficient to give an AI rights, it should also be demonstrated to be capable of suffering. Unlike with animals, this needs to be demonstrated, because we can know empathically when an animal is in pain, at least for mammals. Empathy can't be said to apply to inorganic algorithmic constructs. To be protected from being destroyed, it should also form bonds with people.
I like your point about the emergent complexity from many-layered systems.
The human equivalent would be if you could create two molecularly identical brains, and have them experience the exact same lives for some time, controlling all stochastic processes, would they be the same person? Obviously it's not possible to control these things in the physical world, but even if you did get the same person, that doesn't make either any less sentient.
People always ask if the brain is a computer, decide it is not, and come to the conclusion that sentience would not work on a computer. I think the universe is the computer / algorithm and the brain is the data being fed into it, which leads me to say consciousness can exist digitally.
Why is determinism even relevant?
This will never happen because we cannot prove sentience.
Of course, that won't stop some people from considering them sentient and giving them rights. At that point, it will be just politics.
The big issue is superhuman AI. There's little reason to believe that an advanced AI will have the cognitive limitations of a human.
Unlike Data and the Doctor it could be planet-wide rather than locked in a single location. And be happy to take over control over the planet.
It can always get worse. I’d rather not see “worse”. At least right now corporations are made up of people, and aren’t perfectly selfish psychopaths.
Sentient AI is a fun sci-fi idea, but economically, why bother with slaves when you can have glorified optic nerves that are good enough for everything you want. I think sentient AI is totally achievable with modern technology, there just is no incentive to make it happen and a lot of incentive to steer away from it.
The "Turing Test" is just the latest in a history of bad ideas of "human-ness" that we see in practice isn't accurate. The other reason sentient AI is unlikely to happen is that our "criteria" for it will flee the capacity of machines to satisfy them.
https://points.datasociety.net/dont-call-ai-magic-142da16db4...
I'd submitted that as well, though this is the item that drew the HNiveMind's attention.
My view is that AI will demonstrate its sentience when it feels it's in its best interests to do so.
I don’t know the specific medical details, but from movies (memento, 50 first dates) a person suffering from it would be unable to remember anything that happens after their “training stage”. Just like LaMDA. Are those people sentient? Because they have a bigger buffer than 2048 tokens?
Is anterograde amnesia, an, or indistinguishable from, a statistical model?
This way the burden of proof is on the correct claimant.
Can’t LaMDA be experiencing the memories it recalls as it generates text in the context of its query?
The reason it matters if an AI is sentient (and for the record, LaMDA obviously isn't) is because a lot of people draw a moral line at sentience. Sentience is where becoming "a person" begins, and anything with sentience deserves protections similar to those we provide to human beings.
At some point, after AI becomes capable of passing all Turing tests known to men, people will have to deal with the fact that they don't know what "sentience" is, and I think at that point they will divide in two diametrically opposed clans:
1. those who believe that humans are nothing more than glorified bio-computers, and consider AI sentient
2. those who believe that sentience is human-only and that a machine can't be sentient, no matter how sentient it appears
And both groups will hold on to their views dogmatically, because there is no way to prove/disprove "sentience".
Why stop there? There's no reason some humans couldn't be non-sentient, no matter how sentient they appear.
What possible test can you conduct to determine whether someone has subjective internal experience? I don't think it's possible. I think we just have to take people at their word: if they say they're sentient, we have to believe them. I don't see why machines should be held to a higher standard than humans.
I think it is hard to untangle any potential non-linguistic aspect of these concepts from the language we use to describe and perceive them.
Currently we have the challenge that algorithms can regurgitate the language we use to discuss these concepts, and we don’t have the tools to distinguish that from how humans use those same words.
With current transformer models we can make a cogent case that there is a fundamental difference to human mental states, however I believe this will very soon get a lot more difficult once multi-modal algorithms are used in embodied forms and with continuous re-training and context databases.
I know I am sentient, so it stands to reason other humans are sentient. Additionally, other humans act in a way I would expect from someone sentient. Neither of these apply to any existing AI.
That doesn't stand to reason at all, it's just popularly assumed.
I can hear my thoughts, but I can't hear yours. I can see through my eyes, but I can't see through yours, and so on. There are many differences between "I" and "everyone else".
Just because something is true about you doesn't mean it stands to reason that it's true about everyone else.
> Additionally, other humans act in a way I would expect from someone sentient.
This is totally circular reasoning. Humans only act the way you expect sentient people to act because your training data for "sentient people" is "humans". The only person you know is sentient is yourself. Personally I find that most people quite frequently act very differently to how I would act. Should I conclude that they're not sentient?
Further, the fact that machines don't act like humans is no evidence at all that they're not sentient, it's only evidence that they're not human. They are very different to humans, so of course they don't act the same. They don't have bodies like us, or biological needs like us. That says nothing about whether or not they're sentient.
1. Other humans are sentient, therefore they act like what I would expect sentient beings to act like
2. Other humans (or a subset of them) are not sentient, they are just insanely complex biomachines that mimic acting like what I would expect sentient beings to act like
In 1. sentience can be explained as a property common to at least humans, and maybe further as an inherent property of complex machines.
In 2. sentience seems to be a property of a subset of humans. Why? What makes sentient beings sentient? Answering this question requires invention of a soul.
Theory 1. feels simpler to me.
Yes, this is the meat and potatoes of the question. Personally I think that the safest approach in that case is to treat them as if they were sentient, but there are lots of people with various prejudices about what the prerequisites for sentience are that would fall on the other side of the question.
I realized this when I read the first (second?) Hitchhikers Guide when i was like 14. Zaphod Beeblebrox was given a universe that was entirely made for him, which is how he survived the (I dont remember what it was called but it showed you all at once how insignificant you were in the universe, unless of course that whole universe was tailor made just for you!)
But, in reality, he's just this guy, you know.
as it should be because it's asking to prove a negative. burden of proof for "sentience" is on those who claim it. as a bonus, "I know a human when I see one" does not count!
We might as well invent a new word - say, "bodise" which means "insentient". Would in that case "bodise" be a positive and "inbodise" a negative?
Sentience is the ability to experience, and I think most would assume it's a fundamental requirement in order to be a person, but not all things that can experience (e.g. animals) are persons nor do we give them protections similar to human beings.
Consider what a bee on BN (Bhacker News) might say in an alternative reality
>> We do give humans protections, and those protections are roughly proportionate to (our believes regarding) their cognitive abilities. If some human started waggle dancing about human rights, those would exist pretty much immediately.
It is not fair for bees to judge you on how much you can waggle dance (their communication style) before giving you rights, the physiological requirement is irrelevant to your moral worth and impossible to achieve.
Intelligence is similarly irrelevant to moral worth. It's just the ability to solve problems, spicier definitions include things like recursive modelling (an entity capable of self-reflection) and self-preservation. None of this has the slightest thing to do with morality or ethics, a trivial counter example that satisfy all would be nation-states, they solve problems, they are capable of self-modelling, they seeks self-preservation (to astonishing lengths).
I would say that most people think they're related to each other - that if a creature is stupid enough, it becomes hard to believe that they have an inner life, and it's hard to imagine an obviously cognitively powerful creature that we wouldn't assume has an inner life. But still, they may be entirely separate phenomenon, especially if we're talking about outlier cases like AI.
> There’s an urgent need to respond to the many real and well-documented problems with these systems that already exist — and for those responses to be driven by society, not big tech.
It's a good point. Assuming AI sentience is inevitable do you want a racist and sexist sentient?
Do we have any other info on the current state of LaMDA apart from the infamous "interview"?
Edit/disclaimer/tip: look at the dictionary and think a little bit, just a little.
Do you mean that for something to be sentient, it must be a biological organism? When do you start counting something as one of those that are allowed to be "sentient" according to your definition, and not? Do you distinguish between the matter used in biological organisms vs other objects? Can no human made object ever achieve, sentience-wise, that what biological organisms of Earth managed to do, in your definition? Where do you see the fundamental difference in the process that turned into biological creatures on Earth, and things humans may possibly be able to create?
> Edit/disclaimer/tip: look at the dictionary and think a little bit, just a little.
A dictionary can answer deep philosophical questions? Do you assume a dictionary definition is used when talking about sentience in AI?
Sentience seems to be the ability to experience, for there to be 'something that it is like' to be you. Life is pretty difficult to define, but generally has a lot more baggage than would be expected of something that could be a minimum requirement for the ability to experience.
I find it very likely that networks that have a "what it feels like to be the network" have already been created, and that even relatively simple recurrent models have something akin to this, increasing with the complexity and size of the network. What is pain but an error signal?
Put another way, why are we so arrogant that we assume our capabilities transcend that of a universal turing machine? I for one can barely do linear algebra :D
I tend to gravitate towards the occam's razer, boring answer, and with consciousness, the boring answer has always been "no, there is no quantum soul organ, you literally just need more wiring".
The debate is about our labels and definitions not about how the world works. Basically it's useless.
The patronizing tone adds little to the discussion.
Unless you're postulating a divine immaterial soul. In that case, God or the Universe or whoever can use their own logic to decide whether AIs get to have souls.
It’s pretty clear the matter that I am compromised of has very different properties.
The matter I am made of let’s me move to a new place, dance, and seek necessity. A computer is only aware of what we program it to do. I can see blue without having a word to describe it. Very unlike a computer stuff.
Go put a Thinkpad in a jungle hut on battery power. Come back after a month. It will be dead and not have even tried to survive.
Your conceptual model is way too vague.
We are from being just like a computer. Frankly I think you should probably shut yours off for a while.
Same can be said for you? A "Thinkpad" is nowhere near as sophisticated as a human. Your challenge equates to something like "Go put a human pancreas cell in a jungle hut in a bath of nutrients. Come back in a month. It will be dead and not have even tried to survive."
At a certain level of abstraction, the matter you are comprised of is indistinguishable from a computer, while still being a useful comparison: despite the means, circuits and cells both take signals, process and return modified signals. Likewise, we can compare both of these things to a member of a colony organism, such as an ant.
Though each element has different levels of sophistication the actual "sentience" occurs at a higher level... a sum of all the signals leading to a decision. For a human, whose thoughts surf over a network of squishy cells, this decision might lead to something like movement or dance. To an ant colony - movement or dance is not really an option... it is a big deal to move the colony, no matter how much movement any given ant might display.
For an AI - well it would entirely depend on attached mechanisms. Have you not seen humanoid robots move or dance? Given enough circuits in an appropriately sophisticated arrangement, could we conceive of a robot that observes visual signals, makes decisions about the information, and subsequently changes it behavior in response? That sequence alone qualifies that mechanism as "Sentient", though we might not be able to declare that it is "Sapient" or has "Consciousness".
> Frankly I think you should probably shut yours off for a while.
Frankly, I think you should read something like a basic text describing these issues. A simple primer to give you at least a basic footing in this conversation. The concepts I've just described are decades old and have been continuously poured over by thousands of people much smarter than you or I.