Become a Creator today!Start creating today - Share your story with the world!
Start for free
00:00:00
00:00:01
#14 Alice Helliwell: The Art of Misalignment
28 Plays2 months ago
What if the best AI art doesn’t care what we think? In this episode, we talk with philosopher Alice Helliwell about her provocative idea: that future AI might create aesthetic value not by mimicking human tastes, but by challenging them. Drawing from her 2024 article “Aesthetic Value and the AI Alignment Problem,” we explore why perfect alignment isn't always ideal—and how a little artistic misalignment could open new creative frontiers.
For more info, visit ethicscircle.org.
Recommended
Transcript
Introduction to Dr. Alice Helliwell
00:00:16
Speaker
Today we're speaking with Dr. Alice Helliwell, Assistant Professor of Philosophy at Northeastern University of London.
AI Art and Philosophy - An Overview
00:00:22
Speaker
Her research explores the philosophical implications of ai art and computational creativity.
00:00:28
Speaker
We'll be discussing her new article, Aesthetic Value and the AI Alignment Problem. Thanks for joining us, Alice. Thanks for having me. It's great to be here.
Dr. Helliwell's Academic Journey
00:00:37
Speaker
Yeah, absolutely. um Yeah, just kind of by way of introduction, can you just tell us a little bit about your background, you know, kind of where you're from and maybe how you ended up in philosophy?
00:00:47
Speaker
Okay. yeah and Yeah. So, I mean, where I'm from, i guess, location-based, I'm from the UK. And and I originally went into to philosophy in my undergraduate degree. and So I did my undergraduate degree at the University of Edinburgh. and with i was studying philosophy and psychology, in fact. and So I did a sort of joint honours degree. so that was where my initial ah sort of interest was. and But then I kind of diverged from that. So I actually, for a brief while, went into the psychology, and sort of organisational psychology,
00:01:16
Speaker
and before sort realizing that wasn't for me. and And I wanted to go back in and to work in something and more arts related. And so that took me to doing a master's and in the history and philosophy of art.
00:01:29
Speaker
And during that time, I started working on AI and AI art and back in, this was back in 2017, 2018. So quite so ah quite well, not the earliest to think about AI art by any means, but certainly earlier than perhaps some might might think given sort of more recent advances in things like Dali and stable diffusion.
00:01:48
Speaker
and And so, yeah, so that that was kind of where my interest sort of grew, particularly in this field um relating to philosophy. um ah Because there's just, for me, there's so many interesting questions that we could kind of um look into and try and answer and philosophically with this kind of new technology.
AI Creativity and Ethical Questions
00:02:05
Speaker
That is a great segue ah to my question because, all right, so I have a background philosophy. I have a background also in computer science.
00:02:16
Speaker
And so I know the AI. I know the philosophy stuff. I am woefully inadequate when it comes to any knowledge whatsoever about art. I'm currently teaching a linked course, as a matter of fact, between an ancient philosophy ancient philosophy and also art history. And I'm realizing that undergraduates know more than I do about play history. So ah you're going to teach me a lot today. We'll see. can you Can you give us like an overview of like what this intersection of ai of AI, art, and philosophy looks like? Just some of the questions that are covered there.
00:02:51
Speaker
Gosh, there's so many things and there's there's more coming out all the time and relating to this topic, right? So we have that kind of typical questions, which I find interesting about, you know, ah whether AI can make art either autonomousy autonomously or whether it could be used to make art, which I think we kind of see that it it can in some way now. and Whether AI can be creative. and We have questions about how we should attribute responsibility for AI works or, you know, where credit should lie for some of these new works.
00:03:16
Speaker
We have... we have ah kind of ethical questions about AI arts in the arts as well, um and um you know kind of how AI might be used in creative practices, and what that might mean for practitioners um in their art-making processes.
00:03:32
Speaker
um i'm I'm sure there's more, and right? I guess something as well, you know, if there's anything and particularly unique to images, say, or music or whatever that's made using AI or whether there's something sort of lacking from that, and whether there's something lacking from processes that appear to generate sort of novel products that mean that it's not creative,
00:03:52
Speaker
A whole array of questions. It's a bit of scattergun of some other questions. but um But yeah, I mean, and to me, all of these are fascinating. and So and so hence my hence my continued research area.
00:04:03
Speaker
Right. And we'll get kind of get into your specific question in a moment. I just wanted to ask, like, could you kind of give us a sense of what's currently happening with AI
Technical Advancements in AI Art
00:04:13
Speaker
and art creation? I think at least for me, like what I think of as like Dolly and Midjourney and I guess OpenAI, like making images.
00:04:25
Speaker
But yeah, what what's what's kind of happening with AI and art creation at this current point, like in ah on a technical level, in the sense of like, yeah, what's interesting, what's going on. No, I don't mean technical, like computer science level. I'm just being like, what's- can't tell you that. Yeah, yeah sorry. i did yeah i sort of like I hate to to try and say just because I know that with things like DALI, I was busy working on generative adversarial networks for a long time.
00:04:51
Speaker
and And then suddenly these kind of diffusion models sort of appeared, things like DALI, Mid-Genie and Stable Diffusion, right? And and they were a lot more capable than we kind of thought would be coming anytime soon. So I'm i'm sort of loathe to try and guess what what computer scientists are working on now because I'm sure it's something very impressive.
00:05:08
Speaker
and But I think these kinds of tools, that they certainly have been getting better since they first launched in sort of 2022, I think. and um And so we're seeing a sort of big progression in this sort of text to image generators. There also seems to be a lot more progress made on sort of video generators and with AIs making fairly impressive um videos from prompts, and but I mean, again, that kind of shocks me each time I'm shocked with the progress that people are making with these kinds of new technologies.
00:05:38
Speaker
um And at the minute, it does seem to be that it's kind of just ramping up the abilities of these existing technologies. But I'm i'm not, i'm I'm sure there's something going on that I'm not aware it So is it my mainly like it seems like the visual arts are the ones that are most advanced, whereas is there nothing really that interesting going on with music? Like i is AI being applied very to for success in music? I don't know.
00:06:02
Speaker
I mean, it certainly it certainly is being used in music. I will say i work in mostly in the visual arts and that's my background. So hence, I'm a little bit more tuned into that kind of thing. and i know for... there's already been debates around things like the Oscars, right, in film, about the use of AI systems in film. and So that's also kind of going on. So I imagine there's similar sorts of things bubbling away in music.
00:06:24
Speaker
do know some people that work on that area, but it's just not something that I focus on as much. and but and But yeah, what's on? I was just going to say, I imagine behind the scenes, people are, you know, artists are, you know, ah sorry, like, um what's the word?
00:06:41
Speaker
The artists using the written word, novelists, poets, I'm sure behind the scenes are integrating AI into their workflow to some degree, but it's unclear. Yeah. i mean I mean, I imagine anyway.
00:06:53
Speaker
Yeah. So I'm not clear about that. Actually, it's quite interesting. I mean, at least in the visual arts, there's some really interesting stuff happening with artists using AI in their practices. potentially sort of they're seeing it some sort of collaboration or whether they're just seeing it as something that they use and in order to even draw attention to the AI system and kind of what it can and can't do and so this I'm sure there are poets and you know people who work with the spoken the written or spoken word that are doing some of those practices whether they're integrating it into their workflows sort of more generally I'm not sure I don't I don't
00:07:25
Speaker
I certainly know that there are people who do things like marketing, kind of other creative industries that are using these kind of tools in their workflows. But yeah, i I haven't spoken to any novelists who are doing that, so I wouldn't be able to to say, but and potentially, right?
00:07:39
Speaker
um Yeah. Cool. Well, maybe we can now shift into um kind of your question, which is basically the value alignment problem, but applied to
Understanding the AI Alignment Problem
00:07:51
Speaker
aesthetics. So um Yeah. Can you kind of just introduce to us what the value alignment problem is Yeah. and So I guess the value alignment problem is just this sort of this concern and that we are going to develop AI systems that don't mesh with our values.
00:08:10
Speaker
and So the, I mean, I'm sure people will have heard of, and you know, Nick Bostrom's example of the clay paperclip maximizer, which seems to be a pretty classic example of a system which, you know, is given a ah job to do to to make paperclips.
00:08:23
Speaker
um And it kind of gets sort of hyper-focused on making paperclips. That's its sort of one value, that it sort of values paperclips. And so um everything else is kind of... um below that value, right? It's gonna kill people, it's gonna use various things for energy, all in service of maximizing the amount paper clips it has has made, right? And so the idea is that we could have a system that sort of um doesn't value the same things we do um and has potentially goals in mind relating to that value, like maximizing paper clips, that cause problems for us, like you know using create getting as much en energy as it can or or whatever it is, right?
00:09:03
Speaker
and So we have this concern that if we if we had a system that was sort of super capable in some way, that it could run rampant over the world um and, you know, kill us all, right? That's the existential risk that people often talk about and from these kind of systems.
00:09:18
Speaker
and So one potential sort of solution to that is and this idea that we just need to make sure that the AI that we develop shares our values, either shares our goals or shares our values in some way. And that's going to sort of guard against that kind of risk that an AI will sort of run rampant with values that and aren't the same as ours and thus cause us some sort of threat.
00:09:39
Speaker
and Yeah, it's a rough summary. Yeah. but That raises all sorts of questions. that I'm sure some of them we'll get into. um One of them is like, what is our values? Like how do what is how do we define in quotation marks our values? But before we get into that, you mentioned Bostrom and I believe the title of his book is super intelligence,
Moral vs Aesthetic Values in AI
00:09:58
Speaker
right? So this problem is usually, you know, spoken about in terms of a What if we make a super intelligent AI and it's totally capable of harming us because it's not aligned with our values and we don't even know it because it's so smart that it's beyond us?
00:10:14
Speaker
ah want to present a ah different question to you just to get your thoughts on this. So um many of us are concerned, sure, about super intelligent AI. Some of us are just concerned about regular dumb AI or non-super intelligent AI, right?
00:10:27
Speaker
So before we hit record, we were talking about how Basically, a chat GPT and other large language models, which are not super intelligent, as far as I know, are already causing some, shall we say, instability in the classroom.
00:10:45
Speaker
um So what are your thoughts on narrow AI ah and and the value alignment problem? Is it as relevant or is it even maybe more relevant? What do you think?
00:10:56
Speaker
Yeah, so it's it's not something that I've looked into greatly, but I think there is a relevant sort of value alignment issue there, right? Exactly as you say, there can be times, you know, maybe I want something like Dali to align with my aesthetic values too.
00:11:09
Speaker
When I'm using it, I want it to kind of and create something which meshes which with what I think is valuable. So there there can be an and and extent to which we are going to be concerned about values in AI in general.
00:11:22
Speaker
and So i think I think the concern is often directed to these kind of extreme cases because we have this fear that those AI will be able to just kind of and completely go outside of our control. So with DALI, with ChatGPT, it's doing something that I don't like. I'm just not going to use it or I'm going to switch it off, right?
00:11:41
Speaker
The concern is that and down the line, you won't be able to have that option anymore. um Hence why it kind of gets conflated with this huge existential risk thing. um so So, yeah, I guess they're still broadly within our control. We still have the option to sort of turn them off or whatever it is.
00:11:57
Speaker
um But there still may be value alignment concerns. And people do look at those, right, in research. They are concerned with that in general. um and And you kind of raise that issue of which values, how do we sort of and put those into an AI? How do we articulate them in the first place?
00:12:12
Speaker
And those are certainly questions that we could have around narrow AI as well. Yeah. Mm-hmm. Mm-hmm.
00:12:21
Speaker
Okay, so i guess you know when it comes to thinking about the super intelligent AI, and ah like something like i always think of Hal in 2001. And so that would definitely be an example of not aligned because he doesn't really value human life very much. like He'll kill people on the mission.
00:12:43
Speaker
And then he also lies to a lot of people. Anyway, so that, that he like, he doesn't value honesty. He doesn't seem to value human life anyway. So he seems misaligned and they're like the upshot is people are dying and stuff.
00:12:55
Speaker
So what about, you know, what if someone's like, well, that's really what's important is more alignment. um But yeah, you're talking about aesthetic alignment. So anyway, could you kind of just motivate like thinking about the issue of aesthetic alignment? Why should we be also thinking about that?
00:13:14
Speaker
that. Yeah. So, I mean, I, I don't necessarily think that aesthetic value alignment should, is quite the same level of problem as something like risk. That's not my position that I'm trying to defend, but it doesn't mean that aesthetic value is sort of trivial and it doesn't mean it's trivial in our lives. Right. So it certainly doesn't mean there's nothing to care about there.
00:13:34
Speaker
And I guess that's what, um, Kind of part of what I'm putting forward and in the paper is and the idea that when we talk about aligning with human values, yes, we're talking about ethical values, moral values, right? But we also might be talking about other kinds of human values, like aesthetic values, and how we decide to sort of, if we manage to solve the value alignment problem, the kinds of AI that we develop are going to have impacts in these other sort of normative domains, right? So things like aesthetics, things like arts.
00:14:03
Speaker
um And I think that, you know, if you're a sort of person that cares about and aesthetic value, who cares about art, cares about kind of aesthetic experiences in some way, then you might care about those that are going to be sort of brought about by AI in the future.
00:14:17
Speaker
and and And so that's kind of what's what's sort of motivating this in some way. It's kind of trying to broaden out the discussion of value and relating to AI um and highlighting the fact that, you know, it's not necessarily always about the moral values, even though they may be very important.
00:14:33
Speaker
There's also a sense in which the two can interact in some ways, right? So as I kind of talk about, if we have a system that is potentially creative, that might manifest in ah an aesthetic domain, but later have impacts in an ethical domain. So there might be some sometimes a connection between the two, right? They're not always so clearly delineated that we can kind of talk about one sort of value without talking about the other entirely.
00:14:57
Speaker
yeah Yeah, actually, just jumping on that real quick, like yeah maybe we could just touch on the whole issue of like what what's the difference between a moral and aesthetic value? I mean, i guess intuitively it's just like, well, beauty is like a different, that's a whole other category than ah right and wrong, or I don't know.
00:15:14
Speaker
ah But anyway, i do you want to comment on that? like what's What's the difference between a moral yeah value? I mean, I'm not im i'm not sure if i if I'm going to have a like a really good answer to this.
00:15:26
Speaker
I suppose some people do think that there is just the good, right? There is just sort of things that are valuable and and there's kind of no distinction between the aesthetic or the moral, right? and I don't think that's that kind of makes sense to me in some ways because I do think there are times when we might be paying attention to one rather than the other or times in which we could have and something with a positive aesthetic value say that's very beautiful but we know that it's unethical right or we know that there's some sort of negative ethical and component to it um I think one of the
00:15:58
Speaker
One of the things that's often brought up is a bit like the novel Lolita or something like that, right? Where there's kind of this the part of why people think that's a great aesthetic achievement is that it's kind of got some ethical problems with it and yet you kind of appreciate it anyway.
00:16:14
Speaker
So I do think there's ways in which the two might be getting at something different. I don't know that I want to try and define and aesthetic value fully. I mean, as you say, things like you know beauty and also the the sort of negative aesthetic value, things like ugliness.
00:16:28
Speaker
We sometimes also talk about aesthetic values as linked to things like cognitive values, and emotional values. Right. So it ends up being sort of a ah contested term in some way. um But I guess what I'm broadly interested in in the paper is things that are relevant to aesthetic experience, to the arts, and that kind of thing, rather than those things that squarely are concerned with, you know, and value of human life or or whatever that normally comes into those discussions of value alignment.
00:16:56
Speaker
ah So, First of all, by bringing up a Nabokov, right? That's his name. It's ah so many thoughts. I almost wanted to you know jump in and say, oh, can you explain that? But it's too dark for for the listeners.
00:17:09
Speaker
It's just one of those examples that always comes up. That's why I point to it. but and yeah um But yeah, yeah they're they're definitely not necessarily aligned on your view. Also in my view, right? but Okay. So given this preamble, ah I want to give listeners sort of... um you know, ah some mental hooks to hang some concepts on so that as we're going through your argument, and they, you know, they can kind of match ideas to to this general.
00:17:39
Speaker
So if you had to give like an elevator pitch for your ah for your thesis, like, so you know, you're saying, i think what you're saying is we want it to be broadly aligned, but not perfectly aligned or or maybe you can give a ah ah better summary of that and and in in your summary you can you know what what what does it mean to for AI to align with aesthetic values and kind of clarify that for us ah Okay, yeah. So um so what I'm, as you said, kind of broadly interested in the paper is just and the alignment problem which relates to values and expanding that out to considering other kinds of human values like aesthetic
Moderate Value Alignment in Aesthetics
00:18:19
Speaker
value.
00:18:20
Speaker
and So my argument is that the value alignment will apply to things like aesthetic values, and that there will be potential for misalignment and between humans and AI in the aesthetic domain and and that sometimes we might want that and sometimes we won't. right So sometimes we might want an AI system that and doesn't fully match with our values and sometimes we we kind of do, because it particularly when it becomes ethically relevant.
00:18:47
Speaker
and So am sort of arguing in defense of um i think Martin Peterson's, um what he calls the moderate value alignment problem. So he frames it, he kind of has his own argument for why he thinks and that his framing of the problem is correct.
00:19:03
Speaker
um And that view basically says exactly as you said, that and some we want AI to align with some human values, right? That's kind of broadly what it's saying.
00:19:14
Speaker
and So not necessarily all human values all the time, or really all human preferences all the time is often how it ends up being cashed out. and But also we don't want it just to align on the cases where there's a clash.
00:19:27
Speaker
and So it's a particular kind of sort of so trying to show that the aesthetic case can interject in this general debate about how to characterize the value alignment problem, as well as making the case that there sort of is an aesthetic value alignment problem, right? So I've kind of got like ah secondary goal there in some way. Yeah.
00:19:43
Speaker
yeah I don't know if that was a clear explanation, but I tried. That's a long elevator ride, I think. Yeah, sorry, you said elevator pitch. I ignored the elevator pitch. So, I mean, I could try and give you an elevator pitch. No, it's great. It's great, Elsie. I'm sorry. That was, yeah, I completely ignored that you said elevator pitch.
00:20:02
Speaker
no one No one likes being in an elevator, or I don't. Yeah, sorry. so that's fine, yeah. No, that that was a good, lot good ah anyway. Long intro. Long summary. yeah No, that's great. I mean, so, yeah, the Martin Peterson thing is interesting. So he's like,
00:20:17
Speaker
he's He's in the context. So it's almost like you're kind of doing something similar to Martin Peterson, but in the aesthetic context. Because he's he's in the ethical context and being like, hey, you actually don't want full alignment. And and what is his reason is his reason? I think you kind of talk about it but isn't it kind of like, well, humans, like we don't want... Because we're kind of mistaken about a bunch of stuff. So if if we made an ai that was...
00:20:42
Speaker
you know see that would have been an interesting better version of 2001. Because Hal is clearly missing some crucial value, which everyone values, which is honesty and human life. right But I guess the kind of thing Martin Peterson is worried about is like, well, what if it's fully aligned with us?
00:21:06
Speaker
And then we realize, oh, wait a second, like we are mistaken about something. Right. Right. So it's not as though Hal is like, it's not like Hal is like, look, i'm just I'm just being who you want me to be. like I'm just fully aligned with your values. And then it leads to chaos. No, it's like he's clearly missing a basic value, which is human life.
Innovation in Art and AI's Role
00:21:28
Speaker
Anyway, so it seems like Martin Peterson is more worried about a situation like, i don't know, if if we had made AI when we were when when slavery was still... you know you know up and running in the you know in the western you know and that Western countries. like that That's the kind of situation probably he's worried about. right I don't know.
00:21:47
Speaker
Anyway, do you want to Yeah, yeah i think that I think that potentially is a kind of situation. and i think one that he points to in the paper, if I remember rightly, is is ah things to do with the environment, right? So sometimes what we think will benefit us and is not the same as what will actually be good, say, for something.
00:22:03
Speaker
and So if we we might value and sort of ah what's what's good for humanity in general, but actually you know what's good for humanity might be like Thanos, right? clicking, you know, snapping his fingers and getting rid of half of the population. Right. So yeah and okay it's kind of complicated and sort of relationships between how we express our preferences versus what's actually good. And so, um you know, that's kind of, i'm I'm doing bad job at explaining this, and but he's kind of interested in these sorts of cases. Right. um So one, just a general one that might come up is something like often humans want to have a nice life, um which might involve having lots of money.
00:22:43
Speaker
Right. and But in general, and sort of having an AI system that's aligned with and giving you lots of money might have lots of unintended negative consequences.
00:22:54
Speaker
So there's that sort of case as well. um I'm sorry, that was a little bit of a bad explanation, I think, of Peterson's own case. and ah But yeah, these... we Yeah, the idea is like, it's so it's almost like, yeah, I mean, we don't you don't always want to get what you...
00:23:09
Speaker
It's not always good for you to get what you wish for in a way? yeah Is that that sort of thing? were anyway i'm just Yeah, exactly. It's not always good for you to get what you you what you think you want. and and Sometimes we're bad judges of what is good for us. and Sometimes we're bad judges of what's good and for us as a whole, say, as humanity as a whole, right? and There can be sort of ways in which ah we are not necessarily reliable sources of our own sense of value. And that seems like pretty clear in the aesthetic case, because it's like there's all these cases that are super famous where something you know that we now think is awesome comes out and everyone hates it. Like Stravinsky's The Rite of Spring. like Didn't that cause a riot or something? and
00:23:54
Speaker
And then... I don't know. ah Like, impressionism. Everyone thinks Van Gogh i mean you know is the man now, but didn't they not really like Starry Night? so um Anyway.
00:24:08
Speaker
so There's a bunch of cases where it's like so obvious that we overall were mistaken. like are what we thought was good art. We all thought that... It was bad art. Yeah, exactly. We all thought van Gogh was bad art. And then it looks like we revised and learned that it was good.
00:24:29
Speaker
Or at least that's what I would think. um Yeah. And so it's really clear there that like... It seems like you don't, if we made an AI that was just like, oh, I just exactly imitate whatever you guys like at this current stage, that would not be good because looks like we get it wrong often.
00:24:48
Speaker
Sam, your impression of super intelligent AI is getting better and better. i think I used to want to say you need some work on it, but I think it's pretty good. That's how I imagine it for sure.
00:25:00
Speaker
Yeah, i'm'm I'm kind of glad that you you say that because that's that's one of the sort of ah things that i defend in the paper, right, is that this particular... Potentially, we don't want AI that kind of and always and gives us exactly what we're after. And that the aesthetic case may be one of these such cases, right? And potentially, if we have an AI that's just feeding us exactly what we think we want aesthetically, then we're going to miss out on the possibility for things like Van Gogh.
00:25:24
Speaker
um And I think it's almost, I mean, it's particularly been relevant recently. If you heard in the the news about these kind of... and it like sycophantic AI kind of telling you exactly what you want to hear. It's a little bit like that, but for the aesthetic diet, right? Like always being given exactly what you want is not necessarily good, and good for you in some way.
00:25:42
Speaker
and Yeah. It emotionally stunts us or something if it's, you know, doing that. um Okay. We should probably um motivate your secondary goal ah for listeners so that we can then, you know, really drive home your, your main point is,
00:25:59
Speaker
But so there might be someone listening who is wondering like, okay, I don't even know how, ah why there wouldn't be alignment between AIs given that they're trained on our data and everything. So can you just tell us, you know, first of all, like why there might be an i alignment problem in the domain of aesthetics? Yeah.
00:26:21
Speaker
um Yeah, so I think the the sort of point that it's trained on our materials is an interesting one. um i mean, in general, the kinds of AI that are generating things like images are not trained on our very best things, right?
00:26:37
Speaker
So there's a sense in which they actually aren't perhaps aligned with our you know our greatest artworks. They're just trained on images on the internet, right? And so um i wonder that even our current AI are capturing any of our aesthetic values, maybe as we'd like to see them.
00:26:51
Speaker
That doesn't preclude the possibility that we could kind of train one to do this, that right which some people have tried to do. and so So it is, I guess, i guess there's sort of a general scepticism that that is kind of what's going to happen in with actual designing of AI systems. and But in in the paper, I kind of draw attention to a couple of different cases that suggest that sometimes AI systems kind of pay attention to different sorts of things than we do.
00:27:14
Speaker
um So i I talk about one case and where... and and AI will, when it's categorizing images, pay attention more to texture than shape, whereas humans ah will pay more attention to shape than texture.
00:27:27
Speaker
um Now, that's just a sort of fairly minimal example, but we could imagine that systems which don't have the same sort of perceptual basis as we do and don't process images in the same ways as we do or exactly the same ways may develop and kind of focuses or interests yeah you know loosely defined in different ways than we will.
00:27:46
Speaker
um And so i see it as ah sort of possible that there might be and AI systems in the future which are kind of um have different ah yeah sort of focuses, attentions and areas of of sort of and where they see as value or difference, then we will.
00:28:05
Speaker
um Now, I think if people wanted focus on that, they could try and sort of ah ah align that in some way with humans. and But I don't think that means that, you know, a future super intelligent AI might and not be aligned in that in that sort of way.
00:28:20
Speaker
don't know if that was clear. Was that kind of clear? I guess I just have a follow-up. So is the concern that, mean, first of all, your point is well taken. 90% I think it's like Sturgeon's Law, like 90% of everything is crap. I don't know what the saying is exactly, but most of what it was trained on is just, you know, picture of Obama, picture of George Bush, whatever, you know.
00:28:46
Speaker
um But and so you mentioned that the the texture versus shape thing. So is, is, it's possible that AI pictures would just focus too much on texture, which is something that we don't even notice and rather than shape, which is what we have a preference for.
00:29:03
Speaker
And as such, there will just kind of be a general, you know, lack of alignment between the sort of shape oriented images that we would like and the texture oriented images that, uh, that the AIs would make. Is that, did I get that right?
00:29:20
Speaker
um Potentially, yeah. That's kind of the the sort of line of thought that I'm trying to go for, right? If we already see that when they're categorizing things, they're privileging different kinds of what we would call visual stimuli, right, like visual features.
00:29:34
Speaker
and If they're already kind of seeing this as more important just for categorization, then they may also end up seeing them as more important for and markers of value, right? so you know, if they see that we love, um I don't know, and paintings and of landscapes, rather than thinking we really like things with trees in, they might think we really like things with green in.
00:29:55
Speaker
And so that just gives us lots of green, right? So something like that is kind of the the line of thought that I'm going for um with with that case. that's That's a great example, yeah. And so and and I guess that that is actually, like you might correct me on this. this this is um do We do have a preference.
00:30:13
Speaker
Is there like an evolutionary preference reason why we have preferences for certain landscapes? and then And so we would rate them as such? Oh, yeah. I've heard that. I don't remember all it, but yeah. yeah i've heard right like that's I mean, everybody loves a good, yeah, like open, you know, kind of landscape. or Anyway, yeah, I've heard i've heard of that, Roberto, yeah.
00:30:33
Speaker
Yeah, i I don't know about that, but I could totally see that being the case, right? I mean, some of some of our preferences for things may be tied to evolutionary basis or, you know, and it just embodied sort of basis, right?
00:30:45
Speaker
You know, and things to do with that us existing as bodies in the world. um There's not like a tradition of like painting holes in the garden, you know? yeah, it's another hole painting. Like, anyway.
00:30:58
Speaker
Quite, yeah. Yeah. Cool. Okay. Well, um so I don't know. Is this a dumb question? Like, what about the whole idea of like, you know, AI doesn't even have experience. So like, how can it have like any aesthetic values? Like it doesn't even have, I guess,
00:31:18
Speaker
This is this is like a based on a misunderstanding, probably. But anyway, rate like yeah what what do you do to so with... like but I guess someone could say, well, you know AI has no lived experience. It's not conscious.
00:31:28
Speaker
So it's you know it's it's very likely to go off the rails aesthetically because it can't even use the lived experience to see if...
00:31:41
Speaker
something sounds or looks good. I mean, I don't know. I just, that's how you test if something's aesthetically valuable, right? You have to like experience it and see how it either sounds or looks or reads, but AI doesn't even have,
00:31:56
Speaker
experience, right? So like, isn't that, is that a problem for alignment? Like, isn't, i don't know. Is that, does that make sense? No, it's it's not a silly question at all. Cause I think that that is potentially a problem, especially today, right? in With today's artificially intelligent systems, and they are not experiencing things and as, at least as we are probably I think lots philosophers it agree they're not experiencing things at all.
00:32:18
Speaker
um So if we talk about aesthetic value as relating to aesthetic experience, which we do, and then they're potentially not going to have access to aesthetic value. I think that doesn't mean that they might not generate something which is aesthetically valuable for us, right?
00:32:32
Speaker
um And that that still might not be aligned with our values, particularly because they're not having those kinds of experience. You know, that might also be additional way that this could cause a problem. um So, yeah, i think I think that might be possible.
00:32:44
Speaker
I guess and in the paper, at least, I'm i'm somewhat engaging with that. you know Well, I am very much engaging with that literature on superintelligent AI. So sort of hand wave these problems a little bit. You know, a future superintelligent AI may have something like aesthetic experience.
00:32:57
Speaker
It may, of course, not. Right. and And it may have no concerns for aesthetics at all. and And we might wonder what an AI system with no aesthetic values might result in. Right. Even worse than maybe misalignment is that it just doesn't.
00:33:11
Speaker
you know, doesn't align in that he doesn't recognize it at all as a form of value. and Yeah, I don't know. I'm sort of ah hypothesizing there a bit. Yeah, yeah, yeah. Yeah, i do think it's relevant for sure.
00:33:22
Speaker
So, oh ah real quick though, another thing, did we, I mean, maybe, should we talk about like, There's this whole issue of like, does the alignment problem even apply to the aesthetic domain? I'm just, and don't know if this is connected yet, but you kind of bring up the bug-eyed monster or someone brings it up. Anyway, you quote in your paper, but like basically air of Pulp Fiction, the magazine covers would have, you know, you might have like a monster, bug-eyed monster, and it's carrying off
00:33:54
Speaker
an attractive female in a torn dress. And so the idea there is like the bug eyed monster. Oh, it looks like they're, they they have a similar like preference as we do in terms of like, yeah they're, they're, they're also stealing the attractive female rather than like stealing, i don't know, whatever, like,
00:34:15
Speaker
old shoes or something. Anyway, sorry but anyway, so what that, that example is showing us that that kind of shows us of like this, it's kind of illustrative of the whole idea that like there can be misaligned.
00:34:29
Speaker
There was a potential for misalignment. Maybe AI is generating, sorry, I don't know. Is this, is this making fun sense? Like that whole example. Yeah, so I think made that that and part of my point, I think, if I remember rightly, um in talking about that example is to say that um aesthetic values are kind of already coming up in some of these cases that are used to illustrate the potential of alignment problem. um And I think...
00:34:56
Speaker
um I think that that that case in particular is and often used to say, you know, we're goingnna we're going to have these assumptions potentially, just like in that case, and that this completely different creature to us will align with us. It will think that, you know, we'll go for the same attractive woman um rather than shoes, as you said, right? and Old shoes.
00:35:18
Speaker
Yeah. So that's sort of something that we assume will be the case, but there is no reason to think that. And as a creature that's so very different from us, like the bog-eyed monster, like an AI system, very well could have completely different value values, could go for completely different things than we would, right?
00:35:36
Speaker
um So yeah, so part of what I'm interested in there is just sort of saying already when we're thinking about such cases, where we're bringing up aesthetics, right? We're bringing up aesthetic examples. and so like so And so maybe like, for example, it might really be into to texture because you're talking about like AI relying more on texture than shape. So maybe like that would be like a potential thing is like actually when it's creating art, it's really into like novel things.
00:36:04
Speaker
Textures. Tactile. I guess. Is that the kind of thing that happen? you know like that could happen is Sure, yeah. I mean the other yeah the other case that gets brought up that I bring up and to do with that is is that sort of the so shift in the threshold for boredom, right? So just minimally changing something is is sort of valuable to the AI system.
00:36:26
Speaker
and So to us that's kind of, if every time you asked an AI for a painting it gave you like almost the exact same image, you'd be just sort of quite quite upset. You'd think it was a bit useless I guess.
00:36:37
Speaker
um so So that's something that we could see happening. It's kind of generating constantly the same thing, but because for that system, its threshold is set differently to ours. and It's threshold for novelty or boredom or whatever it is.
00:36:49
Speaker
um So yeah, so these kinds of ways we could we could sort of imagine such a case where an AI system would be kind of favoring a feature or quality or a difference between things that we just wouldn't care about.
00:37:03
Speaker
So that takes us actually into sort of your main reason for wanting, so, i mean, you're basically um arguing for your position, but want to like make it very explicit, right? Like this is exactly why we want only moderate alignment.
00:37:19
Speaker
So i I know we're, we're jumping ahead a little bit here, but I, I feel like this is a good place to, um, to, to, you know, put this in here. so ah Let's talk ah about, what do you think, Sam? New aesthetic values? Is it time for new aesthetic values?
00:37:38
Speaker
Let's do it. Let's do it. All right. So I wish we had a cool music right now to launch us into. ah into new aesthetic values. Yeah. ah So, yeah, some AI-generated music. So um tell us about how AI might create new aesthetic
AI and New Aesthetic Values
00:38:00
Speaker
values. like what what What does that mean? How would that work How would we know when it's happening? I don't know. I just, you know, this is ah this is the this is it.
00:38:08
Speaker
Bam, this is the part. Here we go. Okay. and Yeah, I mean, disappointingly, I think so in some ways it's going hard for me to say because I can't conceive of them, right, to some extent. The idea is that it might might produce something and that I don't yet recognize the value in, that I don't yet consider it to be an aesthetic value.
00:38:27
Speaker
and But yet, like Van Gogh's paintings, i you know down the line, we realize that there's something really special there. and I can draw upon a case where I think AI is contributing something which maybe it maybe isn't completely new as an aesthetic value and we do seem to recognize it, but that it's a little bit different, a little bit unique to the AI case.
00:38:49
Speaker
and And that's, and I write about weirdness, right? the idea that and some AI images are weird, we see odd shapes, extra fingers, it puts features of the face in the wrong places, etc.
00:39:02
Speaker
um And um anecdotally, at least, it seems that people kind of pay attention to this. So some artists that work with AI are particularly interested in pushing and the boundaries of sort of the weirdness and of of of the generated images.
00:39:16
Speaker
They draw attention to the fact that they didn't and understand why it was doing that. They didn't anticipate it doing that, you know, arranging the features in that way or creating these weird blob-like shapes. and People also draw attention to these things online. They think they're funny, right, when an AI is trying to draw handshakes and it's got all the extra fingers and extra arms everywhere.
00:39:34
Speaker
and So I kind of argue that there's something and particularly that we find interesting in this case that it's um and that it's a kind of sort sort of non-human in some kind of way, right? So there's a particular, well, I argue a particular failure of the AI system that's going on, and it's failing in a way that a human wouldn't do.
00:39:54
Speaker
So it's it's tied to the fact that it's not a human, that it's producing images in these ways. um So that's the sort of thing where we might get to some sort of value, right? We might get to something that's particularly interesting. and It's a particular kind of aesthetic feature aesthetic quality in some way, we might start valuing it. And it's not the sort of thing we necessarily saw and from human made art, for example.
00:40:17
Speaker
I don't know if that gives like ah a bit of a case that an AI might produce something that's uniquely valuable or that's uniquely has unique aesthetic qualities. um And I'm thinking of those sorts of cases when I'm talking about maybe future new aesthetic values that we're not yet aware of.
00:40:35
Speaker
So could i quick follow up here? Because i i I was thinking like, there's this interesting element of your paper, which is like about, it's like a theory of innovative art, maybe like what is it?
00:40:50
Speaker
Like what's going on with Van Gogh or um whatever, like pet sounds by the Beach Boys or like Stravinsky or something, something that's like initially not appreciated.
00:41:04
Speaker
And, you know, is it that there are new aesthetic values? I was kind of wondering about that because I was thinking, well, you know, with like Stravinsky, I guess there's like a lot of dissonance in that. And maybe people were getting upset by the dissonance um but is that a new value, aesthetic value? like I don't know. I guess like I'm kind of wondering, this might be taking us for field, but I'm kind of wondering, like,
00:41:38
Speaker
you know Maybe someone could say, you know actually, you want it to be fully aligned with our values. like You want it to you know pursue something beautiful and you want it to have like grace or complexity or you want it to have like all those like traditional labels. want to strive for those. But maybe innovative art broadens our understanding of what is beautiful or striking or I don't know. I guess, do do you see kind of where I'm going with this or should it?
00:42:14
Speaker
Yeah, i think so I think so. I think in the um in the paper at least, what it what it comes down to to some extent is how an AI system would sort of understand those those values, right? Right. So, i mean, Russell and Peterson, they talk about, say, human preferences, behaviours as a kind of indication of values, that kind of thing. Right.
00:42:35
Speaker
and So regardless of whether we have, maybe we have these sort of buried values that we're not yet aware of and haven't uncovered, and we still don't, the system is not going to be able to access them. So I think the kind of argument might still stand anyway. Right. Because like regardless, we're not going to react as if we value this thing and it may be uncovered later.
00:42:53
Speaker
It's going to come down, I suppose, whether you think they're sort of out there aesthetic values and maybe some of them we discovered, of them we haven't, or whether, you know, aesthetic values could manifest that are new in some way or whether there's kind of a very limited number of them.
00:43:07
Speaker
and I kind of, i don't know what what I would say would be the answer to that. I guess for me, and I think some of these artistic cases are doing something transformation transformational, right? like So to come from the sort of creativity literature, right? They're doing something transformational and to like a conceptual space about what art is, what is valuable about art.
00:43:30
Speaker
And they're shifting that in some way. and so So one way or another, there's something of key that's been transformed there. So I guess that's yeah what I focus on. and and Yeah. But i yeah, I mean, someone probably could could very make a convincing case that, you know, underlying it all, we actually have these values, but I just don't know how a system would be able to access them if we're not even aware of them, right? Yeah.
00:43:52
Speaker
So yeah, if we solve that puzzle, then sure, and I would be willing to concede that, I think. Yeah. Yeah. Anyway, yeah. I mean, at any rate, it seems it makes sense that like, you know, if in 1913, there was super intelligent AI and it was making music,
00:44:09
Speaker
It presumably would not include dissonance because it would know our, if it if it was aligned with our preferences, it would sense that we're not into that much dissonance. I mean, I guess like there was already some kind of, I mean, there had to be dissonance before. I don't know much about classical, but there had to be some kind of, but but maybe it was like, we didn't we didn't like that much dissonance. So it was like, it was going put threshold on that. And at a rate, so if we imagine a situation where it's super aligned with us, it's going to not create something like the Rite of Spring.
00:44:48
Speaker
um and And so... It seems, it seems like it's anyway. Yeah. It seems like it's putting on it Yeah. I just had a thought and it's probably a gobbledygook, but I'm going to just spew it out anyways. And you can just respond as you see fit.
00:45:04
Speaker
um So I'm now, now I'm thinking, okay, sometimes it's like the theme or like a music, there's a motif that is not currently appealing to us. So this, I think it was the minor sixth in that, you know, uh,
00:45:19
Speaker
ah classical piece that caused a riot whatever. And now you hear the minor six all the time and and heavy metal music. That's like, as a matter of fact, if it if you don't play the minor six, I think it does not technically qualify as a metal song.
00:45:33
Speaker
So, so that's going on. But now I'm thinking there's also something else ah that, that machines can do much better than us. And has to do with like sort of, you know, I have, I'm limited, I play the piano, i'm limited to 10 notes at a time, but AI is not. And I'm, you know, what if ah AI starts coming up with like, you know, 20 tone chords that, that occur in the beginning, I'm going to, you know, I'm sure it's going to sound very jarring to me, but over time,
00:46:01
Speaker
I'm going to love these harp sounding chords, you know, and that's exactly, or or maybe really weird time signatures, like that I have a hard time counting, but AI doesn't, you know, even think about it, it just does it automatically. So are these other examples of how, there's no question, I'm just going to stop talking, let you respond.
00:46:23
Speaker
Yeah, I mean, I think potentially, certainly, you know, depending on how much our, you know, these kinds of aesthetic qualities or features of artworks are tied to us as humans, right, we're we're potentially going to see the breaking of that boundary.
00:46:37
Speaker
I mean, I think humans have already done a pretty good job in some ways of being able to develop things that that kind of break some of those boundaries themselves. and But there might be things that we wouldn't wouldn't even have thought of, right?
00:46:49
Speaker
to do. I mean, ah there's a really interesting case of um some people that work with quantum computing to generate and music. and And so sort of like the the qualities that come out from using a system that's kind of ah ah what's the word, probabilistic in that way is is very interesting. it mean, it's a little bit atonal to my ears, but but, you know, there's something particularly there that humans are not going to be able to do that the system is able, you know, this computer is able to do.
00:47:18
Speaker
and And it results in some aesthetic qualities and that we might not have seen before. so its yeah It's because the cord is in superposition, it has not resolved yet into its state.
00:47:29
Speaker
but um I'm sorry. Oh, gosh. Oh, I had a question.
00:47:54
Speaker
oh could you Oh, we have to talk about the the thought experiment. Okay. So you you gave up the you came't came up with this like a thought experiment where the AI is inspired by sensory deprivation dining and it like starts impairing people's vision.
00:48:12
Speaker
ah ah thought that was pretty fun. Do you want to tell tell us about that one?
Thought Experiment: AI and Sensory Deprivation
00:48:16
Speaker
um Sure, yeah, I can i can and tell you about that case. and So I guess the part of the point of that thought experiment was to sort of illustrate the some risks of not recognizing creativity in AI, right? That was part of the point I was making. So and the case is that... um We are potentially interested in aesthetic experiences that involve the limitation of some of our senses, perhaps to heighten other senses. And there's that case, of course, of the the restaurant where you go and you and eat food in the dark, right? So you kind of experiences it sprint is ah experience it as if you were visually impaired.
00:48:50
Speaker
and and people kind of find something um aesthetically rewarding about this experience or they notice different things in the food than they would have normally. and And we can imagine a sort of case where an AI recognizes this um but and thinking creatively thinks they'll do something sort of different and to this and decides instead and to create a similar aesthetic experience but rather than turning off the lights they're going to sort of permanently impair your vision.
00:49:18
Speaker
um And while there may be sort of values in that, and it might be creative and it might be, you know, complex um and aesthetically valuable experience, and which the AI is going to see is far more valuable, particularly as you're going to kind of go about the world in a, in a different way. And in some ways we think that has value, right? After engaging with an artwork, we might think that it changes how we think about things and that that's valuable in some way. Um, but ultimately we as humans are going to look at that and think it's not great that it's come up with this sort of creative way allowing us to achieve these experiences.
00:49:50
Speaker
Right. Um, it's got a lot of negative, negative side effects. Um, and there's a kind of, a, you know big concern with kind of permanently damaging people for an aesthetic experience.
00:50:02
Speaker
and So part of this is just this idea that you know that's that's ah evidence of sort of creative thinking, and which if we hadn't noticed that an AI was able to do any kind of creative thinking, like actual creativity, because it was doing it in in a domain that we didn't recognize or with kinds of um sort of shift in value that we didn't recognize, and then it might grow these kinds of capabilities. It was sort part of the concern there. So kind of linking that concern from you know, growing creativity to um to an ethical concern. and Again, sort of partly showing that link between the ethical and aesthetic domains.
00:50:34
Speaker
yeah Yeah, that makes sense. I would like that example, by the way, just because it it's like it's so hard to predict what, how it could go wrong. And so imaginative examples like that really kind of put it, you know, front and center is like, yeah, it's just going to, you know, you have to, you have to manage, you have to build it. So it's constrained. So that doesn't do that. That's why I feel like, yeah, because yeah, just piggybacking on what you're saying, Richard. I mean, yeah, I mean, that gets to the control problem, right? and I mean, i don't I don't think I really explicitly addressed the control problem in this paper, but but some people say ultimately, you know, that's going to be a key issue, right, alongside this value alignment problem. They say that they're linked, right, but but that that's part of it is that we have AI systems that aren't,
00:51:17
Speaker
suddenly out of our control and they'll do all sorts of things. So perhaps we should design them with, ah with sort of constraints in place in that way. So they never go fully out of our control, partly because we're not sure we're ever going to be able to get them to align with our values.
00:51:30
Speaker
You know, that that example also makes me think of like, uh, I imagine this is something people talk about, but, or maybe you talk about it, but just the whole idea of like hierarchy of values. Cause I feel like when I think about the value alignment problem, it makes me think it's like unsolvable because and fully, because like,
00:51:52
Speaker
It seems like a ah lot of the issue has to do with correctly or ordering values rather than just having the value. And then at that point, I'm like, how do you how would you program something to have the right order? Like, what? Like, that sounds like such a... You know, because I mean, even with how, right? Like...
00:52:10
Speaker
So, you know, the the kind of first thought is, oh, Harold just doesn't have the value of like human life, or maybe he just doesn't value um honesty. But then if you think about it's like, well, actually, you know, we also will be dishonest in certain circumstances, but we know so we are like, what's the word? It's like we,
00:52:34
Speaker
we will we We know the limit of dishonesty. We know like not to go too far, generally speaking. Of course, there are certain people who will... like um yeah they they don't out value honesty enough. And so they end up, you know, really being dishonest throughout their life. But like, most people will agree that there are certain circumstances where you can be a little bit deceptive, you know, like, you know, someone's crying and they're like, what do you really think about like what I'm wearing? And it's like,
00:53:05
Speaker
It's great. you know you It's like, I'm going to lie to you at this moment. You know what i mean So it's like, we we know how to like... Ravishing. you' like Yeah, yeah. Yeah, exactly. So whatever. So point being is that like, we do engage in like, you know, little bit deception to some degree.
00:53:21
Speaker
And so it's like almost like, it's not that Hal just lacks a value maybe. It's it's that he doesn't know how to throttle it right and he like goes overboard. Anyway, it's just interesting to think about like, that seems to suggest it's like a hierarchy issue of like,
00:53:34
Speaker
when some another value overrides one and at that point i feel like you know becomes really tricky because i don't know how you yeah yeah no i mean i i think that's sort of like um think people call that a sort of threshold issue right at what point it becomes okay and what point it does or doesn't matter um gonna like sort of take this question very broadly though because i think that this is where philosophy comes in with with talking about AI, right? Because this is such a central philosophical question about kind of the nature of values, um the nature of kind of ethical concerns, you know, what kinds of ah system of evaluating goods we should be using.
00:54:16
Speaker
um And I think like, but you know, the whole of philosophy has been aimed at some of these questions. ah wouldn't I wouldn't want to try and sort of adjudicate on the the best response to that now. but But that's where I think that, you know, philosopher philosophers and philosophy has a real...
00:54:30
Speaker
utility, like a real usefulness in engaging with questions about AI, right? and Because a good person to ask about this might be someone that kind of has studied this. and They still might not have the answers, but maybe they can get you know some way towards an answer about how we might want to approach these kinds of issues. you know Is it about where we set a threshold? Is it about and a hierarchy of values? Is it possible to have a hierarchy of things with intrinsic value? you know That kind of thing.
00:54:57
Speaker
Yeah. And it all comes back to ethics, right? Ethics becomes imperialistic. um So ah maybe by way of closing, you can tell us, so i know this paper was, ah when was this paper published? Last year, 2024. So 2024, a year is an infinity in academic ah time. So what what are you working on now or what are you planning on working on next?
00:55:23
Speaker
and Yeah, I mean, lots of things. and But at the minute, there's sort of two main things that I'm working on. and One is around responsibility for AI art. So I'm actually on a fellowship at the moment and at FAU in ah Germany and working on the problem of kind of how we should attribute responsibility for positive things that AI is.
00:55:44
Speaker
sort of produces, so art being my so key case, and but potentially broadening that out to other sorts of cases, and right praiseworthy rather than blameworthy cases.
00:55:55
Speaker
um Because I think this might have some implications for things like how we credit people that are involved in the process ultimately of producing images with AI. um So there's that's sort of part of what I'm working on. And the other thing I'm continuing to work on is um the sort of issues around ah creativity and AI. So I'm trying to sort of think about if we could develop an account of machine creativity. So creativity that doesn't just sort of rely on the underlying understanding of how humans work, and but has something that's a little less chauvinistic to it.
00:56:28
Speaker
right and that project has been long ongoing so it's still in the works um but yeah those are sort of two of my main areas of research at the moment that's very cool we hope you uh you might consider coming back on and chatting about machine creativity it would be my pleasure awesome yeah well thanks a lot for coming alice alice has been awesome really like your paper and yeah hope oh all listeners you should definitely check it out so thanks so much it's been great to to talk to you