0:01

This podcast is brought to

0:03

you by Knowledge at Wharton.

0:06

Welcome to Knowledge at Wharton.

0:08

I'm Andrew Bassini. I'm here

0:10

today with Kevin Warbach. He

0:12

is professor and chair of

0:15

the Department of Legal

0:17

Studies and Business Ethics

0:19

at Wharton. He's also

0:21

the faculty director of

0:23

our new Wharton Accountable

0:25

AI Lab. which is dedicated

0:27

to advancing responsible development of

0:30

artificial intelligence. And that is

0:32

what we're going to talk about today.

0:34

Kevin, welcome aboard. Thanks so much. Appreciate

0:36

you having your hair. So let's just jump

0:39

right into it. What is accountable AI? Why

0:41

did you start this lab? Accountable

0:43

AI is about understanding the

0:45

challenges that AI poses. The

0:48

starting point is that AI

0:50

is an incredible innovation, has

0:52

tremendous potential to create value

0:55

for businesses, and to do

0:57

a great deal of social

0:59

good. But we can't realize

1:01

that potential. We can't achieve

1:04

the benefits of AI without

1:06

acknowledging and mitigating the risks.

1:08

thinking about the potential dangers

1:10

and harms and problems with

1:12

AI. So accountable AI is

1:14

about not just thinking what

1:16

could happen, what are the risks,

1:18

although that's part of it, not

1:20

just asking from an abstract perspective,

1:22

what principles should organizations have about

1:24

what they're doing with AI? Again,

1:27

that's part of it. Not just

1:29

saying generally we should be responsible

1:31

about AI or have well-governed AI,

1:33

although that's part of it. It's

1:35

saying systematically. How do we put

1:37

into place the kinds of

1:40

practices and the understandings that

1:42

it takes to ensure that

1:44

AI systems are deployed and

1:46

developed in the ways that

1:48

maximize their benefits

1:51

and appropriately mitigate and

1:53

address or redress the problems and

1:55

harms? And accountability is chosen. by

1:57

something I chose intentionally, it's about

1:59

making those connections, the connection between

2:01

the risks and the potential or

2:03

real harms, and what actually happens

2:05

to prevent them, to mitigate them,

2:07

to understand them, to address them,

2:09

having all those practices in place

2:11

and doing it in a thoughtful,

2:13

systematic, structured, rigorous way, which of

2:15

course is very consistent with how

2:17

we think about things at work.

2:19

So let me ask you, is

2:21

the lab going to, are you

2:23

going to develop sort of like

2:25

a best practices or prescriptive information

2:27

for business leaders for tech companies

2:30

about how to use AI, how

2:32

to deploy it? One of the

2:34

things that I have found in

2:36

speaking with companies in the research

2:38

that I do in this area,

2:40

and as we were putting together

2:42

the plans for the lab, is

2:44

that most of them are really

2:46

struggling to get on top of

2:48

these issues. They don't understand what

2:50

other organizations are doing. There's a

2:52

few companies who are very far

2:54

advanced, especially some of the big

2:56

technology companies have invested significantly in

2:58

responsible AI or AI governance. But

3:00

even they have questions about what

3:02

should they be doing? What are

3:04

other companies doing? Are they appropriately

3:06

addressing all of the issues? What

3:08

is a data show about what

3:10

kinds of governance mechanisms are effective?

3:12

And most companies are not even

3:14

at that point. So we are

3:16

certainly not going to say, we'll

3:18

tell companies what the best practices

3:20

are. AI is so diverse, and

3:23

there's so many different kinds of

3:25

AI. There's machine learning systems, there's

3:27

generative AI. It's a different thing

3:29

if we're talking about a company

3:31

that is doing hardcore technical development

3:33

of AI models. versus a company

3:35

that may be a very large

3:37

enterprise, but is deploying a system

3:39

that they are procuring from elsewhere,

3:41

versus a small startup that is

3:43

involved in this area, and it

3:45

depends on what industry you're in

3:47

and so forth. So we are

3:49

first going to try to understand

3:51

what organizations are actually. doing, what's

3:53

successful, what's not successful, what are

3:55

the gaps, to try and synthesize

3:57

some of that to help organizations

3:59

understand what the possibilities are. And

4:01

it's a moving target. It's going

4:03

to be an ongoing process of

4:05

understanding what can be done, what

4:07

are all the problems that are

4:09

most concerning, and how can they

4:11

be overcome. That is a tall

4:13

task, but I know that you're

4:15

upward. I want to tell people

4:18

a little bit about your background.

4:20

You have a law degree from

4:22

Harvard. You came to Wharton in

4:24

2004, so going on 21 years

4:26

now. But you also worked in

4:28

the Clinton administration, the Obama administration,

4:30

you worked with the FCC on

4:32

emerging technology. You've been at this

4:34

for a long time. You have

4:36

four books about technology, including blockchain.

4:38

You've seen the emerging technology. You've

4:40

worked on the business implications, the

4:42

ethical implications. Here's AI. Is it

4:44

different? How is it different from

4:46

the concerns that we've dealt with

4:48

in the past? Or is it

4:50

the same? Some of both. As

4:52

you note, I have been working

4:54

on emerging technologies my whole career

4:56

when I started in the 1990s.

4:58

That was the internet. And I

5:00

wrote a paper on internet policy

5:02

at the Federal Communications Commission. This

5:04

was early on before I was

5:06

an academic. And at that point,

5:08

there were something like less than

5:10

50 million people on the internet

5:13

in the entire world. And the

5:15

vast majority of them were people

5:17

dialing up on their telephone to

5:19

the proprietary American online service. There

5:21

was not a single person. in

5:23

all of China who had a

5:25

private internet connection at that point.

5:27

And yet we could see the

5:29

issues that were coming up. We

5:31

could see that this has a

5:33

technology that has the potential to

5:35

change the world and we needed

5:37

to understand what the issues were.

5:39

And so all throughout my career

5:41

I've tried to get engaged on

5:43

major important technology developments early enough

5:45

to identify the issues to work

5:47

on helping to develop. the regulatory

5:49

strategies, work with government, identify and

5:51

highlight what the problems are before

5:53

it was too late. And so

5:55

I did that with broadband technology.

5:57

I did that with something. called

5:59

gamification, which is applying psychological techniques

6:01

and other techniques from video games

6:03

to motivate people in different contexts.

6:05

I did it, as you mentioned,

6:08

with blockchain, which was another field

6:10

that I saw coming that had

6:12

this diverse potential, but it was

6:14

still poorly understood. If, frankly, it's

6:16

still poorly understood today. AI, I

6:18

put in a similar bucket. We

6:20

are in some ways very far

6:22

along with AI. The AI if

6:24

you're talking about in terms of

6:26

machine learning technology is decades old.

6:28

In some ways, though, we're just

6:30

at the beginning. We're just a

6:32

couple years after the kind of

6:34

chat GPT shot heard around the

6:36

world announcement that kicked off this

6:38

incredible race to exploit the potential

6:40

and understand the potential of generative

6:42

AI. And we know there are

6:44

all these problems. We know there

6:46

are issues about privacy and bias

6:48

and intellectual property and manipulation and

6:50

so on and so forth. And

6:52

yet we don't have good solutions.

6:54

So AI is similar to these

6:56

earlier technologies in that it starts

6:58

at a point where it has

7:01

tremendous potential and generates a lot

7:03

of excitement, but there is a

7:05

lack of understanding broadly about really

7:07

whether it will realize this potential

7:09

and what the impacts will be.

7:11

But every technology is different. And

7:13

with each of these waves we

7:15

build on what came before. So

7:17

AI leverages the fact that we

7:19

have the internet. And we have

7:21

these incredible networks and technical capabilities,

7:23

which allow things to be deployed

7:25

and scaled very fast around the

7:27

world. And we see this tremendous

7:29

amount of activity and investment going

7:31

into this space. So it's different

7:33

than it was back 30 years

7:35

ago when I was looking at

7:37

dial up internet. But it's similar

7:39

in that we have this period

7:41

of uncertainty. And I think that

7:43

is the point where it's most

7:45

important to really dig in. Think

7:47

about the ethical issues. Think about

7:49

the governance issues. regulatory issues and

7:51

so that's really the genesis of

7:53

the accountable AI lab. There are

7:56

a number of issues and in

7:58

my experience interviewing people about a

8:00

which I've been doing quite a

8:02

bit of in the last year,

8:04

I find that there are three

8:07

camps, right? So there are the

8:09

people who fear it, the people

8:11

who celebrate it, can't wait for

8:13

more of it, and then those

8:16

are, those are, there are folks

8:18

who are just proceeding with caution,

8:20

right? Yellow light, green light, red

8:22

light. What camp do you fall

8:24

into? Why? And you know, what's

8:27

your overall message about AI, especially

8:29

heading up this lab? It means

8:31

that you believe AI has this

8:33

incredible potential that it's going to

8:35

be deployed and going to have

8:38

real impacts. And similarly, you can't

8:40

celebrate it without recognizing these challenges,

8:42

a whole range of challenges. And

8:44

some of them are very speculative,

8:47

but many of them are very

8:49

real. I talked to lots of

8:51

companies that say, our focus is

8:53

not on regulation. Our focus is

8:55

on whatever the government tells us.

8:58

we know we're going to deploy

9:00

the systems that might have problems.

9:02

And if we build and deploy

9:04

something that breaks, it fails, the

9:07

generated AI system hallucinates and gives

9:09

false information, that could be a

9:11

big problem for us with our

9:13

customers in the marketplace. So, and

9:15

these are companies that are deploying,

9:18

these are companies that are excited

9:20

about it, but they realize they

9:22

need to understand the problems. And

9:24

then the reality is, there are

9:27

some aspects of this where speed

9:29

is absolutely essential. companies need to

9:31

invest, things are developing so fast,

9:33

there's so much potential, you don't

9:35

want to get left behind. But

9:38

you need to understand where there

9:40

are points, where care is warranted,

9:42

where there is the opportunity and

9:44

the need to slow down and

9:46

ask and answer these questions. And

9:49

even if the technology is moving

9:51

really fast, there's going to be

9:53

regulation. they're going to be lost

9:55

past. They're going to be court

9:58

cases addressing these issues. And so

10:00

you can't just ignore all of

10:02

that. You have to appreciate that

10:04

development of the legal process and

10:06

the development of, frankly, the kinds

10:09

of deeper understandings that come out

10:11

of research in lots of different

10:13

fields, not just in large. What

10:15

are the technical capabilities? What can

10:18

we do to mitigate bias? What

10:20

is the potential for explanation of

10:22

generative AI systems? It's a fascinating

10:24

area of advanced research. And what

10:26

is the development of ethical and

10:29

psychological behavioral understandings and what's going

10:31

on here? That is happening over

10:33

time, not at the same speed

10:35

as the technical development of AI,

10:38

but it's going to have a

10:40

really being impact, all those things

10:42

on being able to. realize the

10:44

full potential of the technology. Yeah,

10:46

the podcast is an interview show.

10:49

I spend 30 to 40 minutes

10:51

on each episode talking with a

10:53

guest and it's a range. I

10:55

speak with senior government officials from

10:58

multiple countries. I speak with technologists.

11:00

I speak with academics. I speak

11:02

with business executives who are leading

11:04

the responsible AI groups or AI

11:06

governance groups at some of the

11:09

largest companies. And I speak with

11:11

startups that are building tools to

11:13

address some of these problems. an

11:15

educational journey on how this broad

11:17

area of accountable AI is developing

11:20

and trying to help people understand

11:22

what the state of the art

11:24

is and also what the questions

11:26

are that they should be thinking

11:29

about. It definitely goes deep. I

11:31

appreciate it. Thanks for being here.

11:33

Absolutely. Really pleasure to do it

11:35

and thanks so much for the

11:37

interest. Kevin Warbach, everyone. Professor and

11:40

chair of the Department of Legal

11:42

Studies and Business Ethics here at

11:44

Wharton. He's also the faculty director

11:46

of our new Wharton Accountable AI

11:49

Lab. If you'd like to learn

11:51

more about that initiative, type in

11:53

Wharton Accountable AI Lab in your

11:55

browser. I also invite you to

11:57

check out his podcast. road

12:00

Accountable AI. AI.

12:02

For Knowledge at at I'm

12:04

Angie I'm Thanks for joining

12:06

us. Thanks for For more insight

12:08

from Knowledge at Wharton,

12:10

please visit at Wharton, please visit .edu.