#427 — AI Friends & Enemies
Episode Stats
Words per Minute
192.95934
Summary
In this episode, Dr. Paul Bloom joins me to talk about his work on AI, morality, and the origins of our ideas about the nature of morality. We also discuss his new book, "The Origins of Morality," which he is writing about.
Transcript
00:00:00.000
Welcome to the Making Sense Podcast. This is Sam Harris. Just a note to say that if you're
00:00:11.740
hearing this, you're not currently on our subscriber feed, and we'll only be hearing
00:00:15.720
the first part of this conversation. In order to access full episodes of the Making Sense
00:00:20.060
Podcast, you'll need to subscribe at samharris.org. We don't run ads on the podcast, and therefore
00:00:26.240
it's made possible entirely through the support of our subscribers. So if you enjoy what we're
00:00:36.660
I am here with Paul Bloom. Paul, thanks for joining me.
00:00:42.180
Yeah, great to see you. It's been, we were just talking off mic, it was, it's been years,
00:00:50.460
I've been following you, though. You've been doing well.
00:00:52.500
Yeah, well, I've been following you. I've read you and the New Yorker recently writing
00:00:59.340
about AI. I think you've written at least two articles there since we, we actually wrote
00:01:04.000
a joint article for the New York Times seven years ago, if you can believe that.
00:01:09.140
About Westworld and the philosophical import of watching it and realizing that only a psychopath
00:01:15.960
could go to such a theme park and rape Dolores and kill children, et cetera. And I think we
00:01:22.900
predicted no such theme park will ever exist because it'll just have, it'll be a bug light
00:01:26.960
for psychopaths and normal people will come back. And if they do anything like that, they'll
00:01:32.540
scandalize their friends and loved ones and be treated like psychopaths appropriately.
00:01:37.480
We'll see. We may be proven wrong. Who knows in this crazy time? But yeah, that was a fun
00:01:44.700
article to write. And I think we wrestled with the dilemmas of dealing with entities that at
00:01:50.720
least appear to be conscious and the moral struggles that leads to.
00:01:54.440
Yeah. Well, so I think we'll start with AI, but remind people what you're doing and what kinds
00:02:00.860
of problems you focus on. I think though we haven't spoken for several years, I think you still
00:02:05.700
probably hold the title of most repeat podcast guest at this point. I haven't counted the
00:02:12.240
episodes, but you've been on a bunch, but it's been a while. So remind people what you focus on
00:02:16.900
as a psychologist. Yeah, I'm a psychology professor. I have positions at Yale, but I'm
00:02:22.360
located at University of Toronto and I study largely moral psychology, but I'm interested in development
00:02:29.140
and in issues of consciousness, issues of learning, notions of fairness and kindness,
00:02:35.380
compassion, empathy. I'm actually been thinking, it's funny to be talking to you because
00:02:40.500
my newest project, which I've just been getting excited about has to do with the origins of
00:02:45.380
morality. And the spark was a podcast with you talking to Tom Holland, author of Dominion.
00:02:56.000
Yeah, I found this a great conversation. He made the case that a lot of the sort of morality that
00:03:01.020
maybe you and I would agree with that, you know, the idea of respecting the right to universal rights
00:03:06.280
and a morality that isn't based on the powerful, but instead maybe in some way respects the weak.
00:03:12.480
It's not the product of rational thought, not the product of secular reasoning, but instead the product
00:03:18.220
of Christianity. And he makes this argument in Dominion and several other places. And I've been
00:03:24.000
thinking about it. I mean, it's a serious point. He's not the only one to make it. It deserves
00:03:28.260
consideration, but I think it's mistaken. And so I think my next project, as I've been thinking
00:03:33.960
about it, and I thank you for the podcast getting me going on this, is to make the alternative case,
00:03:39.420
to argue that a lot of our morality is inborn and innate. I'm a developmental psychologist. That's
00:03:44.540
part of what I do. And a lot of morality is a product of reasoned and rational thought. I don't
00:03:49.660
deny that culture plays a role, including Christianity, but I don't think that's the big story.
00:03:54.520
Nice. Well, I look forward to you making that argument because that's one of these
00:04:00.360
shibboleths of organized religion that I think needs to be retired. And so you're just the guy
00:04:05.300
to do it. So let me know when you produce that book and we'll talk about it.
00:04:10.320
I mean, I got to say, I heard you guys talk and you sort of, you know, you guys engaged properly
00:04:14.560
on the idea. And Holland, I've never met, but he seems like a really smart and quite a scholar.
00:04:20.020
So it's not, you know, and he gets, I think, credence by the fact that he's in some way
00:04:26.000
arguing against his own side. He himself isn't a devout Christian, a secular. But that doesn't
00:04:31.680
make him right. So I'm very interested in engaging these ideas.
00:04:35.500
So how have you been thinking about AI of late? What has surprised you in the seven years or so
00:04:42.840
that we, uh, since we first started talking about it? A mixture of awe and horror. I'm not
00:04:49.460
a doomer. I'm not as much of a doomer as some people. I don't know. I don't know when the last
00:04:54.020
time I checked with you, Sam, what would you say your, your P doom is? That's I've never actually
00:04:58.600
quantified it for myself, but, um, I think it's non negligible. I mean, it's very hard for me to
00:05:05.120
imagine it actually occurring in the, in the, in the most, um, spectacularly bad way of, you know,
00:05:11.520
a very fast takeoff, you know, an intelligence explosion that ruins everything almost immediately,
00:05:16.100
but, you know, fast or slow, uh, I think it's well worth worrying about because I don't think the
00:05:21.720
probability is tiny. I mean, I, you know, I would put it in double digits. I don't know what those
00:05:26.800
double digits are, but it's, I wouldn't put it in single digits given the kind of the logic of the
00:05:31.220
situation. I think we're kind of in the same place. Yeah. I mean, people always talk about,
00:05:35.740
you know, you have a benevolent, super intelligent AI and you tell it to make paperclips and it
00:05:39.500
destroys the world and turns us all into paperclips. But there's another, another vision of,
00:05:43.800
of malevolent AI I've been thinking about with the rise of, um, what's it called? Mech Hitler.
00:05:48.200
Uh-huh. Yeah. Mecha Hitler. Yeah. Mecha Hitler. You know, so, you know, here's a simpler scenario.
00:05:55.000
Some deranged billionaire creates an AI that fashions itself on Hitler to defend,
00:06:00.960
the Trump defense department purchases it and gets it to, connects it to all of his weaponry
00:06:05.840
systems and, uh, hijinks and zoo. How could you come up with such an outlandish scenario that could
00:06:11.040
never happen? There's just no way. It's a bizarre fantasy. And by the way, it also makes porn.
00:06:15.220
That's right. Yeah. So just, just to get to the trifecta. Yeah. Um, so anyway, I, I'm like a lot
00:06:20.720
of people, I, I worry about that. I, I also find at the same time, I find AI an increasingly
00:06:26.080
indispensable part of my intellectual life. I, um, you know, I, I, I have a question,
00:06:31.040
I got it, but, but not, not just a specific, I use it as a substitute for Google. I, you know,
00:06:35.600
say, you know, where's a good Indonesian restaurant in my neighborhood and, you know,
00:06:39.400
how do you convert these euros into dollars? But, but I also have a question like, I don't know,
00:06:44.640
I'm, I got into an argument with somebody about revenge. So what is the cross-cultural evidence for,
00:06:49.320
uh, revenge being a universal and who would argue against that?
00:06:52.080
And three seconds later, boom, bunch of papers, books, thoughtful discussion,
00:06:57.040
thoughtful argument mixed in our hallucinations. Occasionally I find it cites a paper written by
00:07:02.160
me that I've never written, but it's astonishing. I've been credited with things. Yeah. Yeah.
00:07:07.880
It's told me I've interviewed people I've never interviewed and it's amazingly apologetic,
00:07:12.180
uh, when you correct it, but, um, I'm so sorry. Yes, you are totally right. Let me now give you
00:07:18.360
citations of papers that actually exist. Yeah. In your working day, when you, when you write,
00:07:23.160
when you prepare for interviews, how much do you use it? Well, I've just started experimenting with
00:07:27.860
it in, in a way that's, um, probably not the usual use case. So we have, we, we have fed everything I've
00:07:35.700
written and everything I've said on the podcast into chat GPT, right? Yeah. Yeah. So, so we have,
00:07:41.980
well, actually we have two things. We have, um, we've created a chat bot that is me that's,
00:07:47.160
that's model agnostic. So it can be run on chat GPT or, or a Claude or whatever the best model is.
00:07:51.660
We can swap, you can swap in whatever model seems best. But so this is like at the, you know,
00:07:56.560
a layer above, uh, at the system prompt level. And it has, again, access to everything. It's
00:08:03.540
something like 12 million words of me. Right. So it's a lot is, you know, that's, that would be a
00:08:08.400
lot of books. We've just begun playing with this thing, but it's, um, it's impressive because it also is,
00:08:15.180
it's using a professional voice clone. So it sounds like me and it's every bit as monotonous
00:08:20.780
as me and it's delivery. I mean, so I'm, I'm almost tailor-made to be cloned because I already
00:08:25.800
speak like a robot. Must be agonizing to listen to sound. It is, but it hallucinates and it's,
00:08:31.100
it's capable of being weird. So I don't know that we're, we're ever going to unleash this thing on
00:08:35.260
the world, but it's, um, it's interesting because it's like, so even having access to every episode of
00:08:41.280
my podcast, it still will hallucinate an interview that never happened. You know, it'll still tell
00:08:46.920
me that I interviewed somebody that I never interviewed. And so it's, that part's still
00:08:51.320
strange, but presumably the, um, you know, this is as bad as it'll ever be. And the general models
00:08:56.340
will get better and better. One imagines. So we're, we're looking at each other on video and I imagine
00:09:02.180
it doesn't do video yet, but if we were talking on the phone or without video, would I, um, be able
00:09:07.920
to tell quickly that I was talking to an AI and not to you? Uh, only because it would, uh, be able
00:09:13.700
to produce, uh, far more, uh, coherent and comprehensive responses to any question. I mean,
00:09:19.200
if you said, cause it's, because it's hooked up to whatever the best LLM is at the moment, you know,
00:09:24.680
if you said, give me 17 points as to the cause of, of the great depression, it would give you exactly
00:09:30.980
17 points detailing the cause of the great depression. And I, and I could not do that. So you would,
00:09:35.680
it would fail the Turing test as all these LLMs do by passing it so spectacularly and instantaneously.
00:09:43.300
Yeah. I mean, that actually, that's a surprise. I want to ask you about that. That's a surprise
00:09:46.580
for me, how not a thing the Turing test turned out to be. It's like the Turing test was the staple
00:09:54.520
of our, of, you know, the cognitive science literature and just our, our imaginings in
00:10:01.340
advance of credible AI. We just thought, okay, there's going to be this moment where we
00:10:05.660
are just not going to be able to tell whether it's a human or whether it's a bot. And that's
00:10:10.300
going to be somehow philosophically important and culturally interesting. But, you know, overnight
00:10:17.680
we were given LLMs that fail the Turing test because of how well they pass it. Yeah. And
00:10:23.220
this is no longer a thing. It's like, it was never a Turing test moment that I caught.
00:10:27.680
All of a sudden you and I, we have, I have a super intelligent being I could talk to on
00:10:32.000
my iPhone and I'll be honest. And I think, I think other psychologists should be honest
00:10:36.120
about it. If you had asked me a month before this thing came out, how far away we were from
00:10:40.880
such a machine, I would say 10, 20, 30 years, maybe a hundred years. Right. And now we have
00:10:45.820
it. Now we have a machine we could talk to and sounds like a person and except like you
00:10:50.360
say, it's just a lot smarter. Yeah. And it is mind boggling. It's, I mean, it's an interesting
00:10:55.640
psychological fact, how quickly we get used to it. It's as if, you know, aliens landed,
00:11:01.020
you know, next to the Washington monument and now they walk among us. Oh yeah. Well,
00:11:05.080
that's the way things go. Yeah. Oh, you develop teleporters. Now we got teleporters and we
00:11:10.180
just, we just take it for granted now. Yeah. Well, so now what are your thoughts
00:11:14.600
about the implications, you know, psychological and otherwise of, of AI companionship? I mean,
00:11:20.560
what is, so at the time we're recording this, there's been recently in the, in the news
00:11:24.740
stories of AI induced psychosis. I mean, people get led down the primrose path of their delusion
00:11:31.620
by this amazingly sycophantic AI that, that just encourages them in their Messiah complex or whatever
00:11:37.840
flavor it is. And I think literally an hour before we came on here, I saw an article about
00:11:46.060
chat GPT, um, encouraging people to, uh, pursue various satanic rituals and telling them how to do a
00:11:52.760
proper blood offering that entailed slitting their wrists. And, uh, as one does, uh, in the presence
00:11:58.820
of a super intelligence, I know you just, you just wrote a piece in the, in the New Yorker that people
00:12:03.640
should read on this, but give me your sense of, of, uh, what we're on the cusp of here.
00:12:08.640
I have a, I have a mild form of that delusion in that I think every one of my sub stack drafts
00:12:14.280
is brilliant. I'm told is just, you know, Paul, you have outdone yourself. This is sensitive,
00:12:18.640
humane as always with you. And no matter what I tell it to, I say, you don't have to suck up to
00:12:23.180
me so much, a little bit, but not so. It just, and now I kind of believe that I'm much smarter
00:12:28.060
than I used to be because I have somebody very smart telling me what, what I, my, my article is
00:12:33.300
kind of nuanced in that I argue two things. One thing is there's a lot of lonely people in the world
00:12:38.760
and a lot of people suffer from loneliness and particularly old people, depending on how you count it,
00:12:44.280
you know, under some surveys about half of people over 65 say they're lonely. And then you get to
00:12:50.180
people in like a nursing home, maybe they have dementia, maybe they have some sort of problem
00:12:54.280
that makes them really difficult to talk with. And maybe they don't have doting grandchildren
00:12:59.060
surrounding them every hour of the day, maybe they don't have any family at all. And maybe they're
00:13:02.840
not millionaires, so they can't afford to pay some horse mo to listen to them. And if chat GPT or
00:13:09.440
Claude or one of his AI companions could make their lives happier, make them feel loved,
00:13:13.520
one that are respected, that's, that's nothing but good. I, I'm, I, you know, I think it,
00:13:19.420
in some ways it's like powerful painkillers are, are powerful opiates, which is, I, I'm not sure
00:13:24.960
what, I don't think 15 year olds should get them, but, but somebody who's 90 in a lot of pain,
00:13:29.700
sure, lay it on. And I feel the same way with this. So that's the, that's the pro side.
00:13:34.620
The con side is I am worried and you're, you're touching on it was this illusion talk. I'm worried
00:13:39.580
about the long-term effects of these syncopathic sucking up AIs where every joke you make is
00:13:47.640
hilarious. Every story you tell is interesting. You know, I mean, the way I put it is if I ever
00:13:52.680
ask, am I the asshole? The answer is, you know, a firm, no, not you, you're the asshole.
00:13:57.280
Yeah. And I think I'm, you know, I'm, I'm, I'm an evolutionary theorist through and through and
00:14:02.280
loneliness is awful, but loneliness is a valuable signal. It's a signal that you're messing up.
00:14:07.160
It's a signal that says, you got to get out of your house. You got to talk to people. You got to,
00:14:11.520
you know, got to open up the apps. You got to say yes, yes to the brunch invitations. And if you're
00:14:16.040
lonely, when you interact with people, you feel not understood, not respect, not love, you got to up
00:14:21.520
your game. It's a signal. And like a lot of signals, like pain, sometimes there's a signal
00:14:26.660
that where people are in a situation where it's not going to do many good, but often for the rest
00:14:31.220
of us, it's a signal that makes us better. Yeah. I think I'd be happier if I could shut off.
00:14:35.840
I generally as a teenager, I'd be happier if I could shut off the switch of loneliness and
00:14:39.560
embarrassment, shame and all of those things, but they're useful. And so the second part of the
00:14:46.060
article argues that continuous exposure to these AI companions could have a negative effect because
00:14:51.960
while for long saying you're not going to want to talk to people who are far less positive than
00:14:57.800
AIs. And for another, when you do talk to them, you have not been socially entrained to do so
00:15:02.440
properly. Yeah. It's interesting. So, so I believe in the, the dementia case aside me, I totally agree
00:15:08.580
with you that that is a very strongly paternalistic moment where you, anything that helps is fine.
00:15:14.000
It doesn't matter that it's imaginary or that it's, that it's encouraging of delusion. I mean,
00:15:18.980
this, we're talking about somebody with dementia, but so just imagine in the, the normal, healthy
00:15:24.860
case of people who just get enraptured by increasingly compelling relationships with AI. I mean, you can
00:15:31.780
imagine, so right now we've got chatbots that are still fairly wonky. They hallucinate. They're
00:15:37.500
obviously sycophantic, but I just imagine it gets to the place where, I mean, forget about Westworld and
00:15:43.360
perfectly humanoid robots, but very shortly, I mean, it might already exist in some quarters
00:15:49.760
already. We're going to have video avatars, you know, it's like a zoom call with an AI that is
00:15:56.760
going to be out of the uncanny valley. I would imagine immediately. I mean, I've seen some of
00:16:00.800
this, the video products, which, um, uh, you know, like a sci-fi movie trailers, which are,
00:16:07.580
they don't, they don't look perfectly photorealistic, but they're, they're getting close. And you can
00:16:11.760
imagine six months from now, it's just going to look like a gorgeous actor or actress talking to
00:16:16.320
you. That's going to be, imagine that becomes your assistant who knows everything about you.
00:16:22.320
Uh, he or she has read all your email and kept your schedule and is advising you and helping you
00:16:27.420
write your books, et cetera, and not making errors and seeming increasingly indistinguishable from just a
00:16:38.320
super intelligent locus of conscious life, right? I mean, it might even seem, it might even claim to be
00:16:43.700
conscious if we build it that way. And I mean, let's, let's, let's just stipulate that for at least
00:16:49.780
for this part of the conversation, that it won't be conscious, right? That this is all an illusion,
00:16:53.680
right? It's just a, it's no more conscious than your, than your iPad is currently. And yet it, it becomes
00:17:00.060
such a powerful illusion that people just, most people, I mean, the people, I guess philosophers
00:17:06.940
of mind might still be clinging to their agnosticism or skepticism by their, their fingernails, but most
00:17:13.840
people will just get lulled into the presumption of a relationship with this thing. And the truth is
00:17:20.020
it could become the most important relationship many people have. Again, it could, it's so useful,
00:17:25.740
so knowledgeable, always present, right? They might spend six hours a day talking to their
00:17:31.460
assistant. And what does it mean if they spend years like that, basically just gazing into a,
00:17:40.620
a fun house mirror of fake cognition and fake relationship?
00:17:46.660
Yeah. We've seen, I mean, I'm a believer that sometimes the best philosophy, our movies do
00:17:51.480
excellent philosophy and the movie her came out, I think in 2013 is an example of this guy, you know,
00:17:58.820
lonely guy, normal, lonely guy, but, uh, gets connected an AI assistant named Samantha played
00:18:04.880
by Scarlett Johansson and falls in love with her. And she does all of, she is the first thing she says
00:18:10.640
to him. And what a, what a meet cute is I see you have like 3000 emails. I haven't been answered.
00:18:17.480
Yeah. I fell in love with her there, but, but, you know, but the thing is you're watching a movie
00:18:23.560
and you're listening to her talk to him and you fall in love with her too. I think we've evolved
00:18:29.440
in a world where when somebody talks to you and, and, and acts normal and gives you, and has,
00:18:34.380
and seems to have emotions, you assume there's a consciousness behind it. You know, evolution has
00:18:38.440
not prepared us for these, you know, extraordinary fakes, these extraordinary, you know, golems that,
00:18:44.780
that, that, that, that elude all of the behavior associated with consciousness and don't have it.
00:18:49.840
So, so we will, we will think of it as conscious. There will be some, some, you know, philosophers
00:18:54.800
who insist that they're not conscious, but, but even they will, you know, sneak back from their,
00:19:01.480
from their classes. And then in the middle of the night, you know, turn on their, their phones
00:19:05.320
and start saying, you know, I'm lonely, let's talk. Yeah. And then the effects of it. Well,
00:19:09.720
one effect is real people can't give you that, you know, married, very happily married, but
00:19:17.380
sometimes my wife forgets about things that I told her. And sometimes she doesn't want to hear my long
00:19:22.940
boring story. She wants to tell her story instead. And sometimes it's three in the morning and I could
00:19:27.780
shake her away because I have this really interesting idea I want to share, but maybe that's
00:19:31.540
not for the best. She'll be grumpy at me. And because the thing is, she's a person. And so she has
00:19:37.160
her own set of priorities and interests. So too of all my friends, they're just people and they have
00:19:43.080
other priorities in her life besides me. Now, in some way, this is, I think what makes when you
00:19:50.040
reflect upon it, the AI companion have less value, you know, here you and I are. And what that means
00:19:55.400
is that you decided to take your time to talk to me and I decide to take my time to talk to you.
00:19:59.920
And that's a value. When I switch on my lamp, I don't feel, oh my gosh, this is great. It decided
00:20:04.820
to light up for me. It didn't have a choice at all. The AI has no choice at all. So I think in
00:20:09.460
some part of our mind, we realize there's a lot less value here. But I do think in the end,
00:20:17.100
the scenario you paint is going to become very compelling and real people are going to fall
00:20:21.640
short. And it's not clear what to do with that. Now, there's something I think you've just come up
00:20:29.040
with a fairly brilliant product update to some future AI companion, which is a kind of Pavlovian
00:20:35.920
reinforcement schedule of attention where it's like the AI could say, listen, I want you to think
00:20:43.000
a little bit more about this topic and get back to me because you're really not up to talking about
00:20:48.780
it right now. You know, come back tomorrow, right? And that would be an interesting experience to have
00:20:55.800
with your AI that you have subscribed to. I've wondered that. Like you asked the AI a question
00:21:00.840
that says, is that really like a good question? Does it seem like a question you couldn't just
00:21:04.860
figure out just by thinking for a minute? I know everything. That's really what you want to ask me?
00:21:10.320
Is that don't you have something deeper? You were talking to a super intelligent, you know,
00:21:15.460
God and you want to know how to end a limerick. Right. Really? Yeah. I would wonder if these things,
00:21:22.860
how people would react if these things came with dials, you know, obviously not maybe a big physical
00:21:28.300
dial, you want a big physical dial and the dial is pushback. So when it's set at zero, it's just
00:21:35.080
everything you say is wonderful and just, and, but I think we do want some pushback. Now, I think in
00:21:43.000
some way we really want less pushback than we say we do. And it's this way of real people too. So
00:21:48.500
everybody says, yeah, Oh, I like when people, people argue with me. I like when people call
00:21:52.020
me out on my bullshit. But what we really mean is we want people to push back a little bit and then
00:21:56.800
say, ah, you convinced me, you know, you, you really showed me or, you know, I thought, I thought
00:22:01.780
you were full of it, but, but now upon reflection, you've really out argued me. We want them to fight
00:22:07.100
and then, and then acquiesce. But, but I turned the dial even further. We'll say, you know, we've been
00:22:13.180
talking about this for a long time. I feel you are not very smart. You're just not getting,
00:22:16.860
I'm going to take a little break and you, you, you mull over your stupidity. Yeah.
00:22:21.140
Paltry human. I don't know. Recalcitrant style. That's a, that's what we could build in. All
00:22:27.360
right. We're going to, I feel this could be the worst business. We're going to get rich,
00:22:30.260
Paul. This is with AI that calls you on your bullshit. That's, that's really the business
00:22:35.460
they do this century. But so what are we to think about this prospect of, of spending more
00:22:43.940
and more time in dialogue under the sway of a pervasive illusion of relationship, wherein
00:22:50.560
there is actually no relationship because there's nothing that it's like to be
00:22:54.340
chat GPT-6 talking to us perfectly. If you'd like to continue listening to this conversation,
00:23:00.860
you'll need to subscribe at samharris.org. Once you do, you'll get access to all full length episodes
00:23:06.960
of the Making Sense podcast. The Making Sense podcast is ad free and relies entirely on listener
00:23:12.660
support. And you can subscribe now at samharris.org.