Latest / The Road to Accountable AI / Derek Leben: All of Us are Going to Become Ethicists

Transcript

Derek Leben: All of Us are Going to Become Ethicists · The Road to Accountable AI

• 0:01
  Derek, welcome to the podcast.
• 0:03
  Thanks, it's a pleasure to be here.
• 0:06
  Like me, you are someone who teaches about AI ethics and responsibility in a business school.
• 0:13
  So what has been your experience engaging these issues with business students?
• 0:20
  It's been extremely positive.
• 0:23
  Business students are excited about the prospects for AI, both internally within the company and also in the products themselves.
• 0:32
  But they are legitimately concerned about the kinds of risks that AI also poses.
• 0:39
  And they're hungry for a serious and honest discussion about those risks in a way that enables them to be realistic.
• 0:49
  about the way that AI can transform business.
• 0:53
  And that's essentially what we do in my classes on AI ethics is we try to take a very honest approach to the kinds of challenges that come up when you're going to be
• 1:05
  incorporating AI in business and ways that companies can be upfront about the policies that they've developed for their AI systems.
• 1:15
  The first book that you wrote, Ethics for Robots, came out in 2018, which is now a number of years ago.
• 1:21
  So, so you've been in this field for some time.
• 1:25
  In general, what have you seen in terms of the development of AI ethics?
• 1:31
  I sort of jumped into this field when it was just getting started.
• 1:35
  I was very excited as a young person about AI and the relationship between what humans are doing when they understand language and what machines are doing when they understand
• 1:46
  language.
• 1:46
  That's what I wrote my dissertation about.
• 1:48
  And I was also very interested in ethics and looking for a way to put these two things together.
• 1:53
  In the early days of the field, AI ethics was about predictive machines and
• 2:00
  Embodied machines autonomous systems and in the 2010s I really focused more on the embodied side of things on things that move around in the world industrial robots drones
• 2:11
  self-driving cars and the kinds of ethical challenges that companies face when developing these systems, but I also shifted more towards predictive systems around 2019 or so 2020
• 2:26
  and of course
• 2:27
  Now after 2022 in the chat GPT moment in November, 2022, there has been more of a shift towards generative systems and eventually agentic systems.
• 2:37
  Now, like you said, these things have been around for a long time and there have been conversations for a long time about them.
• 2:45
  But now that companies are suddenly becoming very aware of them, now that consumers are becoming aware of them, we have a new kind of urgency in our field to
• 2:57
  try to develop concrete recommendations that we can give to companies, but also to regulators, to the general public about how to use these systems responsibly.
• 3:08
  Your new book's about AI fairness.
• 3:11
  Why did you decide to make that the focus?
• 3:14
  So fairness has always been a central issue in my work.
• 3:19
  When I wrote about ethics in 2018, I have an unusual way of thinking about ethics where it is the sort of confluence of what we ought to do as individuals and how goods should be
• 3:35
  fairly distributed to groups or to people.
• 3:40
  And I tend to think about
• 3:42
  ah justice as rightness, which is what Rawls sort of identified in one of his chapters of his big book, Theory of Justice.
• 3:50
  I tend to think that theories of fairness, theories of justice can actually be sort of appropriated and placed into individual decision-making.
• 3:59
  So when I wrote about ethics in 2018 in my first book, I was writing about decision-making as which of these outcomes is going to be the most fair.
• 4:09
  And so when I have subsequently in the years since that book been thinking about fairness, I wanted to come back to this way of thinking about it now applied to traditional concerns
• 4:22
  about how we distribute jobs, how we distribute loans, how we distribute medical resources, insurance claims, all sorts of other kinds of decisions that are now being made
• 4:32
  by AI systems.
• 4:34
  Another answer to that question that's not so personal is that
• 4:39
  AI systems are now being used to make a number of decisions that are affecting the distribution of resources in our society.
• 4:49
  And so this is now becoming a central problem for companies and individuals and regulators.
• 4:56
  Exactly what are the fair allocations in terms of when AI systems make these decisions.
• 5:03
  What's different about asking those questions in the AI context?
• 5:09
  So I like to situate these new problems within the old problems.
• 5:15
  It's true that fairness has always been an issue in business.
• 5:20
  I just taught a class for the undergraduates at the business school at CMU called fairness in business, which was very little about AI.
• 5:29
  And instead it was about these traditional problems whenever companies are making decisions about pricing, about scheduling.
• 5:37
  about resource allocation, they have to make decisions about fairness.
• 5:43
  And so this has always been a problem in business.
• 5:46
  But now that AI comes into the mix, there are a number of changes.
• 5:50
  One change is in the kinds of features that might be used, whereas humans in the past only used a small set of features to make decisions.
• 6:01
  AI systems use many, many features.
• 6:04
  And some of those features might be strange and weird.
• 6:07
  So for example, if you train a big machine learning program to make decisions about lending based on thousands and thousands of features, it might discover some weird
• 6:17
  relationships in that data.
• 6:19
  Like it might discover there's some relationship between how long you charge your phone at night and your likelihood of repaying a loan or between the friends you have on social
• 6:29
  media and your likelihood of repaying a loan.
• 6:32
  And these new features.
• 6:34
  are irrelevant in some ways to the task of lending, but they're not relevant statistically, the machine learning program discovers, to the lending decision.
• 6:45
  And so that's a way where we take these old debates in business ethics and law, say around what used to be called bonafide occupational qualifications in employment discrimination,
• 6:56
  where you say which features are relevant to the company, to the business necessity of the company, which features are not relevant.
• 7:03
  and you apply it to these new cases.
• 7:05
  Another way that AI sort of expands what used to be traditional debate is now there's a new set of ways of measuring things or a new range of measurements that we could apply to
• 7:18
  judgments that were not previously accessible before.
• 7:21
  uh It's also possible to adjust AI systems before the fact, before they make decisions in a way where we couldn't do that with humans.
• 7:30
  Like with a human hiring manager,
• 7:32
  We could sort of do some training and we could do some coaching and then we can measure after he or she makes a decision, a hiring decision or a set of hiring decisions.
• 7:43
  Okay, what were the decisions that this person made?
• 7:46
  But with AI systems, we could do this kind of measurement before we deploy it into the world and we can adjust the system as needed.
• 7:54
  And that's a new way that AI systems actually throw uh or raise fairness problems.
• 8:03
  So let's take that example that you gave.
• 8:05
  A company finds that phone charging behavior predicts loan repayment.
• 8:11
  Is that something that they should be concerned about using?
• 8:15
  How should they make the judgment about whether it's appropriate to incorporate that element?
• 8:20
  claim that yes, they should be concerned about this, which is perhaps surprising because this is not legally a protected attribute.
• 8:28
  So in the U S we have these features that are called protected features.
• 8:33
  And in discrimination law, they include the familiar ones like age, race, gender, and disability.
• 8:38
  And there's a lot of features which are not protected, but are still, you might say, strange or irrelevant.
• 8:46
  Like the shape of your face, uh, like whether you wear glasses.
• 8:50
  like your parents' highest level of education.
• 8:53
  These are not classically protected features, and so they may be legal for companies to use in decisions about a person or about other decisions that affect a person, like how
• 9:03
  much you pay for something or what ads you see.
• 9:07
  But I claim that they may not be relevant to the task itself and that this is a broader sort of ethical mandate that we have as a company, which is
• 9:17
  to use features that are relevant to the task here.
• 9:20
  And I claim that we have sort of broader justification for why we should be doing that.
• 9:25
  So yes, I claim that beyond just the classically protected features that we should care about this broader set of features as well.
• 9:33
  And as a philosopher, what I love is that this gets us into a foundational question of like, well, why are these features protected in the first place?
• 9:40
  Why do we care about these in particular?
• 9:42
  I think that there is something special about things like age, race, gender, disability.
• 9:47
  Um, but that, that what's special about them is that they're historically tools of oppression and that we should care about these irrelevant features, especially, but that
• 9:56
  they share irrelevance in the same way as a broader set with things like, you wear glasses and what's your phone charging behavior?
• 10:06
  Why should relevance, business relevance be the dividing line?
• 10:10
  I could see someone saying, look, I'm running a business.
• 10:13
  My job is to uh have a profitable lending business.
• 10:18
  I totally agree.
• 10:19
  shouldn't make decisions based on things like race, but if I can get uh a more efficient process by using phone charging, who are you to tell me that that's, you can tell me it's
• 10:32
  not relevant, but that it's not ethical.
• 10:35
  Yeah, so this is an old debate and the word relevance is sometimes used in law when we talk about business necessity, bona fide occupational qualifications.
• 10:44
  I like to talk about these two historical cases between Southwest Airlines and Hooters, where both of them refused to hire men for certain roles and they both use the same
• 10:55
  argument.
• 10:56
  They said that it was not
• 10:58
  It was relevant to the business in some way.
• 11:01
  So Southwest Airlines only hired female flight attendants in up till 1981 because they said it was relevant to the business.
• 11:09
  It was relevant to their brand.
• 11:10
  The same with Hooters.
• 11:12
  But Hooters successfully made this claim and Southwest did not.
• 11:16
  Now you could have the approach that the business decides what is relevant.
• 11:20
  It's just up to them.
• 11:21
  If they say it's relevant, then it's relevant.
• 11:24
  But Southwest failed to make that case.
• 11:28
  they did not successfully plead that case.
• 11:30
  said this is relevant and the court came in and said, no, you're mistaken about what is relevant in your own business.
• 11:36
  And we see this sometimes in other cases where say a company like Hobby Lobby says, well, we think that our religious beliefs actually allow us to not give contraception or
• 11:50
  contraception care to our employees.
• 11:52
  And some justices in the Supreme Court, like Ruth Bader Ginsburg came in and said, no, you're just mistaken about this.
• 11:57
  But she also said that companies can't have religious beliefs.
• 12:00
  So there's a number of issues there.
• 12:02
  But the reason I'm bringing up these cases is that what's relevant, at least in discrimination law, historically is not just whatever the company thinks is relevant.
• 12:12
  But then there has to be some objective sense of what is actually relevant to this task.
• 12:16
  And that's a really difficult problem.
• 12:18
  I tend to agree with the philosopher Michael Sandel.
• 12:21
  that there are certain objective facts about the nature and purpose of certain kinds of tasks that we can appeal to, that some activity like lending or hiring ah essentially
• 12:34
  involves certain kinds or reference to certain kinds of features that we can call qualifications for paying back a loan or for doing a job.
• 12:45
  and that we should be engaged in this conceptual task of saying, what are the features that are part of the nature and purpose?
• 12:51
  Part of that we could actually do a priori when we're engaging in just a sort of neutral discussion about the task itself.
• 12:59
  So in this Rawlsian framework that I prefer, I could imagine, I don't know who I'm going to be, a person doing the job interview or the person on the receiving end of the job
• 13:10
  interview.
• 13:11
  And I could say, well, in a perfect bargaining position,
• 13:14
  which of these features do I think should be used by the party actually doing the interview, which of these should be irrelevant?
• 13:22
  And I claim that in that sort of ideal bargaining position, there are some features that we would all agree are relevant and irrelevant.
• 13:30
  As you point out though, AI systems can take into account many, many different features and these models sometimes are difficult, if not impossible to explain in simplistic terms
• 13:43
  that this was the factor that led to the result.
• 13:46
  So how do we decide what's fair and appropriate in that environment?
• 13:52
  So there's two approaches here.
• 13:54
  One of them is to try to discover the role of features in a decision or in the content that was generated.
• 14:04
  And for that, we have a whole host of methods that computer scientists have developed over the last 10 years, sometimes called explainability methods.
• 14:12
  And there's a lot of different kinds of explainability methods, but the kinds of methods that I think are extremely useful.
• 14:19
  are ones that give us counterfactual information.
• 14:22
  They tell us something like, okay, if you had been a different member or I should say a member of a different group, if you had been a man, would you have been approved for this
• 14:33
  loan?
• 14:33
  Would you have been approved for this job?
• 14:35
  And there's also some basis in the ethics and law of discrimination around this.
• 14:39
  There's a test called the similarly situated person's test, which tries to establish something like this as a basis for discrimination.
• 14:46
  But the problem with that test is historically it's very difficult to find someone who is similarly situated in practice.
• 14:54
  But in AI systems, we can do this fairly easily just by changing features and looking at the impact on the decision.
• 15:03
  And so those are some methods we could use to try to say what role is the feature playing in the decision?
• 15:09
  But there are other kinds of approaches we could take as well.
• 15:13
  And one of them is just looking at the outputs
• 15:16
  and doing certain kinds of measurements on the outputs of a model, either the content that's generated, if it's a generative system, or the decisions, the predictions, if it's
• 15:25
  a predictive system.
• 15:26
  uh And there's lots of different ways of measuring these outputs.
• 15:30
  So for example, in traditional discrimination law in the US, we have measurements like the four fifths rule, which is applied to usually the rate of approval across different
• 15:42
  groups.
• 15:43
  So
• 15:44
  I might say, okay, what's the rate of approval for men versus women in this loan?
• 15:51
  if the rate of approval, excuse me, if the rate of approval for men is, should say the rate of approval for women is less than within a margin of 20 % than the rate of approval
• 16:05
  for men compared to that rate, then it's a fair decision.
• 16:13
  So that's
• 16:13
  another method, but we have more sophisticated measurements with AI systems now where we can look at not just the rate of approval, but the rate of qualified approval.
• 16:23
  can conditionalize on qualifications.
• 16:26
  And of course that depends on our definition of qualifications.
• 16:29
  We can also do different kinds of measurements depending on how we conditionalize.
• 16:35
  So we could say, what's the rate of qualification given that they were approved as opposed to the rate of approval given that they were qualified.
• 16:44
  And there are different measurements as well.
• 16:45
  So in the book, I proposed that the most important or the most justified kind of measurement to use is rate of qualified approval across these different kinds of domains
• 16:59
  and then followed by approval rate and then followed by some of these other measurements.
• 17:04
  So it's a sort of hierarchical recommendation that I have.
• 17:08
  And just so it's clear what exactly does Rate of Qualified Approval mean.
• 17:12
  Right.
• 17:12
  So you could imagine if you're in the lending space and you have two groups, group A and group B, and you say something like, well, there's some historical reason why group B is
• 17:28
  less likely to repay loans because they've been deprived of wealth and opportunity and education and so on.
• 17:35
  systematically their qualifications are shifted.
• 17:39
  down.
• 17:39
  So they're just less likely as a group to repay this loan.
• 17:42
  ah What do we do in that case?
• 17:45
  Well, we could say we're going to give more loans to group B or sort of lower the qualifications for group B.
• 17:52
  And this is one approach.
• 17:53
  And there's an issue with this, namely it bumps up against what's called differential scoring in the U.S.
• 17:59
  indiscrimination law and also in the ethics of affirmative action, where you might say there's some kind of
• 18:05
  both violation of procedural fairness or procedural justice here.
• 18:10
  And it even produces maybe worse outcomes for the vulnerable group, for the group that we're trying to help.
• 18:16
  This is a kind of consequentialist argument that's been made often against fairness, which is that in trying to provide equal rates of approval between group A and group B in
• 18:28
  lending, we're giving out more loans to group B when they can't pay back those loans.
• 18:33
  and then they get into terrible debt and it's worse for them.
• 18:37
  It winds up producing worse outcomes.
• 18:39
  So how do we measure fairness?
• 18:41
  Well, we could say something like of the qualified members of group B, what was the approval rate of the qualified members of group A, what was the approval rate?
• 18:51
  And then we could try to equalize those rates rather than just say, let's equalize the pure rate of approval across groups A and B.
• 18:59
  What's interesting is that you started with ethics and philosophy and asking what's right and what's justified just as a matter of principle.
• 19:09
  And then what you got into is something that's very concrete and practical.
• 19:13
  So for a company thinking about doing this is the ultimate result.
• 19:19
  Therefore, buy some software, use a tool that will implement the following tasks, push a button, and this is how you decide fairness.
• 19:29
  Or do companies still need to uh cultivate the ability to ask the big questions about what's fair here?
• 19:37
  Definitely the latter.
• 19:38
  Definitely the latter.
• 19:39
  So that's my approach that I always take to ethics, which is that it always involves doing the hard work of trying to justify your decisions to others who disagree with you.
• 19:51
  There's never going to be a checklist.
• 19:53
  There's never going to be a switch that you pull and say, okay, we've, we've, uh, we are compliant as my legal friends like to say.
• 20:04
  Instead, it is about.
• 20:06
  trying to defend your decisions and give reasons for why you did what you did.
• 20:11
  And so in this sense, ethics is a never ending task.
• 20:14
  It is a constant task of trying to defend your judgments, defend your decisions, defend your practices as a company to others who are skeptical about it.
• 20:23
  And so especially right now where the regulations around AI are, to put it politely, unclear and uncertain.
• 20:35
  in the U S but also when new technologies are coming along that expand what we traditionally might have thought of as unfair practices.
• 20:46
  Companies need to take a stand and say, okay, here in the absence of tradition and precedent in the absence of clear regulation and guidelines or industry standards, here is
• 20:58
  our decision about what we think is fair, how we are.
• 21:02
  mitigating our models, how we're not mitigating our models and why not.
• 21:08
  So here's an example with image generators, which is that we saw with Google and Gemini a couple of years ago, this problem where Google tried to mitigate their models to achieve
• 21:19
  fairness across race and gender groups.
• 21:21
  And it wound up producing some very surprising and sometimes very silly results.
• 21:26
  Like if you ask for an image of the American founding fathers, it would give like 50 % women and
• 21:32
  20 % of them were black and it was a sort of silly result.
• 21:38
  Right, right, exactly.
• 21:40
  Yeah, but according to, you're adjusting your models for demographic parity, this makes sense.
• 21:44
  If you're just trying to get equal rates of representation across groups, then the US population is 14 % black, it's 50 % female.
• 21:51
  So you would expect, look, we should get these rates in the images that we generate.
• 21:56
  Now, Google walked that back and said, okay, we're not gonna do that anymore, but...
• 22:02
  If you're generating images, have to have a policy about this.
• 22:05
  You can't escape it.
• 22:06
  There's no escaping it.
• 22:07
  You have to say either we are going to mitigate our models or we're not.
• 22:11
  And obviously a more sophisticated approach would be something like, okay, if I ask for an image of a doctor, like a generic term, then we can mitigate the models for age, race,
• 22:19
  gender, disability, those kinds of traits.
• 22:22
  But if I'm asking for historical cases, then maybe not.
• 22:26
  But once again, that itself is an interesting substantial claim.
• 22:30
  And there are some companies
• 22:31
  that say we're not going to mitigate our image generators or our language models to try to achieve fairness, even in generic representations.
• 22:39
  And so this is exactly why companies need to have this debate is to say, okay, here's what we're doing and why, so that they're not just being responsive to criticisms and the
• 22:50
  public when there's some catastrophe that happens.
• 22:53
  I mean, to put it...
• 22:55
  As, as practically as possible, we want to be proactive rather than reactive and actually saying here's what we're doing and why, before people start getting very upset about it
• 23:05
  is, the right way to go.
• 23:07
  In talking with companies, uh what do you find that they struggle with most in trying to address these issues effectively?
• 23:17
  Well, this has changed a little bit in the last year.
• 23:20
  I would say in the last year or so, the big shift has been around what we're going to do with respect to compliance in, in federal law in the U S for obvious reasons.
• 23:34
  The Trump administration has made it clear that they are going to be both trying to change the laws around.
• 23:44
  discrimination, but also not prosecuting certain cases and that they're going to be unfriendly to companies that are mitigating their models to try to achieve fairness in
• 23:55
  ways that were perhaps legal before and even are legally required in the European Union.
• 24:00
  ah So the first thing that I encounter when I actually talk to people in companies is what are we going to do in the legal and the regulatory landscape?
• 24:10
  Now, I unfortunately don't have an answer for that.
• 24:12
  uh, politics and law are not my area of expertise.
• 24:17
  However, I can say that in the absence of clear guidance from the law that we can say, okay, we think we're going to mitigate our models or we have an ethical obligation to
• 24:27
  mitigate our models for these reasons.
• 24:29
  Another challenge I often find when dealing with companies is the, the motivation to actually have
• 24:40
  these uncomfortable, these difficult discussions and to make it clear and transparent to the public exactly what we are doing that is perhaps surprising.
• 24:51
  And so here's an example.
• 24:53
  I was doing an executive education talk a couple of weeks ago and there was an executive who said, I think we should just do the least controversial thing.
• 25:02
  And my response was there is no least controversial thing.
• 25:06
  Every single thing you do.
• 25:08
  is controversial.
• 25:09
  And I would say not just around fairness, but around most of these ethical issues, everything you do is controversial.
• 25:15
  And that's another attitude.
• 25:16
  You know, the, I mentioned the sort of compliance mindset, but also the, let's just do the least surprising, least controversial thing.
• 25:26
  And I'm always emphasizing there is no such thing.
• 25:29
  Everything you do is going to commit you to something surprising or controversial.
• 25:35
  And as much as it pains us to do that, we have to be honest and upfront about it and actually tell our investors and our customers and the public and the regulators.
• 25:48
  Yeah, here are the costs.
• 25:49
  Here are the sacrifices.
• 25:50
  Here are the surprising predictions.
• 25:52
  Here are things that people aren't going to like about what we're doing.
• 25:57
  Here are ways that what we're doing is different from what our competitors in the same field are doing.
• 26:03
  So.
• 26:03
  these are the kinds of challenges that I try to emphasize is actually saying, here's where we draw the line, here's some surprising predictions, here is what we're going to say that
• 26:15
  you're probably not going to be happy about.
• 26:18
  Yeah, and this gets back to where we were starting the conversation.
• 26:20
  You were talking about business ethics in general.
• 26:23
  uh That's not a message that companies necessarily want to hear.
• 26:29
  Maybe it's a message that they need to hear, uh but it's not a pleasant thing to think that there's no easy way out.
• 26:36
  So what's the argument to convince companies that they really need to do that work if, let's put aside, it's not legally required in the current environment?
• 26:48
  So there's a stereotype of business ethics that we're just coming in and telling companies not to do things.
• 26:52
  And that's not true.
• 26:54
  We are telling people how to do things in a responsible way.
• 26:57
  And that's usually what I emphasize is I'm not anti AI.
• 27:02
  I love AI and I always have.
• 27:05
  And I love the kinds of potential that it has for both operations and products inside of companies.
• 27:13
  However, I want to do it in a responsible way.
• 27:16
  When we saw, for example, the Senate testimony where Sam Altman testified before Congress a couple of years ago, the members of Congress on both sides of the aisle in the US said
• 27:27
  that they thought they had failed with social media and they were trying to correct that failure to be proactive now with AI.
• 27:37
  And I think social media is a good analogy.
• 27:39
  So our colleague
• 27:42
  In the field, John Height has argued for many years that social media companies are being reckless and irresponsible and they should be doing more.
• 27:51
  And I think both the companies and all of us as, as consumers and customers of social media should also be sort of navigating this space in a responsible way.
• 28:01
  And we're starting to develop the, norms around social media, despite the kinds of harms that have occurred to say our attention spans and our social relationships over time.
• 28:12
  I think the same applies to developing AI systems where we're going to get to a place, my naive optimism says, where we develop industry norms and standards, eventually perhaps
• 28:26
  regulations around AI systems.
• 28:30
  It's just going to be a hard road.
• 28:34
  And I think about ethics more as navigation along this road.
• 28:37
  It's not coming in and
• 28:38
  constraining us and telling us what not to do.
• 28:41
  It's telling us how to be a good driver.
• 28:44
  My colleague, JS Nelson, who teaches over at Pitt likes to give this analogy between ethics and law where she says that law is kind of like the guard rails on the highway
• 28:54
  where if you're hitting the guard rail, you're already a bad driver.
• 28:57
  That's not good.
• 28:58
  And ethics is in her analogy, the way of driving, the way of being a good driver, right?
• 29:04
  So there's lots of legal ways.
• 29:08
  of being a bad driver and we want to be good drivers along this road and get someplace ah in a way that's not just the fastest route, but in a way that is also a comfortable route
• 29:21
  that we feel proud of and happy about, that we had a good trip, that we did not put any risk on other road users to follow this analogy here.
• 29:31
  And so that's, that's usually how I think about ethics is not as a constraint, not as saying no, but instead as working with companies hand in hand to say, okay, we want to
• 29:41
  help you do this.
• 29:42
  We think there's a lot of potential here, but how can we do this in a way that is respectful, responsible, fair, and so on.
• 29:52
  And how do you see the next few years playing out?
• 29:56
  Because as you mentioned, we have these political shifts, we have different views in different countries, the technology is evolving very fast.
• 30:05
  What's the argument that might give us optimism that that evolution towards norms and standards and ultimately regulation you talked about will actually happen?
• 30:14
  So there's two big things that I think we could say for sure about the future.
• 30:20
  One of them is that in the U S there's not going to be any regulations anytime soon.
• 30:27
  And the second is that the industry is going to push forward with rapid speed accelerated by the sort of geopolitical battles that are happening right now between say the U S
• 30:41
  China, the EU and others.
• 30:43
  So
• 30:44
  It's not like AI is going away anytime soon.
• 30:48
  If anything, we're going to see more and more of it over the years inside of operations and the way that people are doing work and in products and services themselves.
• 31:01
  So those two things create an interesting new landscape where all of a sudden I think there's going to be a great demand for ethics.
• 31:15
  at both the company level and the individual level.
• 31:18
  And by the individual level, I mean things like, do you use ChatGPT to write letters to your mom or respond?
• 31:25
  Nobody writes letters, write emails or texts to your mom.
• 31:28
  Do you use ChatGPT if you're a therapist in responding to your clients, even in just everyday life as a student, how do you use ChatGPT to write your papers, that kind of
• 31:39
  thing, right?
• 31:40
  So at every level, we're going to see more and more of this happening.
• 31:45
  And I don't know what is going to incentivize people to take ethics seriously aside from just the undeniable need for it.
• 31:59
  ah If you have the absence of regulation and precedent, and you have the industry moving full steam ahead in a way that is transforming our lives, I think that all of us are
• 32:14
  going to become ethicists in the near future, whether we want to or not.
• 32:22
  Derek, thank you so much for the conversation.
• 32:24
  My pleasure.