00:00:00
Angie Basiouny: This podcast is
brought to you
00:00:03
by Knowledge at Wharton.
00:00:13
Welcome to Knowledge at Wharton.
I'm Angie Basiouny. I'm here
00:00:16
today with Kevin Werbach. He is
Professor and Chair of the
00:00:18
Department of Legal Studies and
Business Ethics at Wharton. He's
00:00:22
also the Faculty Director of our
new Wharton Accountable AI Lab,
00:00:26
which is dedicated to advancing
responsible development
00:00:29
of artificial intelligence. And that
is what we're going to talk
00:00:32
about today. Kevin, welcome
aboard.
00:00:34
Kevin Werbach: Thanks so much.
00:00:35
Appreciate having you here.
So let's just jump right into it.
00:00:38
What is accountable AI? Why
did you start this lab?
00:00:43
Accountable AI
00:00:44
is about understanding the
challenges that AI poses. The
00:00:49
starting point is that AI is an
incredible innovation. It has
00:00:53
tremendous potential to create
value for businesses and to do a
00:00:58
great deal of social good. But
we can't realize that potential,
00:01:02
we can't achieve the benefits of
AI without acknowledging,
00:01:06
understanding and mitigating the
risks. Thinking about the
00:01:10
potential dangers and harms and
problems with AI. So accountable
00:01:14
AI is about not just thinking
what could happen, what are the
00:01:18
risks, although that's part of
it, not just asking from an
00:01:22
abstract perspective, what
principles should organizations
00:01:25
have about what they're doing
with AI? Again, that's part of
00:01:27
it. Not just saying generally,
we should be responsible about
00:01:31
AI or have well-governed AI,
although that's part of it. It's
00:01:34
saying, systematically, how do
we put into place the kinds of
00:01:39
practices and the understandings
that it takes to ensure that AI
00:01:45
systems are deployed and
developed in the ways that
00:01:50
maximize their benefits and
appropriately mitigate and
00:01:54
address or redress the problems
and harms. And accountability is
00:01:59
chosen by— is something I chose
intentionally. It's about making
00:02:04
those connections, the
connection between the risks and
00:02:06
the potential or real harms and
what actually happens. To
00:02:11
prevent them, to mitigate them,
to understand them, to address
00:02:15
them. Having all those practices
in place and doing it in a
00:02:18
thoughtful, systematic,
structured, rigorous way, which,
00:02:22
of course, is very consistent
with how we
00:02:23
think about things
at Wharton.
00:02:25
So let me ask you, is the
Lab going to— are you going to
00:02:28
develop, sort of like a best
practices or prescriptive
00:02:31
information for business—
business leaders, for tech
00:02:34
companies, about how to use AI,
how to deploy it?
00:02:38
One of the things that I have
found in speaking with companies
00:02:42
in the research that I do in
this area, and as we were
00:02:46
putting together the plans for
the Lab, is that most of them are
00:02:52
really struggling to get on top
of these issues. They don't
00:02:55
understand what other
organizations are doing. There's
00:02:57
a few companies who are very far
advanced, especially some of the
00:03:01
big technology companies have
invested significantly in
00:03:05
responsible AI or AI governance.
But even they have questions
00:03:09
about, what should they be
doing? What are other companies
00:03:12
doing? Are they appropriately
addressing all of the issues?
00:03:16
What does the data show about
what kinds of governance
00:03:19
mechanisms are effective? And
most companies are not even at
00:03:23
that point. So we are certainly
not going to say we'll tell
00:03:27
companies what the best
practices are. AI is so diverse,
00:03:31
and there's so many different
kinds of AI. There's— there's
00:03:33
machine learning systems,
there's generative AI. It's a
00:03:36
different thing if we're talking
about a company that is doing
00:03:39
hardcore technical development
of AI models, versus a company
00:03:43
that may be a very large
enterprise but is deploying a
00:03:46
system that they are procuring
from elsewhere, versus a small
00:03:49
startup that is involved in this
area. And it depends on what
00:03:53
industry you're in and so forth.
So we are first going to try to
00:03:57
understand what organizations
are actually doing, what's
00:04:01
successful, what's not
successful, what are the gaps, to
00:04:03
try and synthesize some of that
to help organizations understand
00:04:07
what the possibilities are. And
it's a moving target. It's going
00:04:10
to be an ongoing process of
understanding what can be done,
00:04:14
what are all the problems that
are most concerning, and how can
00:04:17
they be overcome?
00:04:18
That is a tall task, but I know
that you're up for it. I want to
00:04:21
tell people a little bit about
your about your background. You
00:04:24
have a law degree from Harvard.
You came to Wharton in 2004, so
00:04:29
going on 21 years now. But you
also worked in the Clinton
00:04:33
Administration, the Obama
Administration. You worked with
00:04:35
the FCC on emerging technology.
You've been at this for a long
00:04:39
time. You have four books about
technology, including
00:04:41
blockchain. You've seen the
emerging technology. You've
00:04:46
worked on the business
implications, the ethical
00:04:48
implications. Here's AI. Is it
different? How is it different
00:04:52
from the concerns that we've
dealt with in the past? Or is it
00:04:55
the same?
00:04:56
Some of both. As you note, I
have been working on emerging
00:04:59
technologies my whole career. When
I started in the 1990s, that was
00:05:02
the internet. And I wrote a
paper on internet policy at the
00:05:06
Federal Communications
Commission. This was early on,
00:05:08
before I was an academic. And at
that point, there were something
00:05:13
like less than 50 million people
on the internet in the entire
00:05:16
world, and the vast majority of
them were people dialing up on
00:05:20
their telephone to the
proprietary America Online
00:05:23
service. There was not a single
person in all of China who had a
00:05:26
private internet connection at
that point. And yet, we could
00:05:28
see the issues that were coming
up. We could see that this has a
00:05:31
technology that has the
potential to change the world,
00:05:34
and we needed to understand what
the issues were. And so all
00:05:37
throughout my career, I've tried
to get engaged on major,
00:05:41
important technology
developments early enough to
00:05:45
identify the issues to work on,
helping to develop the
00:05:49
regulatory strategies, work with
government, identify and
00:05:53
highlight what the problems are
before it was too late. And so I
00:05:57
did that with broadband
technology. I did that with
00:05:59
something called gamification,
which is applying psychological
00:06:03
techniques and other techniques
from video games to motivate
00:06:06
people in different contexts. I
did it, as you mentioned, with
00:06:09
blockchain, which was another
field that because— that I saw coming,
00:06:14
that had this diverse potential,
but it was still poorly
00:06:17
understood. In fact, it's still
poorly understood today. AI, I
00:06:20
put in a similar bucket. We are,
in some ways, very far along
00:06:24
with AI. The AI, if you're
talking about in terms of
00:06:27
machine learning technology, is
decades old. In some ways,
00:06:30
though, we're just at the
beginning. We're just a couple
00:06:32
years after the kind of ChatGPT
shot heard around the world
00:06:36
announcement that kicked off
this incredible race to exploit
00:06:42
the potential and understand the
potential of generative AI. And
00:06:46
we know there are all these
problems. We know there are
00:06:48
issues about privacy and bias
and intellectual property and
00:06:54
manipulation and so on and so
forth, and yet we don't have
00:06:57
good solutions. So AI is similar
to these earlier technologies in
00:07:00
that it starts at a point where
it has tremendous potential and
00:07:05
generates a lot of excitement,
but there's a lack of
00:07:07
understanding broadly about,
really, whether we'll realize its
00:07:11
potential and what the impacts
will be. But every technology is
00:07:14
different. And with each of
these waves, we build on what
00:07:18
came before. So AI leverages the
fact that we have the internet,
00:07:21
and we have these incredible
networks and technical
00:07:24
capabilities which allow things
to be deployed and scaled very
00:07:28
fast around the world. And we
see this tremendous amount of
00:07:32
activity and investment going
into this space. So it's
00:07:34
different than it was back 30
years ago, when I was looking at
00:07:38
dial up internet, but it's
similar in that we have this
00:07:41
period of uncertainty, and I
think that is the point where
00:07:44
it's most important to really
dig in. Think about the ethical
00:07:48
issues, think about the
governance issues, think about
00:07:50
the regulatory issues.
And so that's
00:07:52
really the genesis of the
Accountable AI Lab.
00:07:56
There are a number of issues. In
my experience interviewing
00:07:58
people about AI, which I've been
doing quite a bit over the last
00:08:01
year, I find that there are
three camps, right? So there are
00:08:04
the people who fear it, the
people who celebrate it, can't
00:08:08
wait for more of it. And then
those are— those are— there are
00:08:11
folks who are just proceeding
with caution, right? Yellow
00:08:13
light, green light, red light.
What camp do you fall into? Why?
00:08:18
And, you know, what's your
overall message about AI,
00:08:20
especially heading up this Lab?
00:08:23
All three.
- All three.
00:08:25
You can't fear it without
celebrating. Because if you fear
00:08:29
it, it means that you believe AI
has this incredible potential,
00:08:32
that it's going to be deployed
and going to have real impacts.
00:08:35
And similarly, you can't
celebrate it without recognizing
00:08:39
these challenges, a whole range
of challenges. And some of them
00:08:42
are very speculative, but many
of them are very real. I've talked
00:08:45
to lots of companies that say
our focus is not on regulation.
00:08:50
Our focus is on whatever the
government tells us. We know
00:08:54
we're going to deploy these
systems that might have
00:08:56
problems. And if we build and
deploy something that breaks, it
00:09:00
fails, the generative AI system
hallucinates and gives false
00:09:02
information, that could be a big
problem for us with our
00:09:05
customers in the marketplace. So—
and these are companies that are
00:09:09
deploying. These are companies
that are excited about it, but
00:09:11
they realize they need to
understand the problems. And
00:09:14
then the reality is, there are
some aspects of this where speed
00:09:18
is absolutely essential.
Companies need to invest things
00:09:22
are developing so fast, there's
so much potential. You don't
00:09:24
want to get left behind. But you
need to understand where there
00:09:29
are points where care is
warranted, where there is the
00:09:33
opportunity and the need to
slow down and ask and answer
00:09:37
these questions. And even if the
technology is moving really
00:09:40
fast, there's going to be
regulation. There are going to
00:09:43
be laws passed. There are going to
be court cases addressing these
00:09:45
issues. And so you can't just
ignore all of that. You have to
00:09:49
appreciate that development of
the legal process and the
00:09:52
development of, frankly, the
kinds of deeper understandings
00:09:56
that come out of research in
lots of different fields, not
00:09:59
just in law. In terms of, what are
the technical capabilities? What
00:10:02
can we do to mitigate bias? What
is the potential for explanation
00:10:06
of generative AI systems? It's a
fascinating area of advanced
00:10:10
research. And what is the
development of ethical and
00:10:14
psychological behavioral
understandings of what's going
00:10:16
on here? That is happening over
time. Not at the same speed as
00:10:21
the technical development of AI,
but it's going to have a really
00:10:24
big impact, all those things, on
being able to realize the full
00:10:28
potential of the technology.
00:10:30
I know that you and so many of
your colleagues at Wharton, and
00:10:33
beyond Wharton— everybody's
working on these issues. So I'm
00:10:35
really excited to see what's
going to come out of the Lab.
00:10:38
Before we go, I do want to let
folks know about your podcast.
00:10:41
It's called <i>The Road to
Accountable AI</i>. I've listened
00:10:44
to a couple of episodes. You got
some really interesting guests
00:10:46
on there. Can you just tell us a
little bit about it?
00:10:49
Yeah, the podcast is an
interview show. I spend 30 to 40
00:10:52
minutes on each episode talking
with a guest, and it's a range.
00:10:56
I speak with senior government
officials from multiple
00:11:00
countries. I speak with
technologists, I speak with
00:11:04
academics, I speak with business
executives who are leading the
00:11:08
responsible AI groups or AI
governance groups at some of the
00:11:11
largest companies. And I speak
with startups that are building
00:11:15
tools to address some of these
problems. And so it's really
00:11:17
intended as an educational
journey on how this broad area
00:11:24
of accountable AI is developing,
and trying to help people
00:11:28
understand what the state of the
art is and also what the
00:11:32
questions are that they should
be thinking about.
00:11:34
It definitely goes deep. I
appreciate it. Thanks for being
00:11:37
here.
- Absolutely,
00:11:38
really a pleasure to do it, and
thanks so much for the interest.
00:11:41
Kevin Werbach, everyone,
Professor and Chair of the
00:11:44
Department of Legal Studies and
Business Ethics here at Wharton.
00:11:47
He's also the Faculty Director
of our new Wharton Accountable
00:11:51
AI Lab. If you'd like to learn
more about that initiative, type
00:11:55
in Wharton Accountable AI Lab in
your browser. I also invite you
00:11:58
to check out his podcast, <i>Road
to Accountable AI</i>. For <i>Knowledge</i>
00:12:02
<i>at Wharton</i>, I'm Angie Basiouny.
Thanks for joining us.
00:12:06
For more insight from <i>Knowledge
at Wharton</i>, please visit
00:12:10
knowledge.Wharton.upenn.edu.
