What Is “Collapsing” Superintelligent AI? | Joseph Corabi

Añadido: 2026-05-07

[00:00:00]It seems like AI systems today get around the problem of skepticism not by coming to a halt but by just making something up. Yeah. Yeah. Yeah. Know this is certainly the the problems that we have with AI today are almost the opposite, right? The overconfidence um you know large language models like chat GPT right they'll just spout off um completely false things with complete confidence.

[00:00:27][Music] Welcome to Closer to Truth. We are pleased to partner with Mindfest 2025 led by Professor Susan Schneider on this year's theme, sentience, autonomy, and the future of human AI interaction.

[00:00:48]Closer to Truth is conducting a series of in-depth discussions with keynoters and participants. Today I'm pleased to be speaking with Professor Joseph Kurabi, chair of philosophy at St. Joseph University. Great to meet you, Joe. I love your work and look forward to our discussions. Robert, you as well.

[00:01:07]I'm really looking forward to it, too.

[00:01:09]So, let's begin by discussing your presentation at Mindfest in some depth.

[00:01:13]And you call it a collapsing super intelligent AI. So, let's uh AI, we sort of know what it means. Let's define those other two terms and tell me what the paper's about. Yeah. So, a super intelligent AI is an AI that equals or exceeds human beings in basically every capacity that matters, every cognitive capacity. And generally, a super intelligent AI will only equal human beings rather than exceeding them where human beings are already kind of topping out the scales. So, where human beings are already sort of performing optimally. And then a collapsing super intelligence is a super intelligence that is unstable in its super intelligence. So it maybe achieves super intelligent levels of um of ability but then it somehow pairs back those abilities. It either just stops using them or actually disassembles whatever gives it those abilities in the first place. So let's use some specifics where you say where humans are topping out already. What are some of those cognitive uh uh C categories? Yeah, I mean there may not be any um but maybe things like facial recognition, certain kinds of pattern recognition, but there there aren't going to be many very many abilities I think where humans are already topping it out and there's not much room for improvement. Okay, we say improvement that's from a human point of view, but AI can do better face recognition than I can. I can, you know, barely remember faces. So, oh yeah. So, no, I was thinking may maybe humans at are are really that good at at facial recognition that, you know, if you give us thousands and thousands of faces that we know, we can differentiate them very reliably. But I'm not in any way committed to that. I mean, maybe humans are not close to topping out any cognitive ability. Okay. Yeah. But super intelligent AI would mean that it's at least uh achieving or exceeding each of those categories where humans are and to be super intelligent you're above that and and the articulation between the word super intelligent as you're using it and as others would use singularity where AI surpasses human intelligence uh they're close. They're very similar.

[00:03:37]Yeah. Yeah. So sometimes people use singularity to talk about you know the point at which the improvement is um virtually infinite um across time like the improvement is getting so fast that um it's practically infinite in I and I'm not so f focused on the dynamic nature of how fast is the improvement happening so much as just that whatever we have right now right now at whatever time we're speaking about is way better than humans are. Okay. All right. So your uh argument though um and approach focusing on collapsing. So the assumption that most make that when you achieve super intelligent AI the singularity then it becomes uh exponentially explosive um and uh can't be stopped. But the word collapsing means that that there's something that that could undermine that. Yes. Yeah. So, so I kind of think of the argument and I'm not in my my my talk is sort of an exploratory talk. It doesn't take any really strong stands, but it's really meant to to put some suggestions out there that I don't see being talked about a whole lot. That's important. Absolutely. Yeah. Absolutely.

[00:04:56]Yeah. I mean, I I don't think there's much value in sort of just like defending something to the death. Um Yeah. You have to be humble I think and and you know aware of where the limitations are. But so so basically yes I mean I think you can think of it two ways. You can think of it as the potential that the um the super intelligence sort of starts accelerating and then hits a wall or hits a slowdown and then turns back around. Or another possibility is it's kind of a wall and as the artificial intelligence approaches that wall, the wall stops it from achieving certain kinds of abilities before it ever really reaches what we would strictly call super intelligence. And do you have categories that this would apply to or is it just a general idea? In other words, could it be coding or mathematics or searching for data um or um um dealing with abstract ideas in uh in in in human philosophy? I mean are all do you differentiate categories? I do. Yeah. So it's specifically in moral reasoning that I think these these obstacles would hit. Yeah. not so much in um sort of your more your non-filosophical um non-scientific topics, but also not in your regular scientific topics, but I think specifically the issues might come up with with moral reasoning. Okay. And and what is your standard for judging the AI? Because maybe you're thinking it's collapsing when the AI is actually seeing something deeper than you're able to do. Ah, yeah. So, that's a good question. I I think the way we would recognize that it was collapsing is it just very clearly ceases to engage in that sort of activity. Um and just no longer even attempts to exceed what human beings would be doing in the same area. Um it's not just disagreeing with what your position is and you define that as collapsing. Anybody who disagree with you would be a collapse. Yeah.

[00:07:01]Yeah. This won't be like the chess playing computer that uh comes up with an unorthodox move and the the chess experts say, "Oh, no, it's making a mistake. It it doesn't know what it's doing." And then eight moves later, it's a it's a genius uh position that it creates, right? This this will be like the chess playing computer that just says, "No, I'm not going to play chess anymore." And what would what would drive that in terms of the internal mechanism? Yeah. So, I think it really depends on how the AI is set up. Um so when people talk about sort of the dangers of super intelligent AI they often talk about the AI having goals and there are different ways of specifying those goals. Now how the sort of how this this matches up with the actual computer science architecture is another set of issues but generally you know people talk about direct and indirect specification of AI goals and a direct specification would be you know program into the AI or directly train the AI to have very specific goals you know maximize human happiness something like that. Yeah. And none of this requires AI consciousness, right? Yes. In fact, yes. None of this um all of the the arguments I'm making are really about sort of the the practical behavior of the AI, right? So, so AI consciousness issues, those kinds of deeper philosophical questions don't even really come up except in tangential ways, I would say. Right. I agree. um you you've approached u um uh super intelligent AI um from a different perspective uh in terms of the dangers of AI and the assumption of uh of of linear progression or and even exponential progression. But you you have approached and used the word skepticism that you have you you embedded the word skepticism in this whole way of of of thinking about the dangers and so how does that articulate? Yeah. So so I mean issues I think there's another set of issues about a skepticism in the sense that AIs themselves could be trapped in skeptical concern. So people who are familiar with philosophy, I'm sure a lot of your viewers have have seen um episodes that discuss kind of related issues, but philos no no secret to those who have um some familiarity with philosophy, there's a lot of concern about skepticism in philosophy, not being able to reach answers to various questions. Now, we human beings, and this is an idea that goes way, way back in the history of philosophy, at least to people like David Hume in the 18th century, the idea that the reason that we human beings don't get that bothered by our inability to answer questions is we kind of have a cognitive architecture that allows us to ignore our failure to answer questions. We don't become paralyzed when we fail to answer those questions. we just kind of have behavioral systems that kick in and allow us to to kind of keep going. David Hume had this this famous passage um I believe it's in um his treatise of human nature where he says and this is a paraphrase but when I go to my study all of these questions about how do I know anything? Is there a world out there?

[00:10:27]How do I know that other people have minds? The these questions vex me. But I leave my study and I have a drink with my friends and I play back gammon and it all recedes into the background. Yeah.

[00:10:40]Um so we human beings have that but that there's no guarantee that a super intelligence will have that sort of backup system that it can that that will kick in when it gets stuck in various worries. Now, you might you might say, well, if it's super intelligent, it won't get trapped in these worries. It will solve the problems that human philosophers don't solve. Um, that's debatable. I mean, the the super intelligence might help it or the super intelligence might not be that useful because these are perspectival problems that no matter how much intelligence you have, that intelligence beyond a certain point just doesn't really crack the nut.

[00:11:25]Um, I I think today when we play with the leading AI systems, they they always, at least when I put in questions that have the things I know a lot about, they always in some degree hallucinate and come up with statements that are I I know affirmatively not true. For example, I ask about myself in certain ways and they and they state it with exactly the same level of certainty as all the things that are true. Yes. And so it seems like AI systems today get around the problem of skepticism not by coming to a halt but by just making something up. Yeah. Yeah. Yeah. Know this is certainly the the problems that we have with AI today are almost the opposite right the overconfidence um you know large language models like chat GPT right they'll just spout off um completely false things with complete confidence. Um but I think as these AIs improve now we may not be talking about large language models. I don't know the future of large language models versus other sorts of paradigms but um some of these other paradigms or maybe even large language models as they mature.

[00:12:36]They they may have sort of more um they they may begin to retreat and start going in the other direction. start getting very concerned about responsibility, about being um not going beyond their evidence, really being humble. Um it's hard to know exactly how their architecture will influence this, but it's very it's very possible we'll see the opposite sort of Well, that that would be I would think a net positive if if they retreat rather than hallucinate.

[00:13:05]Oh, yes. Yeah. But but they I mean the rational response at a certain point might be to retreat so far that they're not able to really say or do anything.

[00:13:16]Um if they become hyperrational in a certain way but you can program around that I would think.

[00:13:25]Um maybe but I think it will depend on the architecture because I mean I so I you can certainly program around it when the AI has a certain level of sophistication a certain low level of sophistication but as the AI sophistication grows those attempts to program around it might might you know reduce to the point where they're just no longer effective at all. the AI just looks at it and and you know red does its programming. H Joe, I I I don't want to pry personally, but as as a chair of philosophy at a Jesuit university, how would one's theistic beliefs affect perspectives of anything have to do with AI consciousness, AI transhumanism, mind uploading, virtual immortality, all of these uh sort of uh advanced ways of thinking, but all involving consciousness which impinge on certain theistic doctrines. Yeah. No, it's a good question. Um, I mean, the Jesuits in particular tend to be pretty open-minded and have not generally enforced a lot of um sort of um you can't explore doctrinally questionable or doctrinally um problematic positions. So um yeah, I don't feel a lot of pressure uh to to conform to any sort of preset um expectations about where the arguments are leading or where the conclusions are leading. Um now I do think there are substantive issues that sometimes just aren't that appreciated by sort of casual religious people or even casual uh casually religious intellectuals. It's sort of like, oh, there's religion, there's Christianity, or whatever particular religion that I believe in, and then there's, oh, this this cool AI stuff where a bunch of things are happening and never do the twain meet. But I think, you know, it's it's very easy to kind of silo things in that way. um and not see that oh there there may be implications here for my specific theistic beliefs or my specific confidence or lack of confidence in AI right and uh you know I I wasn't so much focusing on the external pressures that would would be case because I agree that's that's a Jesuit position is to be open-minded about those things that that are not sort of core doctrally uh but one's personal beliefs I'm not talking about specifically, but if if one had a theistic belief or a Buddhist belief, um how might those affect your approach to these issues of AI consciousness or mind uploading and Oh, I see. I see. Yeah.

[00:16:14]So, it's a good question. I don't have a fully worked out position on that. I mean, I do think I have a I have a colleague at St. Joseph by the name of John Keller who's writing a paper specifically about um whether AI whether the the existence of strong AI human level AI would be evidence against Christianity or not.

[00:16:37]That's good. Yeah. I his tentative view which I share is um that it would be some evidence against Christianity. I think um it's sometimes useful in these situations to think about how did people within the religious community think about these issues in the past before they were sort of realistic threats or what were parallel issues historically.

[00:17:02]So in Christianity, you know, a great example of some of this, I think, is um when Europeans came to the new world, uh how did Europeans think about the natives when they encountered them? Did they think that they were human beings?

[00:17:17]Did they think they had souls? Were they affected by original sin? All sorts of core theological questions. And there were there was debate about that. But um so so anyway to you know before uh going down some sort of rabbit hole with that huge topic I'll say the reason I think it's some evidence um is that if there were strong AI there'd be some decent reason to think that it would be conscious and the idea that humans could create conscious beings from scratch not by manipulating pre-existing living matter or something like that but could make it from scratch.

[00:17:56]doesn't I mean it's not impossible if God is in charge and God is the God of Christianity but it strikes me as not fitting in quite as well with the picture of that God then you know some than a world where God is the only sort of fundamental source of that level of um of impressiveness in the world that sophisticated life I think that's an excellent analysis I mean we could use the analogy which is much stronger which is the problem of evil which you have to deal with other issues of astrobiology.

[00:18:30]If there are other yeah conscious entities on all different kinds of worlds do you have to have multiple incarnations or one incarnation solves all of them? You know, those are the kinds of issues. And what you're saying basically now is AI consciousness is another kind of category that can impinge on your um estimation of the validity of of certain religious traditions. Yes. It's just it's just another data point. Yeah. Uh in in that and and it would be a data point on the negative side. So I I I think that's a a good answer. Very very quickly as a philosopher of of religion, how do you analyze some of the other themes at the conference on AI? What strikes you?

[00:19:14]Yeah. The transhumanism themes for instance. Yeah. Yeah. Um I mean some of the themes are sort of at a nitty-grittier level um sort of like well will will this particular type of weapon get out of control? And those I think have sort of less evidential impact. But the sort of transhumanism uh or at least direct connections with kind of the religious themes the transhumanism transhumanism issues I you know it's a good question. I don't have fully worked out views about um you know is uh pursuing sort of really radical human enhancement compatible with a Christian worldview say.

[00:19:58]Um, I mean, my my quick take on that would be it's definitely an area for extreme caution. Um, but I don't know that I'm convinced that it's completely ruled out by settled sort of doctrinal commitments or views that are well supported by by the evidence that's evidenced for Christianity.

[00:20:20]Good. Joe, let's do some short questions and some very short answers to uh some things we're asking uh everyone. So, question one, can AI ever be conscious as we are conscious? Zero to 100, zero meaning impossible, 100 meaning it's absolutely sure.

[00:20:40]35 two is virtual immortality possible which is uploading our first person consciousness not a digital clone but actually you where it's your first person consciousness is it possible in principle given a trillion years doesn't matter how long the time period zero to 100 um sorry what type of upload are we a a a a a firsterson consciousness not a digital clone but actually you where you feel it's your same consciousness in this new environ term. Okay. And it is a new environment. It's not uh you know in your brain. Uh yeah. Yeah. Yeah. Um 10 three. Will AI generate some very unpleasant surprises for humanity? Zero to 100.

[00:21:26]80.

[00:21:30]Four. The ultimate human future is transhumanism melding with AI. 0 to 100.

[00:21:39]I'll say 25.

[00:21:41]Finally, the ultimate future is nonhuman. All AI, no humans, zero to 100.

[00:21:51]20.

[00:21:53]Many thanks, Joe. Viewers can watch more mindfest 2025 videos on the Closer to Truth website and YouTube channel. Also over 1500 videos in more than a 100 TV episodes on consciousness, all facets, free will, personal identity, pansychism, dualism, idealism, parasychology, life after death, cosmic consciousness, but all infused with critical thinking. Thanks everyone for watching Closer to Truth. Thank you, Robert.

[00:22:23]Thank you for watching. If you like this video, please like and comment below.

[00:22:28]You can support Closer to Truth by subscribing.

[00:22:33]Closer to Truth is now accepting your taxexempt donations. Please come to closertoruth.com/donate. Thank you very much for supporting us and thanks for watching.