Popular in Law

Human Rights Law

Human Rights and Corruption: Problems and Potential of Individualising a Systemic Problem

Max Planck Institute Anne Peters

Law

A statistical approach to law school citation rankings

University of Virginia Joshua Fischman

Constitutional Law

What is constitutional ideology?

University of Oxford Nick Barber

Civil Procedure Law

Politics, Identity, and Class Certification on the U.S. Courts of Appeals

University of Pennsylvania Sean Farhang

Law

Diversity and Social Mobility in the UK legal profession

Law Society I. Stephanie Boyce

Human Rights Law

Transforming Restorative Justice for Transitional Settings

University of Nottingham Kerry Clamp

Law

Surrey’s Silence: Subpart F and the Swiss Subsidiary Tax that Never Was

Boston University Steven Dean

The Law of AI for Good

View more insights in Professional - Law - Internet Law

View more figures from University of San Diego

View the page of Orly Lobel

Next in Internet Law

Antitrust High-Tech Exceptionalism

Rebecca Haw Allensworth | Vanderbilt University

The False Choice Between Digital Regulation and Innovation

Anu Bradford | Columbia University

Legal policy and scholarship are increasingly focused on regulating technology to safeguard against risks and harms, neglecting the ways in which the law should direct the use of new technology, and in particular artificial intelligence (AI), for positive purposes. Orly Lobel argues that the focus on AI wrongs is descriptively inaccurate, undermining a balanced analysis of the benefits, potential, and risks involved in digital technology.

Image courtesy of interviewee. July 25, 2024

Log-in or Sign-up to Faculti
Currently viewing this subject insight as a guest. You have insight(s) remaining for this month. Login to view 8000+ figures on the platform.

Generating Workspace...

Export as PDF

This video discusses the themes and arguments presented in the author's new book, "The Quality Machine," which aims to shift the narrative surrounding artificial intelligence (AI) and its integration into society. The author critiques the prevalent alarmist discourse that focuses on the risks and harms posed by AI, arguing that this type of pessimism detracts from the potential benefits of harnessing technology for positive societal impact.

The author emphasizes the need for a rational, comparative analysis of AI, urging that discussions should not only highlight its flaws but also consider how AI can improve upon existing systems. They argue that AI's capacity for learning and improvement, especially when compared to human cognitive biases, presents an opportunity for progress rather than fear.

Additionally, the author calls for more balanced regulations and policy interventions that facilitate the use of AI for socioeconomic benefit, rather than simply limiting its deployment due to potential risks. They stress the importance of public discourse and the role of policymakers in guiding responsible AI development, ensuring that technology serves the public good and promotes equity and safety.

Key themes include the promotion of AI's positive potential, the risks of technology being viewed through an overly critical lens, the importance of empirical evidence in policy-making, and the call for collaborative efforts to engage all stakeholders in shaping the future of AI. The overall goal is to encourage a future that leverages AI to enhance various aspects of human life while addressing its challenges effectively.

Language:

Export as PDF

X-TIMESTAMP-MAP:LOCAL=00:00:00.000,MPEGTS=900000

00:00.495 --> 00:02.775
I wrote The Quality Machine, my new book,

00:03.145 --> 00:06.695
harnessing Digital Technology for a better, brighter Future.

00:07.595 --> 00:11.455
As a response to this tunnel vision

00:12.145 --> 00:16.335
discussion, or what I call a very skewed conversation

00:16.515 --> 00:20.095
and debate that we've been having about,

00:20.305 --> 00:23.255
human machine future, about the future of ai

00:23.715 --> 00:26.855
and how it's affecting every aspect of our lives.

00:27.395 --> 00:31.815
Uh, what I argue is that we are indeed at the cusp

00:31.835 --> 00:33.375
of a, a huge revolution.

00:33.835 --> 00:38.655
Uh, every, human activity work,

00:39.475 --> 00:43.775
family, our, health, our

00:45.015 --> 00:48.855
intimate relations, they're all being affected by the

00:49.505 --> 00:51.855
rapid integration of algorithms

00:51.855 --> 00:55.695
and machine learning into our decision making processes,

00:56.205 --> 00:59.255
into, just our, our sorting

00:59.255 --> 01:01.775
and what we know about ourselves, about the world.

01:02.435 --> 01:05.735
Um, but what we've been seeing is a lot

01:05.735 --> 01:09.255
of alarmist discussions, a lot of alarmist reporting

01:09.635 --> 01:11.215
by, journalists.

01:11.795 --> 01:15.495
Um, and also that kind of trickles into the policies

01:15.695 --> 01:17.295
that we're seeing on both sides

01:17.295 --> 01:20.375
of the pond all over the world, in the EU

01:20.515 --> 01:24.935
and, here in the United States, where there is a,

01:25.135 --> 01:30.055
a skewed focus on AI as wrongs, as harms, as risks,

01:30.475 --> 01:32.495
as fails.

01:33.275 --> 01:37.935
And what I argue in both, my new article, the Law of AI

01:37.955 --> 01:41.815
for Good, and in my new book, the Equality Machine, is that

01:43.215 --> 01:46.575
absolutely undeniably there are, wrongs,

01:46.575 --> 01:50.215
there are risks, but the way that we need

01:50.215 --> 01:54.415
to analyze what's happening has to be much more rational,

01:54.525 --> 01:58.775
much more, comparative, not kind

01:58.775 --> 02:02.215
of in absolute terms of, you know, how AI doesn't do well.

02:02.275 --> 02:04.695
We need to compare it to the status quo and,

02:04.695 --> 02:08.095
and what, you know, what, needs to be fixed,

02:08.165 --> 02:10.575
what is imperfect right now.

02:11.395 --> 02:14.695
And, also celebrate and really discover and,

02:14.715 --> 02:17.575
and, develop a vision of

02:18.475 --> 02:23.055
how do we harness this technology for good, for, more

02:23.615 --> 02:24.775
equality for all the values

02:24.775 --> 02:26.535
that we want in a democratic society.

02:26.555 --> 02:28.855
And there is vast potential,

02:28.995 --> 02:32.295
but if we're kind of just focusing on the dark side,

02:32.805 --> 02:37.255
it's almost like a self-fulfilling prophecy that we get,

02:37.565 --> 02:40.535
very, fearful, alarmed,

02:40.675 --> 02:43.175
and disengaged with the potential.

02:43.625 --> 02:46.335
There is kind of this pendulum in the way

02:46.335 --> 02:50.335
that we think about technology in general, where,

02:50.785 --> 02:53.455
there is a tendency to,

02:53.995 --> 02:56.535
either be overly optimistic

02:56.595 --> 03:00.925
and have this kind of very rosy, you know, kind

03:00.925 --> 03:05.045
of blind trust, and maybe that was, true

03:05.185 --> 03:06.485
for a few decades ago.

03:07.065 --> 03:09.845
But then what I call the tech lash, I, I show that

03:10.425 --> 03:14.645
it starts from, specific disenchantment

03:14.665 --> 03:19.525
and specific, uncovering of various

03:20.775 --> 03:21.805
activities and,

03:21.985 --> 03:24.325
and wrongs that were promulgated

03:24.465 --> 03:28.485
by certain big tech corporations,

03:28.895 --> 03:30.325
which are, you know, absolutely

03:30.325 --> 03:31.605
something that we need to think about.

03:31.985 --> 03:35.605
Uh, you know, has Facebook done, you know, things

03:35.605 --> 03:37.005
that it should not have done.

03:37.265 --> 03:39.525
Um, you know, how do we think about

03:40.635 --> 03:42.725
what the algorithms prioritize,

03:42.745 --> 03:46.165
or how do we design, robots embodied

03:46.985 --> 03:50.165
AI in, in, a gendered way or,

03:50.265 --> 03:52.605
or not, you know, using stereotypes or not.

03:52.665 --> 03:57.405
And I, I, I go through so much of that in, in the book,

03:57.705 --> 04:01.045
but what I, show is that

04:02.225 --> 04:06.845
at some point we kind of delink that those specific,

04:07.725 --> 04:12.085
concerns about specific fails of, of,

04:12.085 --> 04:13.565
of certain corporations.

04:14.105 --> 04:18.325
And it becomes this mindset where we are,

04:18.585 --> 04:22.005
overly distrusting of anything that has to do

04:22.005 --> 04:24.765
with technology and specifically

04:24.875 --> 04:26.645
with artificial intelligence.

04:27.145 --> 04:31.765
And we kind of adopt a mindset of simply like,

04:32.505 --> 04:34.845
you know, stopping everything disengaging

04:35.585 --> 04:39.965
and, preferring, just, you know, the,

04:39.965 --> 04:44.365
the status quo over, what is inevitably becoming,

04:44.705 --> 04:46.285
you know, very important in our lives.

04:46.855 --> 04:51.165
There is, something that I've found very frustrating,

04:51.545 --> 04:54.845
particularly with, reporters, with the media,

04:55.665 --> 04:59.205
is this rehashing of the same exact stories

04:59.545 --> 05:03.325
of an AI fail that are no longer relevant

05:03.425 --> 05:05.725
or representative of what the technology actually is.

05:06.145 --> 05:07.885
So I'll give you two examples.

05:08.505 --> 05:09.685
Um, one,

05:09.685 --> 05:11.605
and I think these examples should be familiar

05:11.665 --> 05:14.445
to our audience because they are so rehashed

05:14.505 --> 05:15.685
and, you know, so kind

05:15.685 --> 05:19.725
of coming up all the time when you know every other,

05:20.095 --> 05:23.725
media story about, you know, how we should fear ai,

05:23.985 --> 05:25.125
brings this up.

05:25.825 --> 05:30.005
Um, so the first is how,

05:32.275 --> 05:35.915
probably a decade ago at this point, there was a finding

05:36.105 --> 05:40.115
that, facial recognition technology,

05:40.375 --> 05:45.155
was doing better when it comes to white male faces

05:45.855 --> 05:50.195
and doing worse when it had to screen,

05:50.895 --> 05:53.475
for, people of color or women.

05:54.335 --> 05:56.715
Um, that's absolutely,

05:57.325 --> 06:00.205
A big problem has to be corrected, has to be addressed.

06:00.625 --> 06:03.925
But in fact, it has been addressed in the newest

06:03.925 --> 06:07.245
technologies, and it's actually on the spectrum of problems.

06:07.275 --> 06:08.405
It's not a hard problem.

06:09.025 --> 06:12.645
It, it, it has to do with how much data

06:12.985 --> 06:16.405
and how representative the data, is

06:16.435 --> 06:18.845
that the algorithm is trained upon.

06:19.185 --> 06:22.965
And so you just simply have to, you know, feed it more,

06:22.965 --> 06:25.645
rather than less and, and make sure that it, you know,

06:25.645 --> 06:28.765
and audit it and, and make sure that it's not in,

06:29.185 --> 06:31.685
having this biased performance

06:31.745 --> 06:33.605
and across different demographics.

06:34.585 --> 06:37.445
But then what, what the problem is that, you know,

06:37.445 --> 06:40.885
that we're not updating, this is kind of the, the,

06:40.905 --> 06:42.965
the thinking about machines a static.

06:43.575 --> 06:47.885
We're just kind of, we being, different audiences,

06:47.885 --> 06:50.485
including policymakers, including regulators

06:50.485 --> 06:53.525
that have interacted with, but,

06:53.585 --> 06:57.205
but certainly, you know, the public, you know,

06:57.275 --> 07:02.255
just readers, businesses, they kind

07:02.255 --> 07:04.135
of hear these stories and they say,

07:04.275 --> 07:05.975
we can't use facial recreation.

07:06.215 --> 07:09.215
'cause it is, per se, you know, inherently biased,

07:09.475 --> 07:11.735
and it just doesn't do well, that, you know,

07:11.735 --> 07:13.255
that's, that's simply not true.

07:13.315 --> 07:16.455
So we have to keep, you know, insisting on,

07:17.095 --> 07:19.495
updating our critique or our concerns to

07:19.805 --> 07:22.495
what is actually the technology that's available right now.

07:22.495 --> 07:24.895
And I know it's hard because there, it's, you know,

07:25.395 --> 07:29.375
so rapidly developing, but, we need to remember,

07:29.375 --> 07:31.575
and this is kind of a message in the equality machine

07:31.595 --> 07:35.015
and the law of AI for good, that, machine learning,

07:35.315 --> 07:39.455
you know, as, as the term suggests, it has the potential

07:39.455 --> 07:40.775
to learn and to improve.

07:41.275 --> 07:44.255
And in fact, if you think about it comparatively, we know

07:44.255 --> 07:49.215
that it has more potential to improve than our cognition.

07:49.315 --> 07:52.335
So again, I kind of take that, adopt that comparative lens

07:52.385 --> 07:55.495
where, in my research, I do a lot

07:55.515 --> 07:56.855
of behavioral research.

07:57.055 --> 07:59.375
I collaborate with social psychologists, with

08:00.015 --> 08:02.775
business school professors, and we're very, you know,

08:03.365 --> 08:06.335
well aware of the, our cognitive biases.

08:06.875 --> 08:08.775
So we document irrationalities

08:08.995 --> 08:13.095
and the limits, of our human cognition.

08:13.715 --> 08:17.415
Um, I teach discrimination law and policy.

08:17.895 --> 08:20.815
I, I serve and as an expert witness in discrimination cases,

08:21.555 --> 08:24.055
and with discrimination, you know, and,

08:24.115 --> 08:28.095
and human bias, it's very hard for us as humans to de-bias,

08:28.125 --> 08:31.135
even when we're kind of shown that we

08:31.765 --> 08:33.015
tend to have certain biases.

08:33.295 --> 08:34.535
A lot of them are unconscious.

08:34.605 --> 08:38.655
It's very hard to learn, to improve, compare that

08:38.795 --> 08:41.335
to the potential of kind of really auditing

08:41.395 --> 08:46.215
and detecting problems in the, in an algorithm

08:46.435 --> 08:47.575
and then improving it.

08:47.915 --> 08:50.695
We really need to insist on, you know,

08:50.695 --> 08:53.735
that comparative advantage not only in its kind

08:53.735 --> 08:55.805
of static form of, you know, the potential

08:55.985 --> 08:59.565
or the, the, the current state of the technology,

08:59.905 --> 09:03.085
but also its potential to, grow and,

09:03.145 --> 09:05.005
and become better and more accurate.

09:05.385 --> 09:09.885
We have, this, expectation

09:10.195 --> 09:14.285
that we won't find any problems in AI

09:14.425 --> 09:17.525
or algorithmic decision making,

09:17.745 --> 09:22.005
or automation that we don't have from human behavior.

09:22.585 --> 09:24.045
And so, another example, I mentioned

09:24.045 --> 09:26.405
that there's several examples, of like, stories

09:26.405 --> 09:28.005
that are rehashed,

09:28.425 --> 09:31.525
and are the, don't kind of match up to where we are

09:32.135 --> 09:35.805
right now in 2024, or where we will be in 2026.

09:36.385 --> 09:40.285
Um, another example is,

09:40.485 --> 09:41.605
self-driving cars.

09:42.345 --> 09:46.565
Um, we a have a tendency to do

09:46.565 --> 09:51.505
what you were just mentioning of kind of require perfection

09:51.725 --> 09:53.825
for, from an automated system

09:54.255 --> 09:56.265
that we don't require from humans.

09:56.845 --> 09:59.025
So in the example of driving, I think is very,

09:59.245 --> 10:00.625
easy to explain.

10:00.765 --> 10:04.225
And once we kind of, you know, insist on this, people,

10:04.565 --> 10:08.145
can grasp this, you know, much, much more clearly.

10:08.885 --> 10:12.865
Uh, it's not enough to just be aware, made aware that

10:13.605 --> 10:15.145
in one instance,

10:15.215 --> 10:18.985
that a autonomous vehicle was deployed on the roads,

10:19.485 --> 10:21.185
it got into an accident.

10:21.695 --> 10:24.465
What we need to ask is not whether it's perfect

10:24.725 --> 10:26.465
and, you know, a hundred percent safe,

10:26.885 --> 10:29.785
but whether it's safer than the human driver.

10:30.605 --> 10:33.465
And we have the data, this is an empirical question,

10:33.645 --> 10:35.745
and the data will change from year to year,

10:35.745 --> 10:37.585
depending on the state of the technology.

10:37.805 --> 10:40.385
But we need to look at that data and,

10:40.605 --> 10:42.545
and insist on the comparative advantage,

10:42.885 --> 10:44.225
not insist on perfection.

10:44.765 --> 10:49.505
So this is, specifically a, an FTC report that came out,

10:49.765 --> 10:53.345
a couple of years ago where, Congress,

10:53.365 --> 10:56.705
the United States Congress asked the Federal Trade

10:56.705 --> 10:59.385
Commission to investigate

10:59.385 --> 11:02.745
and report on how can we use AI

11:02.745 --> 11:04.425
to tackle online harms.

11:05.285 --> 11:09.065
And the report for me is really paradigmatic of the kind of,

11:09.645 --> 11:10.785
irrational

11:11.245 --> 11:15.145
and kind of skewed aversion of AI

11:15.335 --> 11:19.505
that I see in a lot of other, policy fields and,

11:19.605 --> 11:23.545
and debates, where the FTC

11:24.175 --> 11:26.065
doesn't adopt a comparative lens.

11:26.445 --> 11:31.305
Um, it doesn't think, you know, kind of take the data or,

11:31.305 --> 11:33.625
or look empirically at,

11:34.565 --> 11:39.265
how our human moderators, outperforming

11:39.325 --> 11:43.305
or underperforming automated content moderators.

11:43.605 --> 11:45.785
And, you know, what, what is, what is available?

11:45.935 --> 11:48.465
What are the tools that are out there in the industry?

11:48.535 --> 11:51.625
What are companies doing? What are the best practices?

11:51.695 --> 11:54.005
What are bad practices? Doesn't do any of that.

11:54.635 --> 11:58.765
What it does is, go to kind of the same usual,

11:59.485 --> 12:03.045
a lot of them outdated suspects of,

12:03.475 --> 12:06.605
here are instances where AI didn't get it right,

12:06.705 --> 12:11.365
or AI was limited in its, ability,

12:11.465 --> 12:14.205
or presented some, you know, wrongs or risks.

12:14.745 --> 12:15.965
Um, it cites those,

12:16.305 --> 12:18.925
it says the technology is too rudimentary.

12:19.535 --> 12:23.805
There are all these, hypothetical, you know,

12:23.805 --> 12:25.325
sometimes panned out risks.

12:25.505 --> 12:29.245
Um, so Congress do nothing and that do nothing.

12:30.105 --> 12:32.805
Status quo approach is something that concerns me

12:32.805 --> 12:37.605
because I, I never want to, kind of

12:38.275 --> 12:42.405
call for, you know, just adoption for the sake of adoption.

12:42.705 --> 12:46.765
I'm not, I don't, have I particular,

12:47.545 --> 12:48.685
you know, stake

12:48.785 --> 12:53.005
or stance on, whether we should shift in any per,

12:53.225 --> 12:55.365
you know, one specific circumstance

12:55.465 --> 12:58.925
or context from, human decision making

12:58.925 --> 13:00.165
to automated decision making.

13:00.545 --> 13:05.165
But what I insist on is that we run through what we know

13:05.355 --> 13:06.965
what the empirics are

13:07.425 --> 13:12.285
and analyze this rationally rather than, just kind of

13:12.905 --> 13:16.845
having, you know, what we might call the precautionary,

13:17.785 --> 13:22.045
default of if we don't know enough, we do nothing.

13:22.505 --> 13:24.685
Uh, I think it's just not helpful,

13:24.705 --> 13:26.245
and it's also just not, not correct

13:26.265 --> 13:27.285
in this particular context.

13:27.685 --> 13:28.805
'cause we do know enough.

13:29.025 --> 13:33.365
The reality is that every Fortune 500 company

13:34.025 --> 13:37.805
is automating, certainly with online harms.

13:38.225 --> 13:42.445
Um, the, you know, every major platform,

13:42.625 --> 13:46.725
social media platform, e-commerce

13:47.405 --> 13:50.285
platform is using automated systems

13:51.105 --> 13:55.645
to do content moderation to, detect

13:56.435 --> 13:59.165
illegal, illegal activities and all that.

13:59.245 --> 14:02.285
I actually happen to be very familiar with some of these

14:02.525 --> 14:05.685
because I, consult a lot of these, you know, major,

14:06.105 --> 14:07.645
companies.

14:08.625 --> 14:11.765
And what I argue about the policy is that

14:12.975 --> 14:16.765
perversely and counter productively when, the FTC

14:16.785 --> 14:20.005
as a major regulator, you know, that's impacting the world,

14:20.155 --> 14:21.845
says, we don't know.

14:22.025 --> 14:23.725
So, you know, we won't say anything.

14:24.385 --> 14:28.525
Um, it's too risky when, when they do that perversely,

14:28.665 --> 14:33.485
we get a lot of the snake oil, the bad instances of ai,

14:33.745 --> 14:38.445
out there competing equally with the better, systems,

14:38.945 --> 14:40.885
the, the better services

14:41.185 --> 14:44.325
and products that, are out there in the for-profit,

14:44.425 --> 14:46.285
you know, in the, in the private industry.

14:46.835 --> 14:50.245
Whereas I, I argue that the role of

14:51.125 --> 14:53.805
Policymakers of regulators is to sort out

14:53.805 --> 14:56.325
and give us, you know, tools to differentiate

14:56.905 --> 15:00.605
and to say like, this is the gold standard right now for

15:01.285 --> 15:04.045
adopting a tool that's automated content moderation.

15:04.145 --> 15:07.925
And, you know, this is how we should, perform audits

15:07.945 --> 15:10.045
and, and regulate the market.

15:10.825 --> 15:15.045
Uh, and, and so, so I, I never wanna be heard

15:15.105 --> 15:19.885
or read as, somebody who's averse to regulation.

15:19.935 --> 15:22.605
Quite the contrary, the law of AI for good

15:23.265 --> 15:25.045
is actually an argument,

15:25.305 --> 15:27.805
for more government government intervention,

15:28.105 --> 15:31.205
but government intervention that's not, kind

15:31.205 --> 15:33.525
of all about banning certain technologies,

15:33.705 --> 15:35.205
but actually sorting out

15:35.545 --> 15:38.965
and, propelling the more effective technologies,

15:39.505 --> 15:43.645
the role of, public entities,

15:43.645 --> 15:47.925
which has always been the role of public entities, to,

15:48.425 --> 15:51.165
aid, consumers

15:51.505 --> 15:56.165
and, you know, private corporations in,

15:56.785 --> 16:00.325
choosing, selecting

16:00.745 --> 16:04.685
and, deploying, you know, the best practices,

16:04.915 --> 16:06.605
that, that are available.

16:06.945 --> 16:10.845
Uh, and this is especially important for,

16:11.465 --> 16:13.605
the smaller companies

16:13.785 --> 16:17.485
and, the, you know, lay consumers,

16:17.485 --> 16:19.245
because I think that the, the larger

16:19.845 --> 16:23.925
companies like Google will already, you know, know how to

16:24.435 --> 16:28.605
sort the different, you know, products available or,

16:28.625 --> 16:30.245
or they will de develop things

16:30.245 --> 16:31.725
that internally and audit them.

16:32.465 --> 16:37.245
But, if we want to have competition, which contributes

16:37.265 --> 16:39.765
to consumer welfare, which consumer contributes

16:39.765 --> 16:43.565
to innovation, then it is the role

16:43.785 --> 16:46.885
of regulators to say,

16:47.275 --> 16:51.885
just like they're doing an environmental, regulation

16:52.065 --> 16:54.245
or occupational safety regulation

16:54.505 --> 16:58.125
or transportation regulation, here's the gold standard.

16:58.155 --> 17:02.445
Like, you know, you need a a, a safety,

17:03.305 --> 17:07.125
system for, for this kind of, vehicle.

17:07.865 --> 17:10.805
Uh, you need to have a seatbelt in every car.

17:11.345 --> 17:12.725
Um, tell us

17:13.455 --> 17:15.965
after, you,

17:16.025 --> 17:20.085
the regulator has undertaken some empirical study or,

17:20.225 --> 17:21.765
or looked at the empirical studies

17:21.865 --> 17:24.045
or commission, you know,

17:24.045 --> 17:28.005
and outsourced some of that work, tell us what works,

17:28.465 --> 17:31.165
invest, procure,

17:31.715 --> 17:35.845
incentivize the private sector to develop,

17:36.105 --> 17:39.045
to improve what works,

17:39.425 --> 17:42.885
and what best aligns with our public interests.

17:43.065 --> 17:46.645
So increasing health, increasing access, increasing safety,

17:46.645 --> 17:50.805
increasing equity and equality across sectors.

17:51.265 --> 17:54.205
Um, really kind of having a public looking,

17:54.385 --> 17:58.365
not just a for-profit looking view of how to deploy

17:58.365 --> 18:01.365
that technology and, and develop these standards.

18:01.425 --> 18:03.765
And that, that has always been the role of regulation,

18:03.865 --> 18:07.805
and we're not doing enough of that now that, you know,

18:08.065 --> 18:11.205
AI is, is accelerating and, and shaping our lives.

18:11.705 --> 18:14.125
And, you know, some of this, i I wanna say kind

18:14.125 --> 18:16.445
of cautiously that I understand that it's,

18:16.505 --> 18:19.285
that's a tough thing when, you know things are moving

18:19.385 --> 18:21.445
so fast and they require knowledge

18:21.445 --> 18:25.405
and skill that are not always available to, regulators

18:25.405 --> 18:28.045
that have not been trained in machine learning

18:28.305 --> 18:31.525
and, you know, don't maybe see themselves as expert.

18:32.305 --> 18:35.045
Um, and there is, I think that lag

18:35.715 --> 18:40.325
that always exists in some, shape and

18:40.325 --> 18:43.205
or form, with, you know, the public

18:44.435 --> 18:47.245
regulator always kind of catching up to the private market,

18:47.585 --> 18:49.725
but I think now it's even more pronounced.

18:50.385 --> 18:53.525
We need to accelerate the investment in our,

18:53.815 --> 18:58.525
skill creation in hiring more, professionals in,

18:58.945 --> 19:03.165
the public sector that understand the technology so

19:03.165 --> 19:06.685
that we can do more of that work of harnessing AI for good.

19:07.465 --> 19:10.805
Um, one huge risk is,

19:11.195 --> 19:13.085
increased market concentration.

19:13.465 --> 19:18.325
And so we're seeing that already with, market dominance.

19:18.585 --> 19:20.205
Uh, what I mentioned

19:20.205 --> 19:21.565
before is that the,

19:21.785 --> 19:24.045
the bigger companies will do thi these kind

19:24.045 --> 19:25.805
of things anyway.

19:26.105 --> 19:30.565
Um, so we have, you know, open AI and we have, Google

19:30.865 --> 19:33.805
and Microsoft, all kind of racing ahead.

19:34.505 --> 19:39.005
Um, but the smaller kind of new entry, fresh blood,

19:39.505 --> 19:42.685
more competition over, you know, pricing

19:43.225 --> 19:46.285
and consumer welfare, we'll see less of that

19:46.355 --> 19:50.765
because we don't have the, the public arm to offset

19:51.115 --> 19:55.485
that disparity in abilities to, to sort to, to develop.

19:56.225 --> 19:57.885
Um, so that's one risk.

19:58.145 --> 20:02.405
Um, and so it's kind of another, way

20:02.405 --> 20:04.765
of thinking about competition law

20:04.765 --> 20:05.965
and antitrust policy

20:06.185 --> 20:10.685
to think about investment in technology as a public good.

20:11.135 --> 20:15.005
Technology is not just a private tool for, you know, the

20:15.965 --> 20:18.325
strongest companies to profit from.

20:18.665 --> 20:23.365
It really is a public good that we need to leverage for,

20:23.625 --> 20:24.965
the welfare of everybody.

20:24.985 --> 20:28.245
And that, again, that's the role of, the regulator.

20:29.265 --> 20:33.805
Um, the other thing is that we will see you just more fails,

20:34.025 --> 20:38.485
because, you know, if we don't,

20:38.985 --> 20:43.125
insist on, the fact

20:43.125 --> 20:46.605
that technology can actually correct for stereotypes for,

20:47.065 --> 20:49.925
for example, in, in the book,

20:49.925 --> 20:54.175
in the equality machine, i, I go through, a lot

20:54.175 --> 20:58.455
of the kind of stereotypes that have been injected into,

20:58.995 --> 21:03.295
personal digital assistance, to,

21:04.195 --> 21:07.295
robots that are designed with like our very like,

21:07.415 --> 21:09.375
gendered way of thinking.

21:09.525 --> 21:13.695
Like, so, so the private sector that's not, you know,

21:13.795 --> 21:18.215
too interested in public looking values will likely have

21:18.285 --> 21:22.255
like a, robot

21:22.255 --> 21:25.935
that is a doctor designed as a male and,

21:26.115 --> 21:27.775
and a nurse doctor,

21:28.035 --> 21:32.255
or sorry, nurse robot, you know, as, as a, as a female,

21:32.555 --> 21:35.335
you know, you, you're see, starting to see all those kind

21:35.335 --> 21:37.935
of like, stereotypes that are

21:38.675 --> 21:42.255
age old just being replicated into our,

21:43.175 --> 21:44.535
embodied ai.

21:45.235 --> 21:50.055
Um, but a public, looking, you know,

21:50.605 --> 21:51.815
role is, or,

21:52.075 --> 21:56.055
or, action by, by agencies can be

21:56.395 --> 22:00.895
to develop these kind of unexpected surprising,

22:01.635 --> 22:03.775
if stereotype flipping

22:03.955 --> 22:07.755
or resisting, access or,

22:07.775 --> 22:09.675
or availability of, of these tools.

22:10.415 --> 22:13.755
Um, other examples, and,

22:13.775 --> 22:16.395
and, you know, you might wanna ask me about this

22:16.395 --> 22:20.675
or you, you may have wanted to ask me about this, in,

22:20.735 --> 22:23.275
in next questions, but in the law of AI for good,

22:24.035 --> 22:28.595
I talk about how, another fallacy is,

22:28.935 --> 22:32.115
our over privileging of the value of privacy

22:32.855 --> 22:35.115
versus other, values.

22:36.095 --> 22:40.325
And, all along, you know, in, in all of my research,

22:40.485 --> 22:42.845
I I very much care about privacy.

22:43.365 --> 22:45.885
I think it's, it, it is an important democratic value.

22:46.045 --> 22:50.645
I think we should be concerned about, you know, over,

22:51.005 --> 22:53.365
surveillance and misuse of information.

22:54.025 --> 22:58.165
Um, but I also show that

22:59.075 --> 23:03.045
it's one of many other democratic values that we have

23:03.345 --> 23:08.245
and goals and norms that we, have always,

23:08.515 --> 23:09.925
strived for in a democracy.

23:10.905 --> 23:14.575
And that, if we

23:15.285 --> 23:19.415
just have kind of that lead of, okay, now that we have ai,

23:19.415 --> 23:22.135
there's even a greater risk that our privacy will be,

23:22.675 --> 23:24.135
infringe upon.

23:24.675 --> 23:29.335
So do not collect, do not misuse, do not, there's a kind

23:29.335 --> 23:34.215
of data minimization at, at the eu, GDPR

23:34.315 --> 23:36.415
and the eu, AI act.

23:36.415 --> 23:40.055
This kind of a very like controlling lens of,

23:40.615 --> 23:43.535
minimize the data that's collecting, if that's the kind

23:43.535 --> 23:48.365
of single dominant, rule, then there,

23:48.785 --> 23:50.365
and, and it's not contextualized

23:50.365 --> 23:53.965
and not offset by other, significant

23:54.585 --> 23:55.965
values and interests.

23:56.025 --> 24:00.525
Um, for example, having, representative data sets,

24:01.985 --> 24:06.935
completing missing, information and missing data.

24:07.355 --> 24:09.815
Uh, for example, in, in our health data.

24:10.075 --> 24:12.135
Uh, I have a whole chapter in the equality machine

24:12.135 --> 24:14.855
that starts from this understanding

24:14.955 --> 24:19.175
that's very well documented in the empirical research on

24:19.235 --> 24:24.095
how, clinical trials and, medical device

24:24.195 --> 24:27.615
and pharma, development has have,

24:27.925 --> 24:31.055
have long had this problem of basing a lot

24:31.055 --> 24:35.375
of the development on just kind of the skewed partial data

24:35.555 --> 24:38.095
of what, the white, you know,

24:38.765 --> 24:41.415
male physique biology, is.

24:41.955 --> 24:46.455
Um, now we have this opportunity to race ahead with,

24:47.155 --> 24:50.575
AI finding more patterns across different,

24:51.895 --> 24:56.455
databases that exist in, in hospitals and clinics and,

24:56.955 --> 25:01.295
and finding, you know, what works, what,

25:01.295 --> 25:04.375
what heals what, what is better patient care for all.

25:04.955 --> 25:08.935
Um, so, so again, what we need to kind of insist on is

25:09.325 --> 25:10.535
what are our public goals

25:10.675 --> 25:14.135
and how do we get there with the help of private, public,

25:14.515 --> 25:16.055
regulation.

25:16.785 --> 25:21.095
First and foremost, I want us to all be part

25:21.095 --> 25:22.975
of the conversation in the equality machine.

25:23.055 --> 25:24.975
I use the term skin in the game.

25:25.255 --> 25:27.775
I think everybody, all of us

25:27.775 --> 25:29.975
around the world should have skin in the game,

25:30.235 --> 25:34.775
in envisioning, in designing, in deploying, in,

25:35.025 --> 25:38.335
benefiting from our newfound capabilities

25:38.635 --> 25:40.215
of artificial intelligence.

25:40.715 --> 25:42.655
And so we can only do that.

25:43.435 --> 25:45.975
We actually have rational conversations,

25:46.415 --> 25:50.295
informed conversations about what it is that we want

25:50.315 --> 25:53.255
and what it is that we have now and,

25:53.395 --> 25:55.415
and can have, in the future.

25:56.275 --> 26:00.015
Um, and, I think that, you know,

26:00.885 --> 26:05.215
when we have, these much more balanced conversations

26:05.715 --> 26:09.215
and examples of all the wonderful ways

26:09.285 --> 26:13.815
that AI can actually assist our lives, then that's when,

26:14.235 --> 26:18.295
things start moving along to, you know, the better

26:18.315 --> 26:20.135
and brighter future that I envision.

"Rethinking the Human-Machine Future: A Call for Rational Discourse on AI's Potential and Risks"**

00:00:00 - I wrote The Quality Machine, my new book, harnessing Digital Technology for a better, brighter Future. As a response to this tunnel vision discussion, or what I call a very skewed conversation and debate that we've been having about, human machine future, about the future of ai and how it's affecting every aspect of our lives. what I argue is that we are indeed at the cusp of a, a huge revolution. every, human activity work, family, our, health, our
00:00:49 - rapid integration of algorithms and machine learning into our decision making processes, into, just our, our sorting and what we know about ourselves, about the world. but what we've been seeing is a lot of alarmist discussions, a lot of alarmist reporting by, journalists. and also that kind of trickles into the policies that we're seeing on both sides of the pond all over the world, in the EU and, here in the United States, where there is a,
00:01:30 - as fails. And what I argue in both, my new article, the Law of AI for Good, and in my new book, the Equality Machine, is that absolutely undeniably there are, wrongs, there are risks, but the way that we need to analyze what's happening has to be much more rational, much more, comparative, not kind of in absolute terms of, you know, how AI doesn't do well. We need to compare it to the status quo and, and what, you know, what, needs to be fixed,
00:02:11 - And, also celebrate and really discover and, and, develop a vision of how do we harness this technology for good, for, more equality for all the values that we want in a democratic society. And there is vast potential, but if we're kind of just focusing on the dark side, it's almost like a self-fulfilling prophecy that we get, very, fearful, alarmed, and disengaged with the potential. There is kind of this pendulum in the way that we think about technology in general, where,
00:02:53 - either be overly optimistic and have this kind of very rosy, you know, kind of blind trust, and maybe that was, true for a few decades ago. But then what I call the tech lash, I, I show that it starts from, specific disenchantment and specific, uncovering of various activities and, and wrongs that were promulgated by certain big tech corporations, which are, you know, absolutely something that we need to think about.
00:03:35 - that it should not have done. you know, how do we think about what the algorithms prioritize, or how do we design, robots embodied AI in, in, a gendered way or, or not, you know, using stereotypes or not. And I, I, I go through so much of that in, in the book, but what I, show is that at some point we kind of delink that those specific, concerns about specific fails of, of,

"Overcoming Distrust: The Need to Reassess AI's Evolving Capabilities"

00:04:14 - And it becomes this mindset where we are, overly distrusting of anything that has to do with technology and specifically with artificial intelligence. And we kind of adopt a mindset of simply like, you know, stopping everything disengaging and, preferring, just, you know, the, the status quo over, what is inevitably becoming, you know, very important in our lives. There is, something that I've found very frustrating, particularly with, reporters, with the media,
00:04:59 - of an AI fail that are no longer relevant or representative of what the technology actually is. So I'll give you two examples. one, and I think these examples should be familiar to our audience because they are so rehashed and, you know, so kind of coming up all the time when you know every other, media story about, you know, how we should fear ai, brings this up. so the first is how, probably a decade ago at this point, there was a finding
00:05:40 - was doing better when it comes to white male faces and doing worse when it had to screen, for, people of color or women. that's absolutely, A big problem has to be corrected, has to be addressed. But in fact, it has been addressed in the newest technologies, and it's actually on the spectrum of problems. It's not a hard problem. It, it, it has to do with how much data and how representative the data, is
00:06:19 - And so you just simply have to, you know, feed it more, rather than less and, and make sure that it, you know, and audit it and, and make sure that it's not in, having this biased performance and across different demographics. But then what, what the problem is that, you know, that we're not updating, this is kind of the, the, the thinking about machines a static. We're just kind of, we being, different audiences, including policymakers, including regulators that have interacted with, but, but certainly, you know, the public, you know,
00:07:02 - of hear these stories and they say, we can't use facial recreation. 'cause it is, per se, you know, inherently biased, and it just doesn't do well, that, you know, that's, that's simply not true. So we have to keep, you know, insisting on, updating our critique or our concerns to what is actually the technology that's available right now. And I know it's hard because there, it's, you know, so rapidly developing, but, we need to remember, and this is kind of a message in the equality machine and the law of AI for good, that, machine learning, you know, as, as the term suggests, it has the potential
00:07:41 - And in fact, if you think about it comparatively, we know that it has more potential to improve than our cognition. So again, I kind of take that, adopt that comparative lens where, in my research, I do a lot of behavioral research. I collaborate with social psychologists, with business school professors, and we're very, you know, well aware of the, our cognitive biases. So we document irrationalities and the limits, of our human cognition. I teach discrimination law and policy. I, I serve and as an expert witness in discrimination cases,

"Comparative Analysis: Evaluating AI's Potential Against Human Bias and Decision-Making"

00:08:24 - and human bias, it's very hard for us as humans to de-bias, even when we're kind of shown that we tend to have certain biases. A lot of them are unconscious. It's very hard to learn, to improve, compare that to the potential of kind of really auditing and detecting problems in the, in an algorithm and then improving it. We really need to insist on, you know, that comparative advantage not only in its kind of static form of, you know, the potential or the, the, the current state of the technology, but also its potential to, grow and,
00:09:05 - We have, this, expectation that we won't find any problems in AI or algorithmic decision making, or automation that we don't have from human behavior. And so, another example, I mentioned that there's several examples, of like, stories that are rehashed, and are the, don't kind of match up to where we are right now in 2024, or where we will be in 2026. another example is, self-driving cars. we a have a tendency to do
00:09:51 - for, from an automated system that we don't require from humans. So in the example of driving, I think is very, easy to explain. And once we kind of, you know, insist on this, people, can grasp this, you know, much, much more clearly. it's not enough to just be aware, made aware that in one instance, that a autonomous vehicle was deployed on the roads, it got into an accident. What we need to ask is not whether it's perfect and, you know, a hundred percent safe,
00:10:30 - And we have the data, this is an empirical question, and the data will change from year to year, depending on the state of the technology. But we need to look at that data and, and insist on the comparative advantage, not insist on perfection. So this is, specifically a, an FTC report that came out, a couple of years ago where, Congress, the United States Congress asked the Federal Trade Commission to investigate and report on how can we use AI to tackle online harms. And the report for me is really paradigmatic of the kind of,
00:11:11 - and kind of skewed aversion of AI that I see in a lot of other, policy fields and, and debates, where the FTC doesn't adopt a comparative lens. it doesn't think, you know, kind of take the data or, or look empirically at, how our human moderators, outperforming or underperforming automated content moderators. And, you know, what, what is, what is available? What are the tools that are out there in the industry? What are companies doing? What are the best practices?
00:11:54 - What it does is, go to kind of the same usual, a lot of them outdated suspects of, here are instances where AI didn't get it right, or AI was limited in its, ability, or presented some, you know, wrongs or risks. it cites those, it says the technology is too rudimentary. There are all these, hypothetical, you know, sometimes panned out risks. so Congress do nothing and that do nothing. Status quo approach is something that concerns me

"Rational Regulation: Navigating the Shift from Human to Automated Decision-Making in AI"

00:12:38 - call for, you know, just adoption for the sake of adoption. I'm not, I don't, have I particular, you know, stake or stance on, whether we should shift in any per, you know, one specific circumstance or context from, human decision making to automated decision making. But what I insist on is that we run through what we know what the empirics are and analyze this rationally rather than, just kind of
00:13:17 - default of if we don't know enough, we do nothing. I think it's just not helpful, and it's also just not, not correct in this particular context. 'cause we do know enough. The reality is that every Fortune 500 company is automating, certainly with online harms. the, you know, every major platform, social media platform, e-commerce platform is using automated systems to do content moderation to, detect
00:13:59 - I actually happen to be very familiar with some of these because I, consult a lot of these, you know, major, companies. And what I argue about the policy is that perversely and counter productively when, the FTC as a major regulator, you know, that's impacting the world, says, we don't know. So, you know, we won't say anything. it's too risky when, when they do that perversely, we get a lot of the snake oil, the bad instances of ai, out there competing equally with the better, systems,
00:14:41 - and products that, are out there in the for-profit, you know, in the, in the private industry. Whereas I, I argue that the role of Policymakers of regulators is to sort out and give us, you know, tools to differentiate and to say like, this is the gold standard right now for adopting a tool that's automated content moderation. And, you know, this is how we should, perform audits and, and regulate the market. and, and so, so I, I never wanna be heard or read as, somebody who's averse to regulation.
00:15:23 - is actually an argument, for more government government intervention, but government intervention that's not, kind of all about banning certain technologies, but actually sorting out and, propelling the more effective technologies, the role of, public entities, which has always been the role of public entities, to, aid, consumers and, you know, private corporations in, choosing, selecting
00:16:04 - that, that are available. and this is especially important for, the smaller companies and, the, you know, lay consumers, because I think that the, the larger companies like Google will already, you know, know how to sort the different, you know, products available or, or they will de develop things that internally and audit them. But, if we want to have competition, which contributes to consumer welfare, which consumer contributes to innovation, then it is the role

"Regulatory Standards and the Imperative for Public Interest in AI Development"

00:16:47 - just like they're doing an environmental, regulation or occupational safety regulation or transportation regulation, here's the gold standard. Like, you know, you need a a, a safety, system for, for this kind of, vehicle. you need to have a seatbelt in every car. tell us after, you, the regulator has undertaken some empirical study or, or looked at the empirical studies or commission, you know,
00:17:28 - invest, procure, incentivize the private sector to develop, to improve what works, and what best aligns with our public interests. So increasing health, increasing access, increasing safety, increasing equity and equality across sectors. really kind of having a public looking, not just a for-profit looking view of how to deploy that technology and, and develop these standards. And that, that has always been the role of regulation, and we're not doing enough of that now that, you know,
00:18:11 - And, you know, some of this, i I wanna say kind of cautiously that I understand that it's, that's a tough thing when, you know things are moving so fast and they require knowledge and skill that are not always available to, regulators that have not been trained in machine learning and, you know, don't maybe see themselves as expert. and there is, I think that lag that always exists in some, shape and or form, with, you know, the public regulator always kind of catching up to the private market, but I think now it's even more pronounced.
00:18:53 - skill creation in hiring more, professionals in, the public sector that understand the technology so that we can do more of that work of harnessing AI for good. one huge risk is, increased market concentration. And so we're seeing that already with, market dominance. what I mentioned before is that the, the bigger companies will do thi these kind of things anyway. so we have, you know, open AI and we have, Google
00:19:34 - but the smaller kind of new entry, fresh blood, more competition over, you know, pricing and consumer welfare, we'll see less of that because we don't have the, the public arm to offset that disparity in abilities to, to sort to, to develop. so that's one risk. and so it's kind of another, way of thinking about competition law and antitrust policy to think about investment in technology as a public good.
00:20:15 - strongest companies to profit from. It really is a public good that we need to leverage for, the welfare of everybody. And that, again, that's the role of, the regulator. the other thing is that we will see you just more fails, because, you know, if we don't, insist on, the fact that technology can actually correct for stereotypes for, for example, in, in the book, in the equality machine, i, I go through, a lot

"Navigating Gender Stereotypes and Privacy Concerns in AI Development"

00:20:58 - personal digital assistance, to, robots that are designed with like our very like, gendered way of thinking. Like, so, so the private sector that's not, you know, too interested in public looking values will likely have like a, robot that is a doctor designed as a male and, and a nurse doctor, or sorry, nurse robot, you know, as, as a, as a female, you know, you, you're see, starting to see all those kind
00:21:38 - age old just being replicated into our, embodied ai. but a public, looking, you know, role is, or, or, action by, by agencies can be to develop these kind of unexpected surprising, if stereotype flipping or resisting, access or, or availability of, of these tools. other examples, and, and, you know, you might wanna ask me about this
00:22:20 - in next questions, but in the law of AI for good, I talk about how, another fallacy is, our over privileging of the value of privacy versus other, values. And, all along, you know, in, in all of my research, I I very much care about privacy. I think it's, it, it is an important democratic value. I think we should be concerned about, you know, over, surveillance and misuse of information. but I also show that
00:23:03 - and goals and norms that we, have always, strived for in a democracy. And that, if we just have kind of that lead of, okay, now that we have ai, there's even a greater risk that our privacy will be, infringe upon. So do not collect, do not misuse, do not, there's a kind of data minimization at, at the eu, GDPR and the eu, AI act. This kind of a very like controlling lens of,
00:23:43 - of single dominant, rule, then there, and, and it's not contextualized and not offset by other, significant values and interests. for example, having, representative data sets, completing missing, information and missing data. for example, in, in our health data. I have a whole chapter in the equality machine that starts from this understanding that's very well documented in the empirical research on how, clinical trials and, medical device
00:24:27 - have long had this problem of basing a lot of the development on just kind of the skewed partial data of what, the white, you know, male physique biology, is. now we have this opportunity to race ahead with, AI finding more patterns across different, databases that exist in, in hospitals and clinics and, and finding, you know, what works, what, what heals what, what is better patient care for all.

"Empowering Collective Engagement for a Brighter AI Future"

00:25:09 - what are our public goals and how do we get there with the help of private, public, regulation. First and foremost, I want us to all be part of the conversation in the equality machine. I use the term skin in the game. I think everybody, all of us around the world should have skin in the game, in envisioning, in designing, in deploying, in, benefiting from our newfound capabilities of artificial intelligence. And so we can only do that. We actually have rational conversations,
00:26:18 - and what it is that we have now and, and can have, in the future. and, I think that, you know, when we have, these much more balanced conversations and examples of all the wonderful ways that AI can actually assist our lives, then that's when, things start moving along to, you know, the better and brighter future that I envision.

Faculti AI can make mistakes

Research

Policy

Analysis