John Donvan

Hi everybody, I’m John Donvan, and welcome to Agree to Disagree from Intelligence Squared.
And about that telephone in your pocket. Well, I, actually, let me use much more up-to-date term
the smartphone in your pocket. And yes, it is smart in the sense that it really does have a certain
kind of functional intelligence going on behind that little screen. Artificial Intelligence, AI. And
thanks to AI, your phone is not some static, you know, lump of silicon, it is actually adapting to
you all the time. It is studying you and it is learning from you to the point where it, it’s actually
now advising you on routes and restaurants, and it’s reordering your photos by subject matter and
whose faces are in them. And it’s lining up news stories for you to read.

And in so many ways that I expect you appreciate your phone and its artificial intelligence, they
are helping you and your relationship with all of that, because let’s face it, it is a relationship
between us and AI. Is it good and healthy and problem solving, or does it become worrisome that
the AI has so much info on you and is practically speaking, deciding things for you, and that AI
itself is just moving so fast into so many parts of life, way, way past the phone in your pocket
and into the, the infrastructure of everything? And also that you don’t really know who’s behind it
all and where it’s all headed.

So that, yes, you’re experiencing some of the benefits of AI right now already, but you may also
be wondering in the longer run, will artificial intelligence do more harm than good? Well, that is
exactly what we are asking in this episode, word for word. That question is one that later in the
program we’re gonna actually put to an AI engine and see what answer it comes up with. But
first, we have two humans who are gonna debate from opposing points of view on the question,
will artificial intelligence do more harm than good? We’re joined by Daron Acemoglu, an
economist at MIT, and Darrell West, Vice President and Director of Governance Studies at the
Brookings Institution. Daron and Darrell, thanks so much for joining us on Agree to disagree.

Darrell West

Thank you. Nice to be with you.

Daron Acemoglu

Thank you.

John Donvan

To get things started, we have this central question and we just wanna know who’s the yes and
who’s the no. So on the question once again, will artificial intelligence do more harm than good?
Daron on that question, are you a yes or are you a no?

Daron Acemoglu

I am a yes, and I think AI will do more harm than good on its current path because it is being
developed in a particular way that deepens social, economic and political inequalities in society.

John Donvan

Thank you very much. And that tells us where you’re gonna be on this question, Darrell, but I
wanna do it as a formality. Darrell West on the question, will artificial intelligence do more harm
than good? Are you a yes or are you a no?

Darrell West

I am the no, because I think AI can relieve humans of boring, dirty, and dangerous activities.
There are gonna be tremendous advantages if we get the policy right. I mean, that’s really the
crucial part here.

John Donvan

All right, thank you. So we know who’s yes and who’s no. And I want to go back to, uh, Daron,
you took the yes side, which is yes, do more harm. So take a couple of of minutes and tell us why
you’re on that side.

Daron Acemoglu

So let me sort of first lay the scene by saying, you know, AI is a bunch of new technologies,
extensions of existing digital technologies in many different dimensions. And of course, any new
technology arrives in the middle of existing conflicts, ex- existing power imbalances between
companies, governments, individuals, workers, citizens, users. And if you looked at it at that
perspective, you would start thinking about, well, how is it going to influence those existing
relationships? And who is going to decide the future path of AI and for whose benefit? So the
thing that really worries me is that AI on its current path is in the hands of already powerful
companies and governments and, you know, forget about, you know, super intelligence and so
on. But artificial intelligence as construed right now is a lot about collecting a lot of data and
developing algorithms based on that data.

Now, as the familiar saying goes, knowledge is power, information is power. So who controls
data is very much embedded in the existing economic and social power relations of society. And
if you give more data to the already very powerful corporations, the already very powerful cogovernment such as the Chinese government, you know, this is going to have an impact on how
we are going to divide the gains from new technologies and how we’re going to change the
power dynamics in society both economically and socially. And in fact, when you look at the
details, what you see is that, you know, AI is sometimes sort of presented as reaching human
level intelligence and amazing capabilities.

Sure, perhaps in 200 years time scientist think it could happen, but today it’s really about three
major activities. It’s about advertising, online digital ads, it’s about surveillance. You actually
watch people, monitor people, learn more about their political views, what they’re doing at work,
what they’re doing in their free time, and it’s about automation. And all three of these deepen
existing inequalities in society. And unless you take other measures to counterbalance them, all
three of them are going to have negative consequences on many people. So you see that with
data collection, you see that with surveillance and monitoring. Automation is the more
complicated one because it’s a key economic activity that’s been going on for 250 years at least.
But I will argue, and we can come back to this later in its current form, this type of automation is
also not, not generating benefits for workers in general.

John Donvan

Okay. You’ve given us a lot to start with there, but before we start digging into the points that
you’re making, I wanna let, uh, Darrell West tell us why he says no to the idea that it’s doing
more harm than good. So it’s your turn now, Darrell.

Darrell West

Thank you. Uh, I guess I am more optimistic about the future of AI because all new technologies
almost always create problems for society. You know, when you go back and look at the
printing, press, television, automobiles, any new technology, there always are problems either in
terms of human safety or lack of understanding about the, uh, technology. And, and certainly we
see a lot of that with, uh, AI today. But what always happens is humans have the capability to
control the technology through laws, policies, and regulations. That’s been the history of
technology innovation, and that will be the future of AI. Now, right now, we have had, for the
last several decades a libertarian stance on technology in which we have outsourced all the major
decisions to private companies.

And they have an enormous amount of data. They have misused the information. Uh, there’ve
been problems of racial bias, a lack of fairness, uh, and certainly inequality has emerged from
that. But it doesn’t have to be that way. It’s actually not the fault of the technology, I would argue
it’s the fault of our public policy. I think with new laws, a new social contract, more regulations,
we can actually adapt AI to human needs and move AI towards the public good in the same way
that we did with past, uh, innovations. Like cars were dangerous things, and we came up with
safety belts, uh, airbags and other, uh, safety, uh, mechanisms, uh, and we’re able to reduce the
fatality rate. Uh, we have problems of AI in terms of fairness, bias, and inequality, but we can
alter our tax policies and our social policies to deal with some of those issues and move AI much
more towards the public good.

John Donvan

So what, what I’m hearing each of you saying there’s some, a certain amount of common ground
in that you’re, you’re both saying there’s AI, uh, that’s doing a certain thing that could be harmful.
And Darrell, you’re saying, well, it, it won’t be that harmful if we take measures. And Daron, you
are saying it’s already doing harm and it’s gonna continue to, to do harm unless we take
measures. But what I think I hear being the difference between the two of you is a little bit of a
sense of where we are now in this journey. And Darrell, you’re pretty optimistic that things are
moving in the right direction or likely to move in the right direction. And Daron, you sound
much less optimistic about the moment we are and the trend lines that you see about how things
are unfolding.

So I, I want to talk, you to talk a little bit more about that and what you said in your opening
statement, and especially in terms of, of AI being an engine for inequality and, and and, and why
not be more optimistic in the way that Darrell is that yeah, there are challenges, but we can
address them, we can figure it out. Why not be more optimistic about that?

Daron Acemoglu

Well, I think, uh, you’ve nailed the key point of disagreement. There is a lot of agreement
between what Darrell and I are saying, but there is one subtle difference, which is that I think
Darrell, as several other, many other economists and social scientists, believes that somehow
there are automatic mechanisms for humanity to adjust its policies, its norms, its institutions, so
that we get the better of the new technologies. Whereas my own read of history, my own read of
evidence is that, that sort of adjustment mechanism is much more haphazard and it often requires
major political change and it sometimes failed, it sometimes succeeds and you cannot bank on its
success. And in fact, I would go further and I would say that the moment we start banking on the
success of these adjustment mechanisms, be they policies to redistribute or change the
governance of data or companies or automation, I think the moment we take those for granted,
we make it much more likely that the current path of AI that is inequality boosting, that is not
actually bringing much productivity benefits continuous.

John Donvan

Darrell, your response?

Darrell West

I mean, I do not think there is an automatic mechanism, uh, that is going to correct the abuses of,
uh, technology, but I do believe that humans have the capability to make different choices and
enact different kinds of policies. And in fact, I think we are actually in the early stages of starting
to do that. So as I mentioned before, the last several decades we’ve been libertarian, but there
now is a tech-lash that has emerged, a backlash against technology, against the technology
companies and against the dominant market power of these, uh, companies. And we are starting
to see new laws, new regulations, and new policies designed to deal with particular problems. So
for example, uh, take the issue of privacy.

Like I think, uh, both, uh, Daron and I would agree, uh, that, uh, companies have too much data,
they can misuse the, uh, data, the threat to privacy is, uh, very real. But the state of California
passed a law to impose a lot of new restrictions on technology companies. And California is such
a big state that the tech companies can’t really devise AI for California that’s different from Utah
or Nevada. So when California cracks down on, uh, privacy abuses or the European Union
cracks down on privacy abuses, which, uh, they are, uh, currently doing as well, essentially the
tech companies have to come along and essentially devise better AI that really helps everyone.

And so I think that just really illustrates the fact that at the state and local level, we’re already
seeing a lot of, uh, action, uh, in moving towards greater, uh, regulation. Uh, there’s more
regulation of the gig economy. Uh, there’s more, uh, cities who are starting to regulate Airbnb
rentals because I think it’s harmful to their local communities. So this tech-lash is starting to
change public opinion into public policy, and I think that’s the source of my optimism.

John Donvan

I’m John Donvan, this is Intelligent Squared US. More when we return. Welcome back to
Intelligence Squared US, I’m your host, John Donvan. Let’s return to our discussion.

Daron, Darrell is also saying we’ve been down this road before with technologies. I’m thinking of
obviously automobiles and electricity and, and railroads, um, maybe nuclear power, um, but um,
that, that we sort of got it figured out, and that’s where the optimism is coming from Darrell. And
I wanna know if you think there’s a, a, a difference of, of magnitude or qual- or, or, or kind in the
challenges you’re talking about with AI versus, say, the adoption of electricity or the
automobile.

Daron Acemoglu

Well, to make this less interesting, I think I agree a lot with the, most of what Darrell said. I think
his beginning statement for the last round is, is very much my opinion as well. I think the only
place where there is some residual disagreement is in Darrell’s account of how we have adjusted
to past breakthrough technologies. I would not be as positive about how that adjustment took
place. And I would say it was much more painful, it was much more haphazard, and it was much
longer. Let me give some examples. You know, before the industrial age, we had the medieval
agricultural economy, contrary to what is sometimes implied with labels such as dark ages. The
Middle Ages had a lot of new technologies that improved agricultural productivity massively.
But if you actually look at what happened is that in most cases, these technologies did not
improve the living standards of peasants, farmers, they deepened inequality.

Sometimes they actually reduced the living standards of, uh, uh, of, of the majority of the
population. For example, with much better mills, uh, you could improve many parts of the
production process, but then the mills were the monopoly of religious authorities or the lords that
then compel people to pay exorbitant prices for using the mills and could not, would not even
allow them to use their hen mills. So this actually went on for several hundreds of years before
sort of you see a systematic change in who has property rights, who controls new technologies,
who has rights. And then you come to the industrial age and you see the same thing playing out.

You know, the new technologies that spearheaded the British Industrial Revolution were mostly
in the textile industry at first. Those not only just led to much greater inequality, but also lower
wages, much longer working days for most textile workers, also to much worse conditions in
factories in cities. And this lasted more than a hundred years before things started improving. So
there is nothing rapid about it. And if you look at the details, there’s nothing automatic about it
now. That’s what worries me.

John Donvan

Well, let, let’s talk a little bit about, you were talking about the impact of technologies, uh, on,
on, on the livelihoods and the, the quality of life of individuals hundreds of years ago. I wanna
talk a little bit now about the impact of artificial intelligence on the livelihoods, uh, and life, uh,
quality and the workplace of, uh, the average American. So Darrell, you said at the outset, um,
that, that, um, AI kind of relieves people of certain dreary jobs, which sounds great. Um, how far
does that go?

Darrell West

Well, technology can relieve people of, uh, mundane, dirty and dangerous jobs. So, for example,
when there’s a nuclear accident, like we used to have to send people in to kind of monitor the
radiation, uh, levels, uh, now we routinely send in robots with cameras, uh, with sensors. They
can basically collect the same, uh, data without endangering humans. So that’s an example of
how technology can actually improve the human life, uh, and, uh, kind of protect people in the
process. But I think there also some economic advantages that, uh, people often ignore. So for
example, cell phones are becoming ubiquitous and smartphones are becoming, uh, more, uh,
common around the world as well, including in developing nations.

Mobile technology has enabled farmers in remote areas of Africa to basically expand their
markets. Like in the old days before technology, essentially people who fished, uh, people who
farmed in, uh, rural developing, uh, nations, basically were limited to their very local market in
order to sell their wares. Today because of mobile technology, they have access to much greater
information outside their local city and their local community. They can sell their wares, uh, to a
broader, uh, uh, group. And so it’s an example of how technology can better the lives of people. I
wouldn’t argue it’s going to, you know, solve the inequality problem because you know, that’s
deeper and more, uh, fundamental and there’s a lot more going on there. But technology can be
one of the tools that can provide more information to people, expand their markets, and therefore
create greater economic opportunities.

John Donvan

So, Darrel, you, you just mentioned that you don’t think technology is solely to blame for
inequality. So I want you to talk a little bit about that. Um, w- what, I know it’s very, very
difficult to put a percentage on it, so I’m not gonna ask you to do that. But if you can give us an
overall sense is, is AI, uh, in, in any way, in the way that Daron is explaining it already an engine
for inequality compared to other things that are causing inequality?

Darrell West

It is currently an engine for inequality, but I would argue a bigger source of inequality is our
completely inegalitarian tax policies and our very poor social welfare, uh, policies. I mean, there
are very large and very rich tech companies that are paying five to 10% on their, uh, corporate
tax rates. Like that is ridiculous. That’s completely unfair. That contributes to the inequality. So
if we want to deal with inequality, there are tax policy and social welfare policy changes that
would mitigate the impact of that. Like, if you look at the experience of America, like the period
from like the 1920s through the late ’40s was a period of pretty high, uh, income inequality. But
then through public investments and infrastructure, tax policy changes, social welfare policy
changes, we developed social security, unemployment compensation, and other things that the
period from, let’s say 1945 through the 1970s, America actually got more egalitarian.

But then over the last 40 years, basically since the 1970s, our tax policies moved in a very
inegalitarian manner. And so tech, uh, and so inequality, uh, that expanded. But it’s not
necessarily the technology that caused that. It’s the fact that we don’t tax these companies. We
don’t tax their income or their wealth. We don’t regulate them. Uh, we don’t have, uh, social
welfare policies that help people make the adjustment to automation. So the, the inequality does
not have to coexist, uh, with, uh, technology. If we pair technology with the proper tax policy
and social welfare policies, we can actually bring down inequality.

John Donvan

Well, let me bring that back to Daron then. What, what Darrell is saying is that your argument
for inequality might be overstating the impact of inequality in the overall equation of what’s
causing inequality.

Daron Acemoglu

Well, I think those are complex questions, and there is much, I agree with what Darrell said. I
think our tax policies are horrendous. You know, there are institutional reasons for inequality. In
the US, social safety net is very weak and inefficiently organized, completely agreed. But if you
look at the data, at least the way I have, and you know, of course there are different choices you
can make in approaching the data, but my own research here as well as research by others,
indicates that automation has been a major source of inequality and stagnant wage growth over
the last four decades.

And it’s not so surprising in some sense what automation has done. And this is all pre AI, by the
way. And the question is how AI is continuing this trend, and we can talk about that. But what
automation has done is that it has enabled companies to use equipment, industrial robots for tasks
previously performed by blue collar workers and factories and software, digital tools for tasks
previously performed by office workers. And if you look at where inequality comes from, it’s
actually driven very largely by the following incomes of workers, demographic groups that used
to specialize in those tasks that have been automated over the last several decades. And in fact,
that’s exactly what the historical record also shows.

In other episodes, you also see that automation tends to increase inequality, tends to depress
wages. This does not invalidate the point that Darrell made. He gave a brilliant example. For
example, from the work by the economist Rob Jensen, about how mobile phones have helped,
uh, fishermen in, in Kerala in India. But when you look at that, the remarkable thing in that
example is exactly that the use of mobile phones to help fishermen is not automation. It’s actually
you are augmenting what fisherman used to do. You’re providing them new tools to make better
decision, acquire better information, and you’re actually creating new tasks for them.

John Donvan

But, but isn’t that

Daron Acemoglu

That’s, that’s the key

John Donvan

… isn’t that actually

Daron Acemoglu

… key issue. If I can just say one more sentence.

John Donvan

Sure.

Daron Acemoglu

I think that’s the key issue when it comes to the labor market implications of any new
technologies. How much of it is just automation? How much of it is creating new capabilities for
workers? And I think the current path of AI building on how we have used digital technologies is
going too much in the automation direction.

John Donvan

Why do you think that? T- talk, talk about a couple of industries where you think that’s
happening.

Daron Acemoglu

Well, I mean, the best example is, uh, manufacturing. You know, if you look at-

John Donvan

Well, let me, let me ask you. I, I think manufacturing is also the easiest example. But, but

Daron Acemoglu

Yes.

John Donvan

… what, what about other kinds of professions that, where people feel they might be sitting
comfortably right now, are you saying that they’re also at risk of losing their livelihoods? Not
having their work enhanced by a, a souped up tool, but having their work replaced by, um, by an
artificial intelligence that can lock, stock and barrel do everything that they were doing for a
living before?

Daron Acemoglu

Yeah, I think there’s been a lot of, that has already happened before AI. If you look at where
automation has been, uh, you know, there’s rapid automation on factory floors, but you know,
even, uh, as far back as 1980s, 1970s, you know, a very small fraction of the US workforce was
in blue collar jobs. A lot more of the impact of automation has been in office jobs, clerical jobs.
And you know, if you look at, uh, in the 1930s, 1920s, 1930s, uh, 1950s, 1960s, you see clerical
jobs or blue collar manufacturing jobs, those were the ones that were engines of wage growth in
the 1940s, ’50s, and ’60s. And that’s because there were many organizational and technological
choices that were expanding the functions that these workers were performing, creating new
tasks for them, new capabilities for them.

And now, fast-forward to the post 1980 era, and you see that a lot of the use of digital
technologies is focused on eliminating these workers or eliminating at least a subset of the tasks
that they’re performing. And that’s the cracks of the matter about how we use technology creating
inequality or actually helping workers.

John Donvan

All right, lemme take that to, to Darrell. And, and Darrell, I’m, I’m sure you say, you’re gonna tell
us it doesn’t need to be that way, but also to some degree, I’m sure you’ll, you’ll concede it has
been that way. So I’d like to hear how you process your response to that.

Darrell West

Uh, Daron is, uh, completely correct, uh, that there are going to be lost jobs and there have been,
uh, lost jobs, uh, due to, uh, technology and automation. I mean, we can clearly see that in retail,
in office jobs, in finance, uh, manufacturing, and in other areas. But a few years ago, I wrote a
book entitled The Future of Work, and I argued, and John, you’re exactly right, it doesn’t have to
be that way. I made the argument that we, in this era of digital transformation, we need to invest
in worker retraining, in adult education.

We need to embrace lifelong learning, kind of the old education model where we invest in
education up through about age 25, and then people are on their own after that has to give way to
new models where basically people re-skill at ages 30, 40, 50 and 60 virtually, uh, throughout,
uh, their entire lives. So again, it’s not the technology that is creating the problem, it’s our lack of
good tax policies, social welfare policies, education policies, et cetera, et cetera. So we need a
new social contract. And if we do those things, I’m actually still very optimistic about the future
of AI and technology in general.

John Donvan

So I mentioned at the beginning of the program that we went to an AI engine, uh, called
Anthropic, and we actually entered the question that we’re debating here, will artificial
intelligence do more harm than good? And it gave us an answer. (laughs), I’m gonna read it. Ai…
this is the answer. I’m, I’m now quoting, I will not do the robot voice, but I’m so tempted.

Daron Acemoglu

(laughs).

John Donvan

“Um, AI has the potential for both positive and negative impacts. AI can help solve many of the
world’s problems and improve lives with applications such as disease diagnosis, increased
productivity, and automation of dangerous jobs. However, there are valid concerns around
privacy and bias and job displacement. There are also longer term concerns about the existential
risks associated with advanced AI, such as the control problem and the misuse of AI by
malicious actors. Overall,” now it’s reaching a conclusion, it’s taking sides in this part. “Overall,
AI will likely have a net positive impact, but in order to ensure this, we need to be proactive
about addressing the risks and continue to develop AI responsibly.” Well, it sounds like

Daron Acemoglu

Are you sure Darrell didn’t write this?

John Donvan

Yeah, it sounds (laughs)-

Darrell West

That is exactly my view. I completely agree with that. (laughs).

John Donvan

(laughs). Uh, I, I’m, I’m just curious about your take that we’re actually listening to an AI answer
that was constructed in pretty darn good English and made sense. Um, what, how does that, how
does that just hitch, hit you, uh, Daron, that, that, that a, a machine made up of wires and lights
and flashing silicon can do that? Somebody’s gonna make so much fun of what I, how I just
described (laughs) the circuitry

Daron Acemoglu

No, I mean, I-

John Donvan

… but go for it.

Daron Acemoglu

You know, o- of course, many of the capabilities we are developing are quite impressive. The
cameras on our phones are amazing. Uh, the remote work communication technologies are just
out of this world. You know, you have much better imaging possibilities, much better ways of
data crunching. I think all of those are very real. That’s the reason why I agree with Darrell that
we have the means to expand our capabilities. I also tend to think that there is often quite a lot of
hype about some of these technologies. You know, if you ask these fancy sort of machine
language models questions that can be built on the basis of things that they have access to from
the internet, they can give very sophisticated answers.

But if you probe and try to understand whether they actually comprehend what they are saying,
even the concepts, I think the, uh, the, the case is very clear that there’s a huge lacuna of, of
understanding, and it’s just like parroting things that you sort of get from the web and put
together or from their training data and, and put together. So I’m not actually, uh, of the view that
we are close to anything like human capabilities in most tasks. But on the other hand, data
crunching can get you very far. But the question again comes down to what are you going to use
that data crunching for?

John Donvan

D- Darrell, wh- why, why is AI so good with data? What, what advantage does it have over the
human brain?

Darrell West

Well, its main advantage is just the tremendous, uh, processing power that it has. And as super
computers get even better, that processing power, uh, is going to, uh, improve as well. But I
mean, what impressed me about the AI engine answer is AI and machine learning are actually
good at summarizing what other people have come up with because it’s basically just a more
sophisticated form of content analysis that we’ve had around, uh, for, uh, decades. You know, it
can look at the literature, uh, look at books, look at articles, and basically kind of pull out what
are the most frequent types of arguments.

That’s completely different from human intelligence. Uh, that’s kind of almost a rote type of
thing. I mean, the future AI is going to develop more capabilities to learn, uh, to adapt, uh, and to
basically make, uh, more informed decisions, uh, based on data. But I do think the part of the
answer that we do need to worry about is the data control part of, uh, that. It is a huge problem,
uh, right now, basically because of the dominant market position of the large, uh, technology, uh,
platform, uh, companies. And so that’s the part that we need to get a handle on.

John Donvan

I’m John Donvan, this is Intelligent Squared US. More of our conversation when we return.
Welcome back to Intelligence Squared Us. Let’s get back to our debate.

So where do you see us being in 20 years on all of these issues? You know, on the whole, again,
the, the question is whether we think AI will do more harm than good. So we’re talking about on
balance, and it’s obvious now from both of you. We’ve heard about both benefits, uh, real and
soon to come, and also risks. But you know, if you look at some of the potential benefits that AI
can, can make people healthier, that the, that, uh, disease can be spotted sooner, that potentially,
if the dream of, uh, auto, of self-driving cars comes true, that AI can make the roads safer. Those,
those are all good things.

On the other hand, what happens to the taxi drivers and what happens to the radiologists and
what happens potentially to some of the other m- members of the medical profession? On the
whole weighing up the, those sets of benefits and, and, and, and risks Daron, what, where do
you, where do you see us going in 20 years, being in 20 years on the trajectory that we’re on right
now?

Daron Acemoglu

Uh, again, I don’t wanna sound like, I think it’s an open and shut case. I do believe, like Darrell,
that we can change course, we can change direction, but I think it’s a more systemic change that
is needed. Do I believe that such a systemic change is likely in the current US political
environment? Unfortunately, I’m not super optimistic, although I still hope we can do better. I
also think that though, there are technologies that are going to be very promising in terms of
disease detection and a few other things like protein folding. Some of the promises of AI are
exaggerated.

Many of the simplest routine tasks were already automated by using simpler digital tools. We are
now entering the field of human judgment. And I don’t think AI is going to make such rapid
progress that in 20 years time, we’re gonna see amazing productivity improvements. In fact, one
of the remarkable things about the data is that the age of digital technologies has been one of
relatively slow productivity growth. And I think it’s going to continue that way unless there is a
complete overhaul of how we use these digital tools, which again, I hope is feasible, but I don’t
see it as likely.

John Donvan

So, so that’s interesting, Darrell. I think what I’m hearing Daron saying is that in addition to the,
the tension with the policy choices that are made, the, to the, uh, establishment of guardrails to
keep the harm from outweighing the good, also just the suggestion that the actual wizardry of
tech, of, of AI has been over-promised that, uh, the, the potential benefits are gonna take a long,
long time to come.

You know, we were talking, I, I know I was on panels in 2016, 2017, where there were
predictions that there were gonna be a lot of self-driving cars all over the highways by 2020, and
now it’s two years after that, and that hasn’t happened. And so that question of whether the
promise is over promised, uh, is, is out there. And I think Daron has just put that out there, and it
goes, I think it challenges optimism to a degree.

Darrell West

Uh, there certainly has been a lot of hype, nobody can really, uh, deny that. But, uh, the example
on self-driving cars, uh, there was a prediction that by 2020, uh, they would, uh, be present at
least in, uh, major, uh, cities. Uh, that has not happened. If you take a 10 to 20 year time horizon,
though, I do think self-driving cars will become, uh, much more prevalent. It’ll probably start in
the taxi business and the ride-sharing services and in the, uh, trucking, uh, sector before it
becomes a mass consumption, uh, type of thing. But it will then disseminate from those areas.

Now, if we actually get to the point where self-driving cars do become more prevalent, right now
in America, there are about 44,000 highway fatalities every year. 90% of them are accidents
involving drunk driving or distracted driving. Like human error is the big source of these 44,000
fatalities. Now, autonomous vehicles are gonna have their own problems. Uh, privacy actually is
gonna be a huge problem because autonomous vehicles collect an enormous amount of data, but
autonomous vehicles are not gonna get drunk, and they’re not gonna get distracted. If we had a
lot of autonomous vehicles on the road, we would not have 44,000 fatalities. Now, it would not
be zero. Uh, it would be somewhere in between, but there would, uh, but our highways would be
much, uh, safer due to the technology.

John Donvan

All right, so that’s a good, that’s, that’s a pretty optimistic prediction for 20 years from now. I, I
wanted to mention before, when we were talking about taking that question to, uh, the question
of our debate to the artificial intelligence, um, at Anthropic, that we actually did a debate
ourselves in 2019. I moderated a debate, uh, under the aegis of Intelligence Squared between a
human being and IBM’s technology called Project Debater. And we filled a hall, hundreds of
people watched, um, a piece of technology in about 15 minutes, put together, uh, an argument on
a question about providing, um, uh, preschool for children as I, as I recall.

And a human was given 15 minutes to prepare his argument. And then the two of them debated
for about 20 minutes and went back and forth and back and forth with the, uh, the technology
speaking and a voice that I would identify as female, um, making very cogent arguments, even
making some jokes that were programmed into it. And in the end, the audience voted and
decided that the human had been more persuasive, uh, which may show certain human bias, but
it, it was, it was not an overwhelming vote. Um, the, the, the, the computer did pretty well. And
I, I’m bringing that up, first of all, to recommend it’s worth going and listening to on our selection
of, uh, of, uh, podcasts, but, in radio programs.

But I’m bringing it up also because we at Intelligence Squared to think of debate as being so, so,
so central to democracy, to the functioning of democracy, to the expression of ideas, to the
challenging of ideas to free expression, period. And I know Daron, that one of your big concerns
about the negative impacts, as you would put it, of AI as we have, are experiencing it already, is
that it is corrosive to democracy. And now that we’ve put democracy into the conversation, I
really want to ask you to take on that, that point of view and explain where you think AI is, is
being damaging to the very fabric of our democracy. Now and potentially.

Daron Acemoglu

Well, thank you, John. I’m wondering when that was gonna come up.

John Donvan

Here, here it is.

Daron Acemoglu

Yes, absolutely. Absolutely. I think, uh, you know, we need to take a step back. There is, as
Darrell said, always a danger that new technologies are going to disrupt existing methods of
communication, discourse, equilibrium in social affairs. And you see that with online
communication in general before AI. But I think the most pernicious effects of AI on democracy
are linked, again, consistent with Darrell’s emphasis to the business model of tech companies,
which is based on digital advertisements trying to monetize data and information about people
who are taking part in that communication. And it has led to an environment in which you have
many problematic aspects for democratic discourse.

Democratic discourse requires that people have trust in the quality of the information that they
receive, that they hear opposing voices, they can have incentives and means for reaching
compromise. They can have small group interactions, which acts as a school for self-governance
and democratic participation. And I think the current business model centered on digital
advertisement undermines all of these. It’s not just the echo chambers. That’s one aspect of it. It’s
not just the extremism sort of getting a lot of mileage, but the whole, uh, sort of erosion of civic
communication and participation that is coming together with this business model and the tools
that are amplifying this, this business model that’s a problem.

John Donvan

Where are those tools functioning? Gi- give us just a few examples of where the algorithms are
doing that.

Daron Acemoglu

I think the, you see that both on sort of, uh, bilateral communication platforms such as, you
know, uh, Facebook where you are sort of directly sending messages to, uh, a group of followers
versus platforms such as YouTube, where you are sort of broadcasting messages. In both cases,
if you look at the business model, it’s based on trying to maximize engagement. And, uh, the
reason for that is very clear, if people are engaged, that they’re gonna be more receptive to digital
advertisement, which is the source of the, the most major source of revenue, overwhelming
source of revenue for these companies.

And in both cases, you see that that acts as a way of undermining any type of reliable
communication. It amplifies extremism and more sensationalist content, uh, in, in the YouTube
case that is going to to be videos with more sensational material or bigger claims, or sort of more
extremist content, in Facebook cases gonna be related to the way the feed is presented and
amplified and boosted. And in both cases, the evidence is building up that this has come with a
variety of anti-democratic consequences.

John Donvan

And, and are you saying that the AI is figuring out, you know, obviously what you’re saying is
that these, um, algorithms are trying to push our buttons. And is the AI figuring out even better
ways to push our buttons because it has access to all of this data than any, than any human
would?

Daron Acemoglu

Absolutely, absolutely. You know, I think, you know, advertisements again, are as old as, you
know, marketing and industrial production. But if you look at how the ad business worked before
it’s a broadcast, it was a broadcast model. So you weren’t targeting the ads to a specific individual
and, and you weren’t using detailed information about how that individual’s attention is going to
be the drone, how his or her buttons are gonna be pressed, how you can maximize that
individual’s enragement, emotional reactions so that they end up spending more time. All of
these again, are not, you know, this is not like super intelligence or anything like that. It’s really
simple algorithms, but with enough data about how individuals react to various different types of
cues, you can do that quite successfully.

John Donvan

All right.

Daron Acemoglu

The sort of Cambridge Analytica is just one little example of that. If you look at the algorithms
involved in Cambridge Analytica, they’re pretty mundane.

John Donvan

D- Darrell, um, what’s your response on this? I, I, again, I suspect you’re gonna say there’s a lot
of truth to this.

Darrell West

Well, I’m not gonna defend the AI of social media companies because clearly we have an
extremely toxic information ecosystem. Uh, Daron is completely, uh, right about the way social
media fuels polarization, extremism, radicalization, and sometimes even outright, uh, violence.
So that clearly is problematic. Uh, I actually have a new, uh, Brookings book out entitled Power
Politics, that does look at threats to democracy. The threats that I worry about beyond technology
are voter suppression, rigged institutions, gerrymandering, and this archaic institution called the
Electoral College, which basically has allowed republicans to lose the presidential popular vote
most, uh, of the time in the last 30 years, but actually win the electoral college, uh, 40% of, uh,
the time and therefore dominate policymaking. So I make that point just to, uh, make the broader
argument that we need to avoid tech determinism. There clearly are a lot of problems in our
society, but they’re not technology.

John Donvan

Can you define that term?

Darrell West

Yeah. Uh, what I mean by that is there clearly are many problems, uh, in our society, but there
not necessarily based on the technology, you know, inequality I think is based on tax policy and
social welfare policy. The threats to democracy, although technology is making the social media,
uh, atmosphere, uh, much more toxic, they’re much more fundamental threats there in terms of
state legislatures passing rules that are basically making it more difficult to vote. How our
political institutions are rigged against majority public opinion, uh, gerrymandering is skewing
the results of house elections and state legislative elections. So the, the problem with tech
determinism is it actually makes people feel powerless. Because if we’re blaming the technology,
then the lesson that people draw from that is the technology is so powerful, we can’t do anything
about it. And in all of my writing, I always try and make the argument that humans are not
powerless. We are actually powerful, we can control the technology.

Daron Acemoglu

But if I could add just one thing, I mean, I think, uh, the argument that I try to articulate is not
tech determinism. I emphasize specifically that this is a function of the business models that are
prevailing. So absolutely, I don’t think there is anything deterministic about tech. But there is one
aspect that I would like to reiterate. Unless you change the business model, which is very, very
difficult, and it’s again, part of that systemic change that I loosely referred to before, you are not
going to be able to avoid those lopsided effects of data concentration, data control, and
monetization of that data that takes place online in social media and other platforms. So I think
it’s not tech determinism, but it is a set of pernicious effects from the current use of tech based on
a specific set of business practices.

John Donvan

Darrell, you mentioned a few minutes ago, we are powerful and, and, and, and we can control
this technology. Which just throws, throws the door open to the conversation that comes out of
science fiction that I’m guessing neither of you (laughs) thinks we should need to take
particularly seriously, but I gotta put it out there. That whole, that whole notion of the cyborgs
taking over that whole notion, you know, Elon Musk warning about the day will come when the
artificial intelligence really achieves consciousness and then achieves, uh, e- egocentrism and
then just achieves ambition and then achieves greed and doesn’t, and wants us out of the way.
Um, is that just crazy, fun, scary talk? Or is there something real there? And I’d like to ask each
of you that question, but go, go to you first, Darrell.

Darrell West

Uh, Elon Musk is just completely wrong on this as he is in many other areas. I am not worried at
all about cyborgs, super intelligent robots or AI eventually enslaving, uh, humanity. Uh, the,
when you look at AI today, it’s actually not that good. Most AI is what I would call single
purpose AI, meaning it can be really, really good at doing one thing like tabulating data or
analyzing the content of the scientific, uh, literature, uh, things like that. There actually is not AI
that is general purpose AI. Like, it’s hard to have an algorithm that is good at like 10 different
things or 100 different things the way human beings are.

And so the fear that they’re gonna be these super intelligent bodies that are like good at a bunch
of different things, maybe in 500 years we’ll need to worry about that. But today what I worry
about is fairness, racial bias, lack of transparency, impact on human safety, inequality, privacy. I
do not worry, uh, at all about cyborgs or super intelligent robots taking over.

John Donvan

(laughs). And what about you, Daron?

Daron Acemoglu

Well, 1- 100%, 200%, 300% agreed with Darrell. Darrell’s completely, right. There is no feasible
path along the current trajectory to anything even remotely looking like super intelligence. But I
would actually add, I would amplify that point. I actually believe that all this talk of super
intelligence is actually misleading us and taking, making us take off our guard against the more
relevant problems with AI, in two ways. And I’m not saying this is a conspiracy. I’m not, this is
not, nobody’s conspiracy. I hate conspiracy theories. It’s just something that has evolved in the
context of the current tech industry, current tech barons, uh, way of thinking about it.

John Donvan

Well, I’m glad, I’m glad I asked that question. I was a little bit hesitant because I thought you had
both, uh, laugh it out of the park, but you both took it seriously and actually brought some insight
to it as, as you did, um, throughout this whole conversation. Um, I, I, I, it was very interesting to
hear how much you agreed, um, but you had that essential disagreement on optimism about the
moment that we’re in and where it’s leading us to in the world of, uh, artificial intelligence
becoming further integrated into our lives and into the infrastructure of everything we do.

But I really wanna say that I appreciated how you listened to each other and showed respect for
each other, because that’s what we aim for in Intelligence Squared. So, uh, Daron Acemoglu and
Darrell West, thank you so much for joining us for this conversation.

Darrell West

Thank you very much.

Daron Acemoglu

Well, thank you, John. Thank you, Darrell. This was really fun and, uh, I’ve learned a lot. Thank
you.

Darrell West

Yep. Agreed.

John Donvan

And the conversation that you’ve just heard perfectly captures why we do this. Y- you know, that
the way the discourse is happening across the land these days is pretty broken, and it’s why it’s so
unusual, but also so refreshing to hear two people who disagree actually be able to converse
rationally and civilly and to shed light, not just blow smoke. And we know from so many of you
that that’s exactly why you tune into us. And it’s why I’d like to remind you that as you turn to us
for all of that, we turn to you for support. We are a nonprofit and it’s contributions from listeners
like you who keep us going. So please consider sending us a buck or two, or 10 or 50. Go to our
website, iq2us.org, whatever works, whatever you can do, and it’ll give you a stake in what we’re
doing here each week. And it will mean that we’ll be here each week and next week and beyond.
So thanks. Please consider it.

And I’m John Donvan. We’ll see you next time. Thank you for tuning into this episode of
Intelligence Squared, made possible by a generous grant from the Laura and Gary Lauder
Venture Philanthropy Fund as a nonprofit. Our work to combat extreme polarization through
civil and respectful debate is generously funded by listeners like you, the Rosenkranz Foundation
and Friends of Intelligence Squared. Robert Rosenkranz is our chairman. Clea Conner is CEO,
David Ariosto is head of editorial, Julia Melfi, Shay O’Mara and Marlette Sandoval are our
producers. Damon Whitmore is our radio producer, and I’m your host, John Donvan. We’ll see
you next time.