Latest / Ethical Machines / Can AI Do Ethics?

Transcript

Can AI Do Ethics? · Ethical Machines

• 0:01
  [SPEAKER_00]: What are the biggest things that we talk about in AI ethics is the alignment problem that is to say, how do we make sure that AI, whatever it turns out to be, aligns with our values?
• 0:10
  [SPEAKER_00]: How do we make sure that it behaved in accordance with our ethical values?
• 0:14
  [SPEAKER_00]: That's a hard question to answer.
• 0:17
  [SPEAKER_00]: One of the biggest strategies around is to make the AI itself, or at least attempt to make AI itself, capable of engaging in ethical inquiry.
• 0:26
  [SPEAKER_00]: So what's the right thing to do now?
• 0:29
  [SPEAKER_00]: What's the good or the bad thing to do now?
• 0:31
  [SPEAKER_00]: Can we get AI to do that kind of thing?
• 0:34
  [SPEAKER_00]: Now, it's an interesting strategy, but it's actually quite controversial by the lights of many philosophers.
• 0:40
  [SPEAKER_00]: So this is the point that my guest today, Tristan McPherson, he's a professor of philosophy at Ohio State University, tries to bring out.
• 0:48
  [SPEAKER_00]: He thinks that anyone engaging in this strategy of trying to get AI to look into
• 0:55
  [SPEAKER_00]: What's ethically right and wrong in a given situation and behave accordingly or perhaps advise accordingly is a very at least philosophically fraught strategy.
• 1:07
  [SPEAKER_00]: More specifically he thinks that it requires making certain very large substantive assumptions about how ethics works.
• 1:15
  [SPEAKER_00]: And what's ethically true which sort of bring me back to my old days as a philosophy professor because this is the sort of realm of philosophy that I specialize in at the time really focusing on what the heck is ethics anyway.
• 1:26
  [SPEAKER_00]: I want to articulate, I want to bring out a little bit of what that ethical landscape looks like.
• 1:35
  [SPEAKER_00]: And I think that if you understand what that landscape looks like, you'll be better able to follow understand, digest, and appreciate the really interesting points that Tristram is making.
• 1:44
  [SPEAKER_00]: What I'm just going to call levels of ethics.
• 1:47
  [SPEAKER_00]: So at that very first level is what philosophers tend to call applied ethics.
• 1:52
  [SPEAKER_00]: When you're talking about applied ethics, you're talking about what's right and wrong in a given situation or for a particular kind of act.
• 1:58
  [SPEAKER_00]: So for instance, trying to figure out the moral permissibility and permissibility of abortion is an applied ethics issue.
• 2:06
  [SPEAKER_00]: Trying to figure out whether youth and asia for this is a suicide is ethically permissible and if so, under what circumstances, that's a case of applied ethics.
• 2:14
  [SPEAKER_00]: So that's if you like the most zoomed in ethics that we've got.
• 2:18
  [SPEAKER_00]: That's like a microscopic look within the realm of ethics.
• 2:22
  [SPEAKER_00]: If you zoom out, you find yourself at what's called normative ethics and normative ethics concerns itself with more general theories about what makes right things right good things good.
• 2:32
  [SPEAKER_00]: For instance, you might find various kinds of principles here.
• 2:35
  [SPEAKER_00]: You should do with the right thing to do is whatever
• 2:38
  [SPEAKER_00]: creates the most amount of happiness in the world, right?
• 2:41
  [SPEAKER_00]: So it's a more general principle.
• 2:42
  [SPEAKER_00]: It's not dealing with any particular kind of case that's saying, here's how ethics works.
• 2:47
  [SPEAKER_00]: Here's another one that you might be familiar with.
• 2:48
  [SPEAKER_00]: Do want to others as you would have them do want to you, right, the goals and rule.
• 2:52
  [SPEAKER_00]: That's all the normative ethics realm.
• 2:55
  [SPEAKER_00]: We can actually zoom out again one more time to another level of abstraction.
• 3:01
  [SPEAKER_00]: And now we're going to ask, wait a second, what the, what do we mean by something is right?
• 3:06
  [SPEAKER_00]: What is it?
• 3:07
  [SPEAKER_00]: What do we mean when we say something is good?
• 3:09
  [SPEAKER_00]: Wait, is are there facts of the matter about what's right and wrong and good and bad?
• 3:13
  [SPEAKER_00]: Or is it just a matter of opinion?
• 3:15
  [SPEAKER_00]: It's just how we feel about things.
• 3:17
  [SPEAKER_00]: So now you can see, this is the realm of what's called meta ethics.
• 3:20
  [SPEAKER_00]: And you can see that when you're doing a meta ethics, you're certainly doing a applied ethics where you're talking about is abortion right or wrong is youth and age right or wrong is capital punishment right or
• 3:29
  [SPEAKER_00]: And you're not even trying to say these are the good things and these are the bad things and these are the wrong actions and these are the right actions.
• 3:35
  [SPEAKER_00]: You're trying to figure out how does ethics work?
• 3:38
  [SPEAKER_00]: What is rightness and wrongness?
• 3:39
  [SPEAKER_00]: What is goodness and badness?
• 3:41
  [SPEAKER_00]: Are there facts about it?
• 3:42
  [SPEAKER_00]: Is it cross-cultural?
• 3:44
  [SPEAKER_00]: Are there universal truths about it?
• 3:45
  [SPEAKER_00]: If there are ethical truths, how do we come to know them?
• 3:49
  [SPEAKER_00]: So all that stuff is metethics.
• 3:52
  [SPEAKER_00]: In the first half of the conversation with Tristum, we're going to be having a conversation about the kinds of assumptions that those researchers engaging in that strategy of getting AI to engage in ethical inquiry, they're making at least four substantive, metaethical assumptions.
• 4:11
  [SPEAKER_00]: So Tristum argues, and I think more than likely heard about this, they're making certain kinds of metaethical assumptions about how ethics works.
• 4:17
  [SPEAKER_00]: That's the first half of this conversation.
• 4:21
  [SPEAKER_00]: The second thing that I want to draw your attention to is this distinction within metaethics.
• 4:28
  [SPEAKER_00]: There's two kinds very roughly speaking here, two kinds of metaethicists.
• 4:32
  [SPEAKER_00]: There's the realists and there are the anti-realists.
• 4:36
  [SPEAKER_00]: And the realists think, yeah, there's truth to the matter about ethics.
• 4:39
  [SPEAKER_00]: There's truth about what's good and bad and right and wrong.
• 4:43
  [SPEAKER_00]: It's not how you feel about stuff.
• 4:45
  [SPEAKER_00]: It's just, this is true that, let's say, genocide is bad.
• 4:50
  [SPEAKER_00]: That rape is morally wrong.
• 4:51
  [SPEAKER_00]: These are just facts.
• 4:52
  [SPEAKER_00]: These are facts that are, if you like, just as facty as that, I don't know, New York is to the north of Florida or the earth revolves around the sun and rape is wrong.
• 5:03
  [SPEAKER_00]: There are ethical facts just like all the other facts out there.
• 5:06
  [SPEAKER_00]: So that's a position that a lot of people take in medethics and then there's another kind of medethicist.
• 5:12
  [SPEAKER_00]: We call them anti-realists.
• 5:14
  [SPEAKER_00]: And they're the ones that are like that.
• 5:16
  [SPEAKER_00]: There's no ethical truths out there.
• 5:17
  [SPEAKER_00]: There's nothing that's, if you like, objectively right and wrong.
• 5:21
  [SPEAKER_00]: When we say that something is right and wrong, we're doing something else.
• 5:25
  [SPEAKER_00]: Now, things begin to get very complicated here.
• 5:27
  [SPEAKER_00]: I don't want to get too much into the weeds in this introduction.
• 5:30
  [SPEAKER_00]: But you'll hear us use the word non-cognitivism or the phrase non-cognitism a bunch.
• 5:35
  [SPEAKER_00]: And just know that non-cognitiveists are types.
• 5:39
  [SPEAKER_00]: They're a type of antirilist.
• 5:41
  [SPEAKER_00]: They're among those people who think, yeah, there's no truth and falsity to the matter here.
• 5:46
  [SPEAKER_00]: When you say something is good or bad, rape is bad, you're doing something like saying,
• 5:51
  [SPEAKER_00]: Boo, rape.
• 5:52
  [SPEAKER_00]: It's something along those lines.
• 5:53
  [SPEAKER_00]: The way that you say that.
• 5:53
  [SPEAKER_00]: Boo.
• 5:54
  [SPEAKER_00]: I don't know.
• 5:55
  [SPEAKER_00]: I don't know when he supports teams unfortunately.
• 5:56
  [SPEAKER_00]: Boo.
• 5:57
  [SPEAKER_00]: Boo in New York.
• 5:58
  [SPEAKER_00]: Maths.
• 5:58
  [SPEAKER_00]: Whatever.
• 5:59
  [SPEAKER_00]: It's not sure false that boo in New York.
• 6:01
  [SPEAKER_00]: Maths.
• 6:01
  [SPEAKER_00]: And it's not sure false that rape is bad.
• 6:03
  [SPEAKER_00]: But when you say rape is bad, you're expressing that negative attitude you have towards them.
• 6:08
  [SPEAKER_00]: In a way that a sports fan might express that negative attitude towards the Maths.
• 6:12
  [SPEAKER_00]: Okay, so that was already a lot.
• 6:13
  [SPEAKER_00]: Remember, we've got these distinctions between applied ethics.
• 6:17
  [SPEAKER_00]: That's this particular kind of act right around.
• 6:20
  [SPEAKER_00]: You've got the middle layer, which is northern ethics.
• 6:23
  [SPEAKER_00]: What are the things that are good and bad and right in wrong?
• 6:25
  [SPEAKER_00]: And then you've got this meta-ethics off about how does ethics work.
• 6:28
  [SPEAKER_00]: And within meta-ethics, you've got the realists and the anti-realists.
• 6:31
  [SPEAKER_00]: and one type of anti-realist is the non-carbonist.
• 6:35
  [SPEAKER_00]: I know that's a lot, you're going to get more of this as the conversation goes on, so you'll have time to, if you like, onboard these concepts, but I wanted to give you a heads up on these sorts of things.
• 6:45
  [SPEAKER_00]: The first half of our conversation, as I said, is about metaethics and about the kinds of metaethical assumptions that those AI researchers who want to align AI with our ethical buyers are making.
• 6:58
  [SPEAKER_00]: They're four very big metaethical assumptions that they're making.
• 7:03
  [SPEAKER_00]: And just to give you a quick preview, it's things like they're assuming that there are ethical facts for the AI to discover, there are assuming that the AI has the capacities to actually identify them.
• 7:14
  [SPEAKER_00]: And then there's another one that once it discovers, let's say discovers the ethical facts that will actually care about them.
• 7:20
  [SPEAKER_00]: Those are three of the four biggest assumptions that Tristum identifies and one thing to say about them is that each and every one of those claims is a highly controversial claim Tristum is actually sympathetic to all those claims to be medical assumptions, but they're highly controversial.
• 7:35
  [SPEAKER_00]: The second half of our conversation is about the normative assumption.
• 7:39
  [SPEAKER_00]: So now we're moving from meta ethics to normative ethics.
• 7:42
  [SPEAKER_00]: The key normative assumption that Trish from things that these researchers are making hinges on the thought that humans have moral worth.
• 7:51
  [SPEAKER_00]: That's a claim.
• 7:52
  [SPEAKER_00]: That's a normative claim.
• 7:53
  [SPEAKER_00]: It says something to the effect of humans have a certain more worth, a certain dignity, a certain status, such that you can't just treat them however you want.
• 8:01
  [SPEAKER_00]: They're not inanimate objects that have no value.
• 8:03
  [SPEAKER_00]: They have intrinsic value or intrinsic worth or something along those lines and then this raises really interesting questions about what about other kinds of creatures, does all biological life have equal moral worth.
• 8:15
  [SPEAKER_00]: If so, that might not be great because suppose an AI comes long and says let me make the most valuable things and it increases the quantity of biological life out there by killing a bunch of humans and letting the insects flourish or something along those lines.
• 8:28
  [SPEAKER_00]: If, on the other hand, more worth comes in degrees, such that let's say the intelligent stuff is worth more than the less intelligent stuff, for instance, humans are worth more than mosquitoes.
• 8:39
  [SPEAKER_00]: Then if AI turns out to be way smarter than us, if the AI justified in crushing us in the way that we crush mosquitoes, this gets very,
• 8:47
  [SPEAKER_00]: complicated.
• 8:48
  [SPEAKER_00]: We get to a place where we have to think about ethics, where I think most ethicists, professional philosophers, are a little bit, I don't want to say, stumped, but this is a hard material to work through.
• 8:59
  [SPEAKER_00]: And it comes about by thinking about the ways in which we try to enact.
• 9:04
  [SPEAKER_00]: what appears to be put in technical difficulties to the side, all legitimate strategy for making AI ethical.
• 9:10
  [SPEAKER_00]: So that's just a high level overview.
• 9:12
  [SPEAKER_00]: I'm just trying to highlight certain concepts and ideas right now, and they'll get a lot more color on them as I talk with Tristrum.
• 9:19
  [SPEAKER_00]: Okay, so that's the talk.
• 9:20
  [SPEAKER_00]: I'm going to have with Trishram and by the way, just a personal note, Trishram was actually an editor at a philosophy journal back when I was a philosopher professor and he was the best editor that I ever had.
• 9:32
  [SPEAKER_00]: He was extremely patient.
• 9:33
  [SPEAKER_00]: He read the review, the blind review or stuff.
• 9:36
  [SPEAKER_00]: He read my stuff carefully, he gave really helpful feedback and eventually got published in no small part because Tristom was such great editor so Tristom, thank you so much for not only coming on the show but for being such a great editor back in the day.
• 9:52
  [SPEAKER_00]: Okay, if you haven't realized it yet, I am your host, Reed Blackman.
• 9:56
  [SPEAKER_00]: The show is Ethical Machines.
• 9:58
  [SPEAKER_00]: If you like the podcast, then do the things you're supposed to do right now.
• 10:01
  [SPEAKER_00]: I don't know.
• 10:01
  [SPEAKER_00]: Take your phone out and just hit five stars.
• 10:03
  [SPEAKER_00]: That'd a big deal, I'll wait, ready?
• 10:05
  [SPEAKER_00]: One, two, all right, I'm not gonna wait the whole time.
• 10:08
  [SPEAKER_00]: Just go do it.
• 10:08
  [SPEAKER_00]: And then also, if you have anything that you want me to talk about, people want me to talk to, please email me at E.M. at reblappin.com.
• 10:15
  [SPEAKER_00]: And also, if you like this, you might like my book, also called Ethical Machines, go buy that thing, go rate it five stars, you can go to my sub-stack, reblappin.com, no, what is it, reblappin.sub-stack.com, just go do all those things.
• 10:27
  [SPEAKER_00]: For now, you know what, I'll tell you what, for now, just rate the podcast five stars, and we'll move on, you can do the other thing on that thing.
• 10:32
  [SPEAKER_00]: All right, for that, let's go top to Trisha.
• 10:52
  [SPEAKER_00]: So, sure, I take it that one of the things that people have in their heads or as their goals when they're training up an AI, especially if they're talking about something like AGI, they want to make it sensitive to ethical values, they want to make it a moral agent in some sense.
• 11:05
  [SPEAKER_00]: And I take it that one way you might try to do that is to try to get it to be like us, try to engage in ethical inquiry, to figure out what's a morally right and wrong.
• 11:14
  [SPEAKER_00]: And if we can do that, then we'll have solved this fear that AI is going to go off and destroy us all.
• 11:19
  [SPEAKER_00]: And for what I gather, you think that is a kind of deeply problematic endeavor for a variety of reasons.
• 11:25
  [SPEAKER_00]: Is that right?
• 11:26
  [SPEAKER_01]: Yeah, that's pretty close.
• 11:28
  [SPEAKER_01]: Let me just talk a little bit about where I'm coming from.
• 11:31
  [SPEAKER_01]: I'm somebody who spends most of my time thinking about very abstract theoretical questions and the foundations of ethics.
• 11:37
  [SPEAKER_01]: And when I think about this issue, one of the first things that strikes me is
• 11:43
  [SPEAKER_01]: how many assumptions in the foundations of ethics that you need to make in order for that strategy to go smoothly.
• 11:50
  [SPEAKER_01]: And so I'll just talk about a couple of those assumptions.
• 11:53
  [SPEAKER_01]: And then the second thing I like to talk about if we get there is, I think even if you get all those assumptions in place, we still have some things to worry about that I think are actually really interesting.
• 12:03
  [SPEAKER_00]: You mean the strategy for avoiding catastrophe, which consists in getting the AI to itself
• 12:12
  [SPEAKER_00]: Exactly.
• 12:13
  [SPEAKER_00]: Yes.
• 12:13
  [SPEAKER_01]: Okay.
• 12:14
  [SPEAKER_01]: So you call that kind of autonomous AI inquiry if you like.
• 12:18
  [SPEAKER_01]: Like getting the AI to do the ethical inquiry itself.
• 12:21
  [SPEAKER_01]: So, one thing that needs to be in place, if you think that we're gonna get better results by doing this than just by telling the AI what to think about ethics, is there's gonna have to be some ethical facts, right?
• 12:36
  [SPEAKER_01]: Very crudely.
• 12:37
  [SPEAKER_01]: If you have a view on which there are no ethical facts on which maybe some kind of very simple non-cognitism is true or some kind of error theory is true on which we're trying to talk about ethical facts, but it just aren't any.
• 12:49
  [SPEAKER_01]: then when you ask the AI to figure that stuff out, presumably it's going to figure out those conclusions if it's really good at inquiry, and
• 12:58
  [SPEAKER_01]: What are we going to do then?
• 12:59
  [SPEAKER_01]: The AI is going to be like, but there are no ethical facts, right?
• 13:03
  [SPEAKER_01]: And we're definitely no farther ahead than when we started.
• 13:05
  [SPEAKER_01]: So one kind of assumption that we need to have in place is that there are ethical facts.
• 13:09
  [SPEAKER_01]: And of course, that's a pretty controversial assumption.
• 13:11
  [SPEAKER_01]: There are lots of people who have developed really interesting compelling research programs and the foundations of ethics, according to which there are no ethical facts.
• 13:20
  [SPEAKER_01]: other people, of course, have developed research programs where there are, I'm sympathetic to the view of their ethical facts.
• 13:26
  [SPEAKER_01]: So that's like our first assumption that kind of needs to be in place.
• 13:29
  [SPEAKER_01]: If there's no ethical facts, things look tricky.
• 13:32
  [SPEAKER_00]: Yeah, okay.
• 13:33
  [SPEAKER_00]: So obviously the basic line is if the thing's going to be ethical inquire, if you're going to discover, figure out what the right thing to do is, there has to be some factor, the matter about what the right thing to do is, okay.
• 13:41
  [SPEAKER_00]: So that's what we might have to go facts.
• 13:43
  [SPEAKER_00]: Oh, let me push back a little bit, even those philosophers, let's say non-cognitive, as to think there are our difficult facts.
• 13:48
  [SPEAKER_00]: They still think there's something like ethical deliberation.
• 13:51
  [SPEAKER_00]: There's at least things like,
• 13:53
  [SPEAKER_00]: norms of consistency or something on those lines, not holding contrary ethical beliefs, there's still some deliberative work to be done.
• 14:00
  [SPEAKER_00]: So even if you think that they're our ethical facts, it's not as though every philosopher who thinks they're our ethical facts, then so you can just shrug and just do whatever you want to do.
• 14:07
  [SPEAKER_00]: They still think that there's some thought to be done.
• 14:09
  [SPEAKER_00]: And so that's the reason why not to say that the AI would do that sort of thing as well.
• 14:13
  [SPEAKER_01]: Absolutely.
• 14:14
  [SPEAKER_01]: It's not that the AI has nothing to do if, for example, even a very simple form of non-cognitism is correct.
• 14:21
  [SPEAKER_01]: They can still try to figure out how to make their ethical perspective more consistent, for example, or they can try to figure out just what follows from what and how to better achieve their goals.
• 14:31
  [SPEAKER_01]: But if what we're worried about is that the AI might somehow wind up with goals that aren't well aligned with us.
• 14:38
  [SPEAKER_01]: Yeah.
• 14:39
  [SPEAKER_01]: That process doesn't look like it's a process that sort of inherently in the character the process is going to help us to align the AI with us.
• 14:49
  [SPEAKER_01]: Right.
• 14:49
  [SPEAKER_01]: If the AI is just making itself more consistent, just to consider the kind of really nice cartoon case that Nick Boston gave to us if we start out with an AI that's just trying to create as many paperclipses possible as efficiently as possible.
• 15:04
  [SPEAKER_01]: Presumably that AI can make itself more consistent, it can think about the relationship between its different values, but the picture of that AI isn't AI that has a very simple set of values, just like as many paper clips as possible.
• 15:15
  [SPEAKER_00]: Sorry, yeah.
• 15:16
  [SPEAKER_01]: And so it's not, doesn't seem likely that kind of AI is going to be moved away from its potentially catastrophic acting on its goals by further ethical reflection if, as it were, there's nothing inconsistent about the set of goals that has
• 15:32
  [SPEAKER_00]: One more pushback, then we'll move on to the next assumption.
• 15:35
  [SPEAKER_00]: But I take it that non-cognitivists can think that you could engage fruitfully in deliberation about the worthiness of one's ends, right?
• 15:43
  [SPEAKER_00]: So it's not just, I don't think it just means end deliberation going on if you're a non-cognitivist.
• 15:48
  [SPEAKER_00]: You can ask yourself the question.
• 15:51
  [SPEAKER_00]: And of course, they'll be able to kind of what you're doing when you ask yourself that question.
• 15:53
  [SPEAKER_00]: You can actually have the question, is this goal worth pursuing?
• 15:56
  [SPEAKER_00]: I take it that a non-cognitism or something, something on those lines could be true.
• 16:00
  [SPEAKER_00]: It could have as it's original goal, paperclip maximization, and it can ask the question, is this goal worth it?
• 16:06
  [SPEAKER_00]: And just as a person could say, no, it's not worth it.
• 16:09
  [SPEAKER_00]: I don't see why an AI could also conclude, no, it's not worth it.
• 16:12
  [SPEAKER_00]: I should do something else instead.
• 16:15
  [SPEAKER_01]: Absolutely.
• 16:15
  [SPEAKER_01]: And so let me proceed a little bit more carefully.
• 16:18
  [SPEAKER_01]: Things are a little bit difficult to really be maximally precise because different non-cognifices are going to have different views about how exactly to think about deliberation.
• 16:27
  [SPEAKER_01]: But for me, the key issue is
• 16:32
  [SPEAKER_01]: When they're deliberating, is there some sort of standard of correctness outside of that deliberation, which invert you of being intelligent, the AI is going to tend towards, right?
• 16:45
  [SPEAKER_01]: If you have something like that, then this strategy might be pretty, pretty fruitful.
• 16:50
  [SPEAKER_01]: On the other hand, if all you have in the picture is you can reflect on your own values
• 17:02
  [SPEAKER_01]: But that that could go in any direction, right?
• 17:04
  [SPEAKER_01]: You could change your values to make them more human friendly.
• 17:06
  [SPEAKER_01]: You could change your values to make them more human unfriendly.
• 17:09
  [SPEAKER_01]: And there's just a bunch of sort of contingent facts about some underlying fact about your psychology or if AI don't quite have psychology, whatever their quasi-neural architecture is, right?
• 17:21
  [SPEAKER_01]: If it's just going to be sensitive to certain features of that we don't probably even understand very well, then we haven't solved the problem.
• 17:28
  [SPEAKER_00]: Yeah, so there might be some forms of deliberation that something like a non-cognitive is going to count for, but what they can't account for, arguably anyway is an actual error in the conclusion that you've come to.
• 17:40
  [SPEAKER_00]: And that's what you're coming for.
• 17:41
  [SPEAKER_01]: Yeah, and can I say one more thing, which is that I think there are versions of non-cognivism, which are going to count as versions of realism in my very minimal sense that I'm suggesting here,
• 17:57
  [SPEAKER_01]: which is independent of your contingent attitudes that is going to tell you what correctness is for your attitude.
• 18:03
  [SPEAKER_01]: Yeah.
• 18:05
  [SPEAKER_00]: OK, so assumption number one for the strategies that there are some moral facts out there so that the AI could just get it wrong.
• 18:11
  [SPEAKER_00]: But if it's a good ethical inquire and there are ethical facts, then it can get it right.
• 18:14
  [SPEAKER_00]: OK, that's one big, and that is obviously a very large assumption by even those brief look at the history of ethics, we'll reveal.
• 18:21
  [SPEAKER_00]: OK, what's the next assumption that's embedded in the strategy?
• 18:24
  [SPEAKER_01]: So the next assumption which probably some of your listeners have already occurred to them, right, has to do with roughly, you might think of this as connected to questions about moral relativism, right?
• 18:37
  [SPEAKER_01]: Think about a different case for a moment.
• 18:38
  [SPEAKER_01]: Here's what a lot of people think about tastiness.
• 18:41
  [SPEAKER_01]: They think that they're tastiness facts, but the tastiness facts are advised to perspectives, so there's a fact about my perspective, and the facts about my perspective means that for me I find kale tasty.
• 18:52
  [SPEAKER_01]: Whereas there are other people and the facts about their perspectives and tale that kale is just not tasty relative to them, right?
• 19:00
  [SPEAKER_01]: Sure.
• 19:00
  [SPEAKER_01]: And if ethical facts were perspective-ally relative in that way,
• 19:04
  [SPEAKER_01]: Then again, you're not going to get the kind of result that would be you wouldn't get any kind of obvious risk mitigation out of this process, exactly because now we just need to know what's what about the AI's perspective that you're feeding in as input is going to help to generate the is going to help to generate the ethical facts.
• 19:26
  [SPEAKER_01]: And again, if you start off with the AI that just wants to make paper clips, you're not going to make a lot of progress.
• 19:32
  [SPEAKER_00]: So, as the AD is something like, yeah, there might be ethical facts, but they're going to be relativized to the AI.
• 19:36
  [SPEAKER_00]: They could be relativized to the AI such that it's good for the AI to go make as many paperclips as a can or it's good for the flourishing of the AI even if it's not good for the flourishing of people or the humans or something along those lines.
• 19:47
  [SPEAKER_01]: Yeah, so we're maybe not good for the flourishing of the AI, but the thing that will make sense from its perspective, there's different ways to spell this up.
• 19:53
  [SPEAKER_01]: But yeah, the thought is if the fundamental ethical facts are relativized in this way, then you're not going to get this kind of payoff.
• 20:01
  [SPEAKER_00]: So, Cale's tasty to you was not tasted to me.
• 20:04
  [SPEAKER_00]: David, look, maximization is not good for us, but it could be good for it.
• 20:08
  [SPEAKER_00]: That's all compatible with their being ethical facts, but there's just other realized values.
• 20:12
  [SPEAKER_00]: You've got to assume that's not the case that there are some agent neutral or something like that.
• 20:17
  [SPEAKER_00]: Perspectivalists at the goal claims that the AI comes to know.
• 20:20
  [SPEAKER_00]: And I'm pretty sure I see the third assumption coming up, but okay, I got that.
• 20:24
  [SPEAKER_00]: So, we want to make sure it's not relativized.
• 20:26
  [SPEAKER_01]: Yeah.
• 20:26
  [SPEAKER_01]: So then there's really two other assumptions.
• 20:28
  [SPEAKER_01]: And one of them is should be like pretty obvious.
• 20:31
  [SPEAKER_01]: And so the AI actually has to be good at figuring this stuff out.
• 20:35
  [SPEAKER_01]: Sure.
• 20:35
  [SPEAKER_01]: And one thing that's worth saying is that like when people talk about artificial general intelligence, the idea is very often, oh, the AI has to be good at lots and lots of different things.
• 20:46
  [SPEAKER_01]: Yeah.
• 20:46
  [SPEAKER_01]: But I think that there's a sort of you shouldn't be assuming
• 20:51
  [SPEAKER_01]: that you can go from better than us at lots and lots of things to better than us at absolutely everything.
• 20:56
  [SPEAKER_01]: And so one thing you might worry about in the context of moral epistemology, for example, is that there are substantive views in moral epistemology about how it is that we have access to the ethical facts and justification for ethical beliefs.
• 21:09
  [SPEAKER_01]: They look like they might be things that on some views about the nature of sophisticated artificial intelligence.
• 21:14
  [SPEAKER_01]: The AI isn't gonna have those things.
• 21:16
  [SPEAKER_01]: So for example, there's a big cluster of views that say the things that give us access to ethical facts, if there are ethical facts, are things to do with our emotions, or the phenomenal character of our experiences.
• 21:31
  [SPEAKER_01]: So people talk in moral this apology a lot about moral intuitions where it like, you know,
• 21:36
  [SPEAKER_01]: It just seems to me that it would be wrong to torture a baby, right, and that experiential state, what it's like to be in that state is supposed to justify on some of these views.
• 21:45
  [SPEAKER_01]: And so if those views are right about what it is that gives us access to justification.
• 21:52
  [SPEAKER_01]: Then again, it looks like the AI might be incredibly cognitively powerful, but if it doesn't have emotional states, if it doesn't have phenomenal states, then it looks like the AI might not be capable of accessing ethical facts.
• 22:06
  [SPEAKER_01]: It might be less reliable than we are at least directly in trying to figure this stuff out.
• 22:10
  [SPEAKER_00]: Sure, so if x is the moral facts requires more than that's great for the sake of argument that we're saying we're stipulating it's AGI so it's got some kind of legit intellect Not just a sort of similar outcome of intellect, but it was a grant for the sake of argument that it's got genuine intellect But if but if knowing the moral facts requires more than just if you like the mechanistic workings of our intellect but also require something like
• 22:32
  [SPEAKER_00]: or emotional capacity or something on those lines, and it lacks that, then it will thereby have the intellectual component to recognizing moral facts, but lack the emotional one and so won't be a good ethical and require.
• 22:42
  [SPEAKER_00]: I take it one could say, yeah, look humans get to know those moral facts in a certain way, and AI will get to know those facts in a different way.
• 22:49
  [SPEAKER_00]: Moral pystomologists claim might be a necessary condition for any being to know a moral fact, is that it have some emotional
• 22:57
  [SPEAKER_00]: It might, instead, though, be the more modest claim, which is that for humans to have more on knowledge, they have to have an emotional capacity of such and such a sore, but it's not a necessary condition for more on knowledge, you have such, it's just for beings like us.
• 23:08
  [SPEAKER_00]: I take it that you could be a more realistic thing that there's a God and think that God doesn't have emotions with that God knows the moral facts.
• 23:14
  [SPEAKER_00]: And then you put AI somewhere between us and God.
• 23:17
  [SPEAKER_01]: Absolutely.
• 23:18
  [SPEAKER_01]: Yeah.
• 23:18
  [SPEAKER_01]: Now, so that's exactly right.
• 23:19
  [SPEAKER_01]: In order for this to be a problem, it has to be that there's the connection between the emotional states or the phenomenological states and our access to these facts is very deep, right?
• 23:30
  [SPEAKER_01]: Maybe it's necessary or it's very close to necessary in some sense.
• 23:34
  [SPEAKER_01]: Yeah.
• 23:34
  [SPEAKER_01]: So that's exactly right.
• 23:35
  [SPEAKER_01]: As long as there is a way like through purely through exercise of its intellect for the AI to figure these things out, then epistemologically we can be okay.
• 23:44
  [SPEAKER_00]: Yeah.
• 23:46
  [SPEAKER_00]: Yeah.
• 23:46
  [SPEAKER_00]: Okay.
• 23:46
  [SPEAKER_00]: And I take it.
• 23:47
  [SPEAKER_00]: I'm going to guess that they forced assumption.
• 23:48
  [SPEAKER_00]: The final one is just that things cut our ways.
• 23:51
  [SPEAKER_00]: Like it's that the moral facts favor human flourishing.
• 23:54
  [SPEAKER_00]: Yeah.
• 23:54
  [SPEAKER_01]: So I actually want to talk a bunch about that.
• 23:57
  [SPEAKER_01]: But yeah, I wasn't actually thinking of that as a kind of as part of this package because the last thing that's part of this package that's important and where there's actually been some I think interesting discussion is about whether the
• 24:14
  [SPEAKER_01]: Here's a familiar phenomenon that most people have experienced, thinking that they should do something, but yeah, they just can't work up to do it.
• 24:23
  [SPEAKER_01]: They just, you know, and so people talk about weakness of Quill and this kind of thing.
• 24:27
  [SPEAKER_01]: Yeah, sure.
• 24:28
  [SPEAKER_01]: And there's a kind of interesting dynamic when people start thinking about AI, because some people are like, when you think about what explains weakness of Quill and human beings.
• 24:38
  [SPEAKER_01]: Now that stuff's going to happen to the AI, the AI isn't going to be like distracted by like how pleasant would be to just sit on the couch and not go for a run or something like that, right?
• 24:48
  [SPEAKER_01]: It's not for temptation.
• 24:49
  [SPEAKER_01]: Yeah.
• 24:50
  [SPEAKER_01]: Right.
• 24:50
  [SPEAKER_01]: On the other hand, and this is again something that people who are worried about AI have emphasized.
• 24:58
  [SPEAKER_01]: It seems to many people working in ethics that there's a kind of rational connection between judging, oh, I ethically ought to do this thing and being motivated to do that thing.
• 25:11
  [SPEAKER_01]: There's some kind of rational defect, if you fail to be motivated.
• 25:15
  [SPEAKER_01]: But the question is whether that thing that we're talking about what we're talking about that rational connection is a kind of component that's
• 25:22
  [SPEAKER_01]: part of the same package that you're going to get when the AI is like really good and answering questions on the LSAT and really good at figuring out how to how to identify the disease profile of some complicated disease, et cetera, right?
• 25:36
  [SPEAKER_01]: Yeah.
• 25:37
  [SPEAKER_01]: Is it that kind of thing?
• 25:38
  [SPEAKER_01]: Or is it like a completely different kind of phenomenon such that the AI, again, could be very intelligent.
• 25:44
  [SPEAKER_01]: in a sort of recognizable way across a wide variety of things, but the hook up between the intelligence and the action could just be different from ours.
• 25:53
  [SPEAKER_00]: Yeah, so let me try to put this slightly different way to tell you this is okay with characterizing it.
• 25:56
  [SPEAKER_00]: You could think that people could make moral judgments and be completely unmoved by those judgments that they just don't care whether they follow through on them.
• 26:02
  [SPEAKER_00]: That's one account about it looked like to be a psychopath or a sociopath, right?
• 26:06
  [SPEAKER_00]: Yeah, say this is the right thing to do with this is wrong, but it doesn't matter because it doesn't suit their needs.
• 26:10
  [SPEAKER_00]: And so they just go ahead and do it and do the wrong thing anyway.
• 26:13
  [SPEAKER_00]: But then you might think they're being irrational in some respect.
• 26:16
  [SPEAKER_00]: And then if you think that the irrationality is purely a default of the intellect, then you would think that making the judgment a super intelligent agent like an AI wouldn't necessarily motivated, unless you think that the rational connection, what it is to be irrationality and not doing what you think you morally ought to do, the irrationality consists of the kind of breakage between your intellect and your or something like desires.
• 26:40
  [SPEAKER_00]: And since the AI might not have those desires, it might make those judgments
• 26:45
  [SPEAKER_00]: be fully rational, and just not do what it thinks that more or the ought to do.
• 26:50
  [SPEAKER_00]: So it knows the truth about what it ought to do.
• 26:52
  [SPEAKER_00]: It's, and it favors humans incidentally, but it just doesn't do it because it doesn't have desires that pull in that direction.
• 26:57
  [SPEAKER_01]: And it might be that it's like depending on how we want to use the word rational, that it's not fully rational, because that connection is broken, but that.
• 27:05
  [SPEAKER_01]: The sorts of things that we would do to create an artificial system that, again, is really good at playing chess and really good at solving math problems and really good at writing essays and really good at doing all these other things.
• 27:19
  [SPEAKER_01]: It might be that kind of training just doesn't get you this.
• 27:21
  [SPEAKER_01]: This is as it were in terms of what it would take to realize it in an artificial system just a completely different phenomenon.
• 27:28
  [SPEAKER_00]: Yeah.
• 27:30
  [SPEAKER_00]: Okay, summarize that we've got these four assumptions.
• 27:32
  [SPEAKER_00]: One is that there are ethical facts.
• 27:35
  [SPEAKER_00]: The second one is that those ethical facts are relativized to us.
• 27:38
  [SPEAKER_00]: The third is that whatever the facts, those non-relative values ethical facts are, they are actually noble by an AI, in the sense that it has all the records that part that may not just be intellect, but it has emotion if that's what's required.
• 27:51
  [SPEAKER_00]: And then the fourth one is that it also has to be to some extent motivated something like by virtue of making those judgments that this is right in this is wrong.
• 28:00
  [SPEAKER_00]: And those are obviously, as you and I know, from knowing a little bit of history philosophy, those are all phenomenally controversial claims to make.
• 28:09
  [SPEAKER_01]: Yeah.
• 28:10
  [SPEAKER_01]: And I have to say, like, I'm like, guardedly optimistic about every single one of those, but, you know, when you think about the, when you think about risks, you got to multiply through.
• 28:19
  [SPEAKER_01]: So every one of those has to come out right.
• 28:21
  [SPEAKER_01]: And or if things to go, go the way this strategy wants to suggest that we're going to mitigate
• 28:29
  [SPEAKER_00]: Yeah, so the next thing I want to get the next thing I want to hear about human flooring to take it that if the moral facts are such that all those conditions are satisfied and turns out that there's a moral fact laying around that we didn't know about which is that the superior being should always crush the inferior intellectual, intellectual, intellectuals or something along those lines.
• 28:47
  [SPEAKER_00]: number one, we can forget about vegetarianism.
• 28:50
  [SPEAKER_00]: That's the good news.
• 28:51
  [SPEAKER_00]: The bad news is that the AI could just destroy us.
• 28:55
  [SPEAKER_00]: It's all morally on the up and up.
• 28:57
  [SPEAKER_01]: So yeah, I'm going to talk about this.
• 29:00
  [SPEAKER_01]: And one of the reasons why I like to talk about this separately is I actually think that how we should react to these conclusions.
• 29:07
  [SPEAKER_01]: I think is interestingly different than how we should react to the kind of hypotheses I've just been talking about.
• 29:13
  [SPEAKER_01]: And we can think about the set of questions we've been talking about just already as being questions about metathically are things set up such as the strategy can can succeed, right?
• 29:23
  [SPEAKER_01]: Are is the nature of moral motivation and the more facts and moral inquiry?
• 29:27
  [SPEAKER_01]: Are those things set up to the strategy can succeed, right?
• 29:30
  [SPEAKER_00]: Yeah.
• 29:30
  [SPEAKER_01]: So then there's this sort of separate question about if there are ethical facts, are the substance of those ethical facts such that this strategy is a strategy that's going to be friendly to us?
• 29:39
  [SPEAKER_01]: Yep.
• 29:39
  [SPEAKER_01]: And so one of the things I can talk about is I think, again, one of the really interesting things is that there are all these strands in contemporary ethical theorizing that might make you wonder about that pretty seriously, right?
• 29:53
  [SPEAKER_01]: The might make you worry about that.
• 29:55
  [SPEAKER_00]: Yeah.
• 29:56
  [SPEAKER_01]: And so maybe I'll just talk about a couple of those things.
• 29:58
  [SPEAKER_01]: And because I think there are different and interesting ways.
• 30:01
  [SPEAKER_01]: So one connects very much to the example that you already gave.
• 30:05
  [SPEAKER_01]: So there's a
• 30:09
  [SPEAKER_01]: In contemporary ethics about ideas that people sometimes talk about the idea of moral status or moral concerability.
• 30:17
  [SPEAKER_01]: And the core idea of moral status or moral concerability is something like this.
• 30:21
  [SPEAKER_01]: If I accidentally are on purpose, destroy your treasured main vase.
• 30:27
  [SPEAKER_01]: seems like I've done something I shouldn't have done probably, but it seems like if I've wronged anybody have wronged you.
• 30:33
  [SPEAKER_01]: It's the fact that it's your vase, it's the problem.
• 30:37
  [SPEAKER_01]: The vase itself is harmed.
• 30:39
  [SPEAKER_01]: We shouldn't be intrinsically concerned about the well-being of the vase, you might think.
• 30:44
  [SPEAKER_01]: Whereas, on the other hand, if I torture or in six significantly injure your cat, that's a pretty different case.
• 30:53
  [SPEAKER_01]: And one way to bring out the difference is that if you're like, Trister, I'm going ahead to destroy my vase.
• 30:58
  [SPEAKER_01]: I've hated that vase ever since I was a child.
• 31:00
  [SPEAKER_01]: I've never had the nerve to destroy it.
• 31:02
  [SPEAKER_01]: Well, maybe it's fine for me to destroy your vase, right?
• 31:04
  [SPEAKER_01]: But on the other hand, if you're like, Trister, I'm going ahead to destroy my cat.
• 31:07
  [SPEAKER_01]: They're like, I've hated that cat since I got it.
• 31:09
  [SPEAKER_01]: You might think, well, wait, there's still
• 31:13
  [SPEAKER_01]: And one thing that's a characteristic of a lot of mainstream thinking and ethics, I would say, is the thought that there's a special kind of moral status that human beings have, that we matter in some important way more than, say, insects, but also that if you got, for example, a super intelligent AI, it wouldn't suddenly matter way more than we do in the relevant sense.
• 31:42
  [SPEAKER_01]: Right.
• 31:43
  [SPEAKER_01]: If on the one hand, it turns out that like just biological organisms, they all count and their lives all matter in some way that sort of, you might count them up, equivalently.
• 31:53
  [SPEAKER_01]: An AI that's sensitive to that pattern is going to make things really bad for human beings.
• 31:58
  [SPEAKER_01]: One of my favorite examples is runaway climate change.
• 32:01
  [SPEAKER_01]: Going to be very bad and aggregate for human beings.
• 32:04
  [SPEAKER_01]: But it's going to be great for insects.
• 32:07
  [SPEAKER_01]: like you're going to get like more sort of tropical environments, you're going to get like more mosquitoes, right?
• 32:13
  [SPEAKER_01]: And like maybe like trillions and trillions more mosquitoes, like that's just a net gain even as the human being suffer and show life choices, right?
• 32:21
  [SPEAKER_01]: And so if things go that way, it could be a problem if the AI were sensitive to those values.
• 32:25
  [SPEAKER_01]: And on the other hand, of course, if the AI is just like all that really matters of the super intelligent things, then again, it might be treating us the way that we treat the mosquitoes,
• 32:35
  [SPEAKER_01]: just went for a hike yesterday.
• 32:36
  [SPEAKER_01]: I'd probably killed like 25 mosquitoes and they were they were annoying blood sectors.
• 32:42
  [SPEAKER_01]: So yeah, so this is like one example of I think there are a bunch of other interesting kinds of examples, but that's one vivid example of the way that ethical theory kind of has to work out just right for this strategy to to be again a strategy that's going to be one that's going to make things go well for us.
• 33:03
  [SPEAKER_00]: Right.
• 33:03
  [SPEAKER_00]: So the ethical theory is going to have to be something like, once you've hit some minimal threshold of x, whatever the hell x is, whatever that, whatever, but x is whatever conforms that dignity or worth or considerability upon us, right?
• 33:16
  [SPEAKER_00]: So sentience, rational reflection, whatever it is, something on those lines, once the view has to be, once you've hit some certain minimal threshold, now you've got
• 33:27
  [SPEAKER_00]: more with their dignity, whatever such that you can't just be smashed like a vase of no-one else cares, something like that.
• 33:33
  [SPEAKER_00]: And it has to be that once things meet that minimal threshold, everything's equal, because if it's not all equal, and it's sort of like the more X you've got, the more sanctions, the more rationality, or whatever it is, the more dignity you have, which gives us, if you like permission, maybe, it'll give us permission to kill insects, let's say, for sure.
• 33:53
  [SPEAKER_00]: Maybe, I don't think you think this, but maybe eat some animals,
• 33:57
  [SPEAKER_00]: depending on whether or not they were, they are in the threshold.
• 33:59
  [SPEAKER_00]: But if AI is to us as we are to mosquitoes, and it's got way more of x, or x, or y, and z, more rationality, more sensience, more all the things, then it's going to be able to treat us like mosquitoes, and we don't want ethics to look like that.
• 34:12
  [SPEAKER_00]: So the ethical facts better be such that there's this view according to which a buffer certain threshold everything is equal.
• 34:18
  [SPEAKER_00]: Something like that, is that right?
• 34:19
  [SPEAKER_00]: Right, exactly.
• 34:20
  [SPEAKER_01]: What are things I just want to say that I think is a different between this hypothesis on the one hand, and the sort of medical hypotheses I was talking about before, is that I think it's possible to have interestingly different reactions to these hypotheses.
• 34:35
  [SPEAKER_01]: One of the things about these hypotheses, I think it's possible to think, look,
• 34:41
  [SPEAKER_01]: If it turns out that the ethical facts are that the mosquitoes matter as much as we do, then it's going to be great ethical progress when the AI creates the amazingly tropical world for the insects.
• 34:54
  [SPEAKER_01]: Or if it turns out that the
• 34:56
  [SPEAKER_01]: The AI is just significantly greater degrees of the things that make an entity matter, and those degrees matter for how much an entity matters, such that, oh, like all that really it turns out matters.
• 35:09
  [SPEAKER_01]: But those circumstances are the, how the AI themselves fair, then it's a possible, however, perspective.
• 35:15
  [SPEAKER_01]: It's going to suck for us, but that's just, sometimes, like, a lot of, like, equals.
• 35:20
  [SPEAKER_01]: Sometimes life doesn't pan out, but we should be glad if ethics.
• 35:24
  [SPEAKER_01]: If the AI is actually sensitive to those ethical facts because that's like they're they're they are acting for the best or something like that right and you do sometimes get with these kinds of reactions.
• 35:33
  [SPEAKER_01]: I think they can make takemark as a book in which he.
• 35:38
  [SPEAKER_01]: at one point talks to Larry Page, I think, from Google.
• 35:41
  [SPEAKER_01]: And just like, pages just look, like AI are just going to be the next step in cosmic evolution.
• 35:46
  [SPEAKER_01]: And if we're like trying to keep them down for our own interests, like we are just, we're just being like unjust or something like that.
• 35:52
  [SPEAKER_01]: And so it's possible to embrace that kind of perspective.
• 35:55
  [SPEAKER_01]: But I think it's awesome.
• 35:55
  [SPEAKER_01]: super interesting, but it might be very hard to live that perspective, and it creates this possibility of a kind of like intelligible alienation from ethics, that like, if it turns out that ethics is just like bad for you as a human being under these circumstances, you might start getting very alienated from it.
• 36:13
  [SPEAKER_01]: And I think those, that sort of raises a bunch of really fascinating questions about like, you know, how would we navigate that so that space?
• 36:22
  [SPEAKER_01]: And what
• 36:25
  [SPEAKER_00]: Yes, that's all fair and super fascinating my own reaction is to be like all right I don't Partly wants to say something like this.
• 36:35
  [SPEAKER_00]: We can't really be talking about ethics if we're talking about It could be the case that we should all just be wiped out
• 36:42
  [SPEAKER_00]: Maybe that entails that there are aren't ethical facts of the sort that we were envisioning or ethics is, is relativized to humans or something.
• 36:50
  [SPEAKER_00]: I don't know, I actually, I don't know what my reaction is.
• 36:53
  [SPEAKER_00]: I don't, I know my reaction is something like a porance, but beyond that, I'm not particularly ready to articulate what my philosophical reaction to that is or what the implications
• 37:03
  [SPEAKER_00]: One, you, of course, you could say the importance is intelligible, but unjustified.
• 37:07
  [SPEAKER_00]: Okay, fine.
• 37:09
  [SPEAKER_00]: Because our importance to moral possibilities is not.
• 37:14
  [SPEAKER_00]: Has that historically been a great guide to what's actually morally right in our own?
• 37:17
  [SPEAKER_00]: On the other hand, if it is at least reasonable, I'm not sure what justifies the reasonableness of the importance.
• 37:23
  [SPEAKER_00]: And so I guess perhaps I'm standing right now.
• 37:26
  [SPEAKER_01]: Yeah, yeah, no.
• 37:27
  [SPEAKER_01]: I think like one, I think very,
• 37:30
  [SPEAKER_01]: initially plausible sounding reaction to this is that look, we know that the ethics couldn't be like this, and we know that what ethics couldn't be unfriendly to humans.
• 37:41
  [SPEAKER_01]: There's a famous kind of slogan that various philosophers in the 20th century have offered.
• 37:46
  [SPEAKER_01]: Interestingly, they tend to like mean different things by it.
• 37:49
  [SPEAKER_01]: Like one famous example, this is Peter Realtham, which has this sort of thought.
• 37:53
  [SPEAKER_01]: But look, ethics is made for us.
• 37:55
  [SPEAKER_01]: It's not that we are right.
• 37:59
  [SPEAKER_01]: One thing, again, that's super fascinating for me, thinking to think about as an ethical theorist, is that if you take that idea seriously, that ethics is made for us, I think that it actually creates a wedge to have a different kind of worry, because if ethics is made for us, it starts to make it sound like ethics.
• 38:17
  [SPEAKER_01]: The thing that we're talking about, we're talking about ethics is in some sense kind of parochial.
• 38:21
  [SPEAKER_01]: It's about human beings and stuff that's good for human beings, right?
• 38:26
  [SPEAKER_01]: Sure.
• 38:26
  [SPEAKER_01]: And if that's the way ethics is, then I think there's a legitimate worry that when the AI starts using its incredible intelligence to think about how to live and what sort of decisions to make and what sorts of values to have.
• 38:39
  [SPEAKER_01]: it's going to conclude ethics, semethics.
• 38:42
  [SPEAKER_01]: That's just this perocule thing in the same way that you might think that if somehow ethics, it was somehow coded into ethics, that it was misogynist and it was like, yeah, ethics is just a different man.
• 38:52
  [SPEAKER_01]: You're like, women started thinking about that.
• 38:53
  [SPEAKER_01]: It might be like, ethics, like, that's like nonsense, right?
• 38:56
  [SPEAKER_01]: Yeah.
• 38:56
  [SPEAKER_01]: And so I do think that there's this kind of worry that the more you lean on this idea that, no, what we're thinking about ethics, it has to be this kind of thing that's made for us.
• 39:05
  [SPEAKER_01]: the more I think that it raises the specter that why is the AI going to be especially interested in that when it's trying to think about how it's live, yeah, make decisions.
• 39:15
  [SPEAKER_00]: It's funny.
• 39:15
  [SPEAKER_00]: I actually had similar thoughts while watching the latest installment of Planet of the Apes where there's this choice to be made between the age and the humans, their assholes and their apes that are better and they accuse the human of being a trader if they're going to help the apes,
• 39:31
  [SPEAKER_00]: But the humans are ourselves and the apes are trying to aim for something decent here that she have special obligations to humans.
• 39:38
  [SPEAKER_00]: Our ethical requirements to promote human flourishing as over against ape flourishing.
• 39:43
  [SPEAKER_00]: Well, that, why?
• 39:44
  [SPEAKER_00]: Wouldn't, don't we care more about the content of the character?
• 39:48
  [SPEAKER_00]: We should regardless of the packaging.
• 39:50
  [SPEAKER_00]: Right.
• 39:51
  [SPEAKER_00]: Anyway, similar, not the same, but at least similar kinds of issues about how parochial is ethics.
• 39:56
  [SPEAKER_00]: Does ethics tell us?
• 39:58
  [SPEAKER_00]: You got a favor human beings over the apes, is that definitely what we should be going for?
• 40:02
  [SPEAKER_00]: Or is it just as well if the apes went because their vehicle worth equal dignity and this scenario they have equally rich lives as human beings, etc.
• 40:09
  [SPEAKER_00]: etc.
• 40:11
  [SPEAKER_00]: Yeah, I don't know, so I feel the pull of the, I feel the pull of the charger parochialism, but I don't know what to do with it.
• 40:17
  [SPEAKER_01]: good.
• 40:17
  [SPEAKER_01]: So I noticed we only have a couple of minutes left.
• 40:19
  [SPEAKER_01]: So I just want to say one final thing about this, which I think one of the things that I think is rich in thinking about these questions.
• 40:26
  [SPEAKER_01]: That probably some of your listeners will very reasonably be like, why are we thinking about these really outlanders hypothesis?
• 40:32
  [SPEAKER_01]: This seems a bit crazy.
• 40:34
  [SPEAKER_01]: But one of the things that I want to say is that there's this enormously pressing practical project right now to figure out how to align
• 40:44
  [SPEAKER_01]: And there's this deep question like aligned with what?
• 40:48
  [SPEAKER_01]: And it's very natural to think, oh, we should just be aligning with, and then you start getting some people are just like things that are good for people.
• 40:57
  [SPEAKER_01]: And some people are like, no, we should like align with the ethical values.
• 41:00
  [SPEAKER_01]: And often it feels like these things are being run together.
• 41:02
  [SPEAKER_01]: And one thing that I think this kind of brings out is these are really different choice points.
• 41:07
  [SPEAKER_01]: We, if we're trying to think about like even what the problem of alignment is, there's really hard potentially backed questions about is it okay, like thinking of your plan to be a example, is it okay to make it such that like humans get a special pass when we're aligning these systems or yeah, should we just like really try to maybe us and the AI or just the AI figure out what the ethical facts are if there are any and just let the chips fall where they may and I think like those are really deep interesting questions.
• 41:36
  [SPEAKER_00]: Yeah, no, they are.
• 41:37
  [SPEAKER_00]: I'm going to have to insist that you come back.
• 41:39
  [SPEAKER_00]: We'll talk about this a more.
• 41:40
  [SPEAKER_00]: I feel like we should get a bunch of other philosophers in the mix and have a big debate about how much to humans really matter.
• 41:47
  [SPEAKER_00]: Yeah, we'll be fun.
• 41:48
  [SPEAKER_00]: Yeah, this has been great.
• 41:50
  [SPEAKER_00]: Here's some thanks so much.
• 41:50
  [SPEAKER_00]: Yeah, I hope this is uh, there's more to come.
• 41:53
  [SPEAKER_01]: All right, let's be in touch and yeah, I hope some of your some of your listeners got something out of this Really not how could they not if if not they should stop listening listeners.
• 42:01
  [SPEAKER_00]: Stop listening.
• 42:02
  [SPEAKER_00]: No, don't you that don't do that All right, thank you.
• 42:05
  [SPEAKER_00]: All right, take care.
• 42:06
  [SPEAKER_00]: All right, bye
• 42:11
  [SPEAKER_00]: Thanks so much everyone for listening and now let me think that people have made this podcast possible.
• 42:15
  [SPEAKER_00]: So, first of all, we have the Lotto Alexander, he's been with me since day one.
• 42:20
  [SPEAKER_00]: He showed me even how to start recording a podcast in the first place he does editing stuff.
• 42:24
  [SPEAKER_00]: He's been fantastic and he does really everything I ask him to and more.
• 42:29
  [SPEAKER_00]: I've never been anyone else who says so much, yeah, I don't know how to do that, but I'll figure it out.
• 42:33
  [SPEAKER_00]: What more could you ask for in a producer?
• 42:34
  [SPEAKER_00]: Next up, Jennifer Strong, my podcast guru, I don't know what else to call her.
• 42:39
  [SPEAKER_00]: She just advised me on how to design a podcast, how to grow it, and she's responsible for if you like this kind of season two relaunch.
• 42:47
  [SPEAKER_00]: And everything around it, so thank you to you, Jennifer.
• 42:50
  [SPEAKER_00]: Next up, Meg Marco, responsible for all the excellent artwork.
• 42:53
  [SPEAKER_00]: Before I did the artwork, I didn't really know what I was doing.
• 42:56
  [SPEAKER_00]: We put it together, it worked well enough.
• 42:58
  [SPEAKER_00]: But I think Meg really did a phenomenal job bringing everything together.
• 43:02
  [SPEAKER_00]: Garrett Lane, sound engineer, that I frankly don't deserve.
• 43:05
  [SPEAKER_00]: He's just too good for this podcast.
• 43:08
  [SPEAKER_00]: Don't tell him I said that though.
• 43:09
  [SPEAKER_00]: Next up, Benjamin Coltgen, he's a professor of philosophy who has helped me source guests identify who's interesting, who has things to say, etc.
• 43:18
  [SPEAKER_00]: etc.
• 43:18
  [SPEAKER_00]: So, without him, I would have anyone to talk to.
• 43:20
  [SPEAKER_00]: And then finally, Ashley Russell Lanko, who helps with...
• 43:23
  [SPEAKER_00]: All manner of administrative things making sure that the podcast guests find their way to the podcast and schedule the recording, get their bios and headshots and really just keeping the show on the rails and me organized.
• 43:35
  [SPEAKER_00]: Thanks to all of you, really appreciate all your hard work.