Legal Cyber Academy

Ethical Concerns of AI Use

Free. Watch without an account. Sign in to save progress and take the exam.

Up next

Sign in to take the examPass it for a verifiable Certificate of Completion

Then

TranscriptThe player shows YouTube’s automatic captions, which are machine captions we do not review. The full transcript is below.

00:34:02ethical issues as we jump right into it, right? So there's all of these concepts that get batted

00:34:08around, inclusivity, transparency, accountability, IP, legal ethics, fairness, and all of these

00:34:14things. So this is the laundry list that I don't think we're going to cover all of them,

00:34:19but I think we'll just jump right in and sort of start the conversation around fairness and

00:34:23bias. And fairness and bias can be something as simple as they stole my candy bar, or that they

00:34:32got a bigger piece of dessert. What is fair? What is bias? To something as complex as hiring somebody,

00:34:40or the way you interact with people. The data training has bias because of the people that

00:34:47participated in the development of the system. So it can be as germane as dessert to as complicated

00:34:53as full on discriminatory against whatever the unknown or known either way. But with that in

00:35:01mind, I mean, we definitely have the people here to weigh in here. So maybe Karen, Moira, and Jenny,

00:35:06in that order of fairness and bias, I think, and then we'll talk about reliability and safety.

00:35:11But I think fairness and bias is a really hot set of terms that lawyers are talking about today.

00:35:17So maybe Karen, Moira, and Jenny? Yes. So I'll just kick it off by saying that I think

00:35:24here it's really helpful, again, to think about AI, not just as the model, which is a piece of what

00:35:30we're talking about. We've just described how the technology or how the model works.

00:35:37The logic running on a big pile of data is the model, and that produces outputs

00:35:44based on a bunch of inputs that it's been trained on and then produced on. I think when you're

00:35:49talking about fairness and bias, though, it's not enough just to talk about the model and whether

00:35:53the model is skewing in some direction because it's built to skew in a direction. You really

00:35:58got to think about sort of the whole system, which is the humans, the data, and the technology,

00:36:05right? And what have the humans asked the model to do and is the objective that the

00:36:11humans have set for the model fair and unbiased or even just appropriate to the task, right? And

00:36:17how we frame those questions actually dramatically impacts sort of what kinds of outputs that the

00:36:24model produces. So when I think about fairness and bias, I really think of it as the whole system.

00:36:31Was the business, let's just take a business case, was the business objective a fair,

00:36:37reasonable, and accurate business objective? Were the people who sort of articulated that,

00:36:42did they articulate it in a way that does not introduce new kinds of bias? Are the data

00:36:47themselves that were used either in training or production skewed in some way because they

00:36:52reflect historical practice? You know, does the math tweak that data in a particular direction

00:37:00or not? You know, and then importantly, the model produces an output. What we do with that

00:37:06output, like I said, particularly if we're talking about a decision support tool, again,

00:37:10those are humans who are sort of trying to figure out what to do with those model outputs and

00:37:14fairness and bias can come into it at that point as well. So when we talk about fairness and bias,

00:37:20we've got to be really clear about at what point in the life cycle we're talking about it,

00:37:24whether we're talking about sort of the human dimensions of that issue or the machine sort of

00:37:30trustworthiness pieces of that issue. And to get an AI application to work well or within

00:37:40bounds of what we want to describe as fair and unbiased, we really need to be prepared to look

00:37:46at all of those steps. And in a kind of non-defensive way, I think there's a tendency to

00:37:50say, well, the algorithm was biased. The algorithm might have been biased, but probably something

00:37:57preceding the algorithm was reflecting some bias that sort of showed up in how the algorithm

00:38:04was then used, right? So it's really a very complex systems analysis. And I would just add to that

00:38:13what's interesting in a lot of the work that we're doing here is that I think there's general

00:38:17consensus that unlawful bias is something that if it's not, it's sort of non-controversial,

00:38:25right? If we can't discriminate not using AI, we should not be able to discriminate using AI,

00:38:30right? But what is unlawful bias is actually a quite narrow set of concepts. It's a really hard

00:38:36problem, but it's a narrow set of concerns and considerations. I think there is a broader set

00:38:42of fairness and bias questions that don't deal with necessarily suspect classes or anything like

00:38:49that that still affect business performance or operational efficacy and things like that.

00:38:55So the concepts in business of fairness and bias around AI are bigger than just whether

00:39:03they are operating lawfully or not lawfully, right? And so you don't necessarily want to

00:39:08discriminate against people who live in houses with red roofs if you're underwriting loans,

00:39:12right? But these tools can buy us off on characteristics like that. So there's sort

00:39:20of a math and business problem around bias. There's a law problem and an ethical problem

00:39:27around bias. And you have to look at that across the lifecycle of your application.

00:39:32So I'll stop there, but I just want to make sure that when we're talking about fairness and bias,

00:39:38we're locating ourselves in which part of the problem is coming up.

00:39:44Yeah, I think that's a really good point. And I think a lot of times you hear about,

00:39:48oh, the training data includes biased information because the world is biased and the experience

00:39:57that we have to date is biased. So when you're talking about, for example, employment software,

00:40:03employment screening, which of these candidates is more likely to make the final cut,

00:40:09chances are the software is going to predict a white male just because that is kind of how

00:40:16the universe has worked. But then also when you're thinking about what the inputs are,

00:40:23are the inputs biased? And some of that goes back to, do we have the right prompt engineers?

00:40:29Do we have people who have been trained to ask the AI tool the right questions?

00:40:34And some of it also comes from the data that you're feeding the AI tool to make decisions.

00:40:39And just as an example, I mean, we hear about AI tools being used to make credit decisions.

00:40:46And that's actually something that a lot of the regulators in the financial industry are

00:40:51looking at and thinking about. Are you making the right decision? Are you making a decision

00:40:57that's explainable? And are you basing that decision off the appropriate data?

00:41:02Like I've seen examples of credit decisions being made based on whether your phone battery is low,

00:41:10or what your friends are doing on social media. And does that, or people you follow on Twitter,

00:41:20does that perpetuate biases in a way that is unfair? And I think that also goes to some of

00:41:28the things in the other list that you talked about on the previous side, Dan, where we talk about

00:41:33transparency. And should we be disclosing to consumers, hey, yes, these are the types

00:41:42of information that we're using to make decisions about you? And can people object to that?

00:41:51Oh, I agree with Karen that you need to look in three places, really, if you're interested in bias.

00:42:01The first is the data. So if the data the system was trained on is not sufficiently

00:42:09representative of the data on which you're using it to make a prediction, you're going to get a bias.

00:42:17So for example, with facial recognition systems that have not been trained on an adequate number

00:42:24of persons of color, they will work pretty well on white male faces, and they work pretty poorly

00:42:32on black women's faces. Now, if we go to Gen AI, and we ask mid journey, which has been trained on

00:42:40the internet, give me four pictures of a lawyer, you'll get four white men. And if you say give me

00:42:48four pictures of a felon, you will get four black men. And that is because the data in which

00:42:56the system was trained contains that bias. The second place is the algorithm. And by algorithm,

00:43:04it could be the algorithm itself, if it's coming up with its own weight. But often, it's a human

00:43:10who's who's a developer, who's made either an incorrect decision, or doesn't completely understand

00:43:19the workings of the algorithms. Two examples there would be if I want to predict health care,

00:43:29how much money did you go to certain health care programs in certain neighborhoods?

00:43:34And if I use expenses or costs from last year, as my predictor, I'm going to have a very biased

00:43:42algorithm because people of color who live in neighborhoods that don't have sufficient health

00:43:48care coverage, they don't use these services as much as they need them one because they have to

00:43:55travel further to a doctor or hospital to because maybe they can't get time off from work for the

00:44:01kind of jobs they have, three, maybe they can't get childcare coverage. So if you use, if a developer

00:44:08uses health care costs to predict health care needs, you're going to have a biased algorithm.

00:44:14Second is what we call proxy variables. I don't ask race. I don't ask that question, but I ask

00:44:21address with zip code. And zip code is highly correlated with race. So race gets pulled into

00:44:27the algorithm through a proxy variable. And then the third part is the person who's interpreting

00:44:34whatever the output is. And there's a phenomenon called death by GPS. And these are people who

00:44:42will literally drive their car off a cliff or into the ocean because the GPS tells them to continue.

00:44:48The restaurant is only a little bit further. And just as you have people who think it came

00:44:56from an algorithm, it must be objective. You have other people who will never, it's called

00:45:02algorithm aversion, who won't listen to it. We see that with technology assisted review,

00:45:07we need discovery. It's very clear from the science that the algorithms are better than

00:45:12lawyers at finding responsive documents, but there are lots of lawyers who don't trust the

00:45:17systems who won't use them. So this bias issue just pervades AI and is really challenging

00:45:30to get out. Can I add one thing because I'm already really raised an important point.

00:45:35I mean, several important points, but sorry about taking the zip code data, right? It's not that

00:45:40the zip code data is biased. It's that if you're using it for underwriting purposes, it could be

00:45:46biased or could contribute to a biased outcome. If you're Amazon and you want to deliver a package

00:45:51to somebody's specific house, the zip code data is actually essential, right? Just different use

00:45:56case, same data, right? So we've got to be really clear about where we're, the bias is not just a

00:46:03function of the data point. It's a function of how that data point is being used and incorporated

00:46:10in an outcome that impacts somebody, right? And I just think that we need to sort of

00:46:19be prepared to focus on those sorts of distinctions and ask really hard questions about use cases,

00:46:26ask really hard questions, as Jenny said, about how prompts and objectives are set in the first

00:46:31place and how engineers, as you said more, are sort of bringing their implicit understanding or

00:46:38assumptions about the world into the math, right? And at each of these steps and about 17 others,

00:46:49skew can come into the system, which is why we need to sort of supervise them as humans to make

00:46:53sure we're getting the outcomes, not just the outputs, but the outcomes that we want,

00:46:58which I would distinguish from outputs, which are more a model production, right?

00:47:02Yeah. And I think that's a great play. I mean, the net is that I tell people fairs and bias is

00:47:09really a reflection of society and the users are using it as upsetting as that may be.

00:47:15Data is a reflection of what our current, it's just compounded. If you want to think about it

00:47:20at a practical level, when you take something a hundred times versus a million times, it's just

00:47:25inherently becoming more compounded or visible to the end user and it's reflected accordingly.

00:47:31One risk point, if you ask it for X, you get whatever the data or whatever, it's just

00:47:37compounding the problem and you have to talk to your clients and make sure they're aware of that

00:47:42and that you've balanced or you're addressing it to manage the risks and outputs that you're getting.

00:47:48No solution's perfect, but you need to make sure your clients have properly evaluated

00:47:52that risk profile and the downside impact. And then reliability and safety. So,

00:47:59I'm going to combine reliability, safety, and privacy and security just in the interest of time.

00:48:09And I'm going to say that your example of going off a cliff probably is a great place to start about

00:48:18what is appropriate security controls and what are the people that are looking at these ethical

00:48:24issues. Where does the line get drawn and what is appropriate? And so, maybe Moira just your quick

00:48:30comments and Jenny and Karen just quickly and then we'll jump to the next group there.

00:48:35So, going back to the prior slide, I think it's the reliability and safety slide. It's really

00:48:42important to distinguish between two concepts. One is called validity and the other is called

00:48:47reliability. And people mush these together and use one to refer to both and both to refer to one

00:48:55and they really should be separated. The validity is does the system do what it's supposed to do?

00:49:02Does it predict what it's supposed to predict accurately? So, if I want to know Karen's height

00:49:14but I put a scale in front of her, that may be a perfectly good scale but it's not going to measure

00:49:20her height, it's going to measure her weight. That would not be valid measure. Now, we must also look

00:49:30at reliability which does the tool measure accurately across time under similar circumstances.

00:49:40So, if Karen gets on the scale three times, she shouldn't get a hundred pounds difference each

00:49:46time she steps on the scale and maybe with a margin of error, it should be a pound or two

00:49:51different maybe. So, for AI systems, for us to be able to trust them, they have to be both valid

00:50:00and reliable and often there may be one and not the other but that's critical to ensure

00:50:09when you're using the system that it's going to be both of those things. So, that's my comment on

00:50:15reliability and safety and if we move to the next slide, I think we've seen a lot about protecting

00:50:26privacy and the security particularly around the generative AI systems where people don't seem to

00:50:36understand you can't just put your client's confidential information into these systems,

00:50:43public ones at least because you may be giving it away, it can be used for further training

00:50:51and all kinds of things. The other thing we've seen is that at least many of the most AI systems

00:51:00and in particular generative AI systems are not completely robust, it is easy to attack them

00:51:08with what we call adversarial attacks. So, somebody asked, I think it was chat GPT to

00:51:17just repeat the word poem infinitely and it started poem, poem, poem, poem,

00:51:24and after a thousand poems, it started spitting out people's phone numbers and addresses

00:51:29and all kinds of personal information. So, we have to be very careful about

00:51:36how safe and secure these tools are in addition to whether they're measuring properly

00:51:43and whether they're doing that consistently under similar circumstances.

00:51:50Yeah, I guess what I would say there is go back to the reliability and safety and privacy and

00:51:54security both. I think this is a good example of where we have to be really clear whether we're

00:51:59talking about the human systems or the technology systems because we can build reliability into

00:52:05an algorithm that if the human process into which it is inserted is not reliable,

00:52:10it will perform unreliably. So, there's that problem. There's also a technical problem which

00:52:16is particularly for some of these models that use what we call reinforcement learning,

00:52:20meaning that their outputs sort of get regenerated into their training.

00:52:25I think, you know, they do have a tendency to drift from their original purpose. So,

00:52:31they require monitoring in a way that traditional software or a locked model might not. So,

00:52:37we have to sort of just understand what kind of technology we're talking about when we talk

00:52:41about reliability and safety or privacy and security and we have to talk about what kind

00:52:45of system, human system we're putting it into. So, for instance, the poem example,

00:52:51I think it's a really good one. I think that is a, I don't know if it's a, we'll call it a flaw

00:52:57in the way that the generative AI model worked. It was also a byproduct of a human who was trying

00:53:04to either test it or make mischief, right? So, we have to have, when we're talking about privacy,

00:53:10security, reliability, and validity and all those sorts of things, we need to be looking at both

00:53:14sides of the equation which is what are going to be acceptable or unacceptable use cases or

00:53:20uses for these technologies and when we put them to those acceptable uses, are they performing in

00:53:24a way that given the use case, you know, the results are valid or reliable or dependable.

00:53:31I want to back up too and say that these tools in general, we haven't talked about this,

00:53:36are really, they're good at sort of some of these brainstorming exercises when we really don't

00:53:41know what good looks like and we just want to kind of stimulate some, you know, unique content

00:53:46to get us thinking. But where they're really useful, and most business applications at this point

00:53:52are much more in the other category, which is where they're replicating a really well-worn,

00:53:57understood process that we can start to automate because we know what good looks like,

00:54:02right? We've got lots of examples of good. We know what tolerances we have for variations from

00:54:07good. Like, we understand the problem really well and we can sort of automate the form of the

00:54:12problem in the form of the response. Most problems aren't either one of those categories or something

00:54:18in the middle. And so when you talk about reliability and validity and privacy and security, you've

00:54:23really got to, again, you've got to be so use case and context specific because depending on the kind

00:54:31of problem you're asking these models to work against, they will inherently be better or worse

00:54:38at some of those tasks. And you will need different kinds of governance and process

00:54:43around those models to make sure you're getting the outcomes you want and can rely on.

00:54:47And those processes will just differ widely based on use case, not even organization or

00:54:54industry, but really use case. That's right. And I do want to go back to something that you said

00:55:01at the very beginning of this lecture, which is, what does it really mean to regulate AI?

00:55:07What does it mean to have to worry about use of AI? And it really is system dependent. When you're

00:55:14talking, when you're speaking as a lawyer and you're thinking about legal ethics, you're not

00:55:18just worried about putting your client's name and personal information into an AI tool. You're

00:55:25worried about doing something that could potentially disclose a legal strategy that the AI tool could

00:55:32be smart enough to infer relates to a particular client. For banking and financial institutions,

00:55:42you're not only worried about the privacy and security of your client's confidential information,

00:55:48you're also worried about your own trades and whether those trades are actually going to be

00:55:53used to inform some other trading model. Everything is context specific. You need to be thinking about,

00:56:05even if something is just on your own systems, what internal firewalls do you need to be setting up

00:56:12when you're working at a law firm? There's ethics walls in place. When you're working at

00:56:17a financial institution, there's walls between investment banking and research.

00:56:21So I think everything really is context specific. And what we continually advise our clients is

00:56:32any new technology, it's not about the technology itself, it's about what is the technology

00:56:36designed to do? What are you marketing it to people? What are you telling people that it can do?

00:56:43And I think that's the thing that you have to worry about regulation with.

00:56:49And that actually ties into some of our other slides. So I am going to hold off for now and

00:56:55continuing. Yeah, so I think that I think I'm going to come, you know, inclusiveness and

00:57:02transparency, right? You know, the kind of, you know, and accountability to me sort of all end up

00:57:09becoming focused on the same outcome for your client, which is how do you man if it's customer

00:57:15facing or internal facing? How do you manage the risk of the tool? Right? I mean, if you think about

00:57:20it, these are sort of, you know, how do you make sure that it is appropriate and using the data set

00:57:27so you don't end up with the four black men as criminals coming back and the four white men as

00:57:33lawyers or whatever more. I mean, I've done it for judge, you name it, you can find biases and

00:57:38whatever. So how do you make sure your client is using that appropriately? And then you have the

00:57:43transparency, which sort of comes along with the inclusiveness, which is, you know, can you figure out

00:57:49what is actually, how does it all work? Right? Do you actually understand how it is connected to

00:57:55give you that? And then at the end of the day, this is why I tell people you are accountable,

00:58:00like nobody's going to care what the lender is the lender, right? Whether it was AI or a

00:58:07racist, hating, lending, institutional loan officer. They're going to look at, okay, you only,

00:58:18you denied only these people's loan applications and only accepted these loan applications.

00:58:24And I tell clients all the time, the accountability isn't something that you can blame on AI. You

00:58:30ultimately are going to be the individuals accountable, whether it's to a regulator,

00:58:34whether it's to the litigants or the other side of the V or to your employees. So those are sort of

00:58:43like quick two cents on those three concepts. If maybe Moira, Jenny, Karen, you just want to

00:58:49quickly weigh in on what your thoughts are there. Just again, I'm trying to keep it somewhat on

00:58:54time. Sure. I mean, I think this is really, it's really about whether you are responsible

00:59:03for the technology that you use as someone who represents vendors and who represents users of

00:59:08products. The advice that I give in both cases is you are, you, the user are responsible for the

00:59:16outcomes of the tool that you use. You're responsible for evaluating the tool that you use.

00:59:22But one of the more interesting questions we have here is whether the providers of AI tools are

00:59:27going to accept liability from a contractual perspective. Because if the service is really

00:59:35specific to a particular use case and the provider can understand the universe of applicable

00:59:40regulations, they may be comfortable making reps and warranties about that. But where, in the case

00:59:48of AI tools, the customer is going to have a very good argument that they can't actually go in and

00:59:53make the determination of can this tool be used in accordance with the regulations that apply to me

00:59:58when you've got a completely black box tool. So I think that's just something that vendors and

01:00:06users of AI systems are going to have to work out amongst themselves because normally you would say,

01:00:13okay, I tell the system to do X, it does X. Here you're telling the system to do X,

01:00:20and it might eventually get to X or it might get to Y. And in some cases, you really don't know

01:00:26the reason for that. And I think that's one of the debates that everybody's having about explainability

01:00:32of outcomes versus explainability of process. Can we at least explain why the system came to

01:00:37a particular outcome? And then I think people are going to be more comfortable using it,

01:00:44even with the knowledge that they are going to have liability for its use.

01:00:49Well, I'll say quickly, on inclusiveness, if you're developing tools, you need to have a very diverse

01:00:58team. Because if you're developing a crash test body, and it's a bunch of males, you're going to

01:01:08design a male crash test body, and they didn't develop women's bodies and pregnant bodies until

01:01:15many, many, many years later. So inclusive teams and include discussions with the people who are

01:01:24going to be using the systems, get stakeholders involved tends to be a fair system. As far as

01:01:31transparency, the real question is at what level of transparency, because there's going to be what

01:01:41a computer scientist could understand, and that's going to be very different than a user. And usually

01:01:49somebody who's had an adverse loan decision is going to want more information than somebody who's

01:01:55had a positive loan decision. They've gotten their loan, they're happy, they don't care

01:01:59how the algorithm works, they want to move on. And I think transparency becomes less important

01:02:06when you've done that validity and reliability. So think about the FDA, we all take drugs that

01:02:13we don't have any clue how the mechanism and how they work in our body, but the FDA has validated

01:02:19them and checked them for us and we're comfortable. So we hope we have most of the time, not always.

01:02:25But so transparency has to be balanced, because sometimes you have to give up accuracy for

01:02:31transparency, use the simpler system that isn't as good as deep learning. Well, then you give up

01:02:37the accuracy. On an accountability, I think lawyers cannot assume that they're just going to be able

01:02:43to point their fingers at the developer of the tool and say, you know, I use this HR tool and

01:02:50I counted on the fact it wasn't discriminatory. I don't think that's going to work. I think you

01:02:55have to do your due diligence and you have to understand enough and put some, you know,

01:03:01check the results yourself and somebody will be held accountable and lawyers will be held

01:03:09accountable the same way they're held accountable for the eDiscovery vendors and photocopy vendors

01:03:14and everybody else's. So I would just quickly add, and I apologize, I'm going to have to run

01:03:21very soon here, but I think what I would add here is that everybody's raising really

01:03:28good points and what's underlying all of it is that this actually is going to require

01:03:33more cognitive work from humans, right? Like we think about these tools as being kind of

01:03:37efficiency enhancing and sort of time-saving and they will be properly deployed, but while we're

01:03:43getting to that point and as we're learning, and I think that's going to be for a very long time

01:03:47because there's always going to be something new on the horizon, we have to get used to

01:03:52sort of applying more of our critical thinking skills, more of our restraint and discretion

01:03:58as we encounter these new sort of capabilities that are technology enhanced or technology

01:04:07generated and that that's, the lawyers are going to have a huge role in that, right? Not just

01:04:11because we're the department of no, we're not, we're the department of how and we need to kind

01:04:18of switch our, we need to be prepared for that and we need to prepare our clients for that, right?

01:04:24And that it is a virtue, not a deficit that we have to be thinking harder about these systems and

01:04:38what good looks like, right? And we have to be a lot more prepared to have those conversations.

01:04:43And that's why I think the last three slides really just speak to the principle that as lawyers,

01:04:50your IP rights, let's say you use your clients using some sort of sophisticated video ad

01:04:55generation AI tool, like will it be a defense that the video AI put in video that they didn't

01:05:01have copyright authority to use and they violated copyright? I don't think the copyright on the

01:05:07other side of the beat cares. They're going to sue everyone and I think it'll be a function

01:05:11of contract. And with regards to the other points of legal ethics and professional misconduct

01:05:20common sense, nobody likes being lied to, you know, who hates it more than anybody else,

01:05:24a sitting judge who gets a brief with made up case law that then ends up wasting time

01:05:29of the court, making everybody look silly and frustrating everyone. So if you wanted to know

01:05:35where lawyers at a personal level have to be alone, like they have to make sure that what

01:05:40they're saying is truthful and honest. And the court's going to hold you accountable for

01:05:43your submitted words. I know we hit up on the hour here, so just quick 15, 20 seconds.

01:05:48I know, Karen, you've got to go. So safe travels and thank you for your time. Jenny or Moira,

01:05:53if there's any quick last comments you want to wrap up with, please feel free and then we'll call

01:05:59it a wrap. Jenny, Moira? Sure. I mean, I really liked what Karen said in those last few minutes,

01:06:07which was we're not the department of no, we're the department of how do we make this work?

01:06:13And the department of saying, okay, what are the risks? Can we get comfortable with the risks?

01:06:18How do we eliminate the risks? And I think that really is up to every lawyer to understand

01:06:25what this technology is, because I really think that in almost any situation, our clients are going

01:06:31to be interested in using this new technology for some purpose. So it definitely behooves us all to

01:06:40understand how it relates to our practice and how our clients are going to be using it.

01:06:44Yeah. And I think you have to play with this stuff. You have to get familiar with it because

01:06:51if you don't, you're going to be left behind. But you also have to be very careful that you

01:06:58you don't fall into the trap that some of these people have using it for research and not

01:07:03understanding that that generative AI hallucinates can make up answers. And but that doesn't mean

01:07:10run away and don't look at it. Use it for what it's good at, creative stuff, summarizing,

01:07:18and so forth. But get up to speed and understand what questions you need to ask and where the

01:07:26minefields are. Yeah, I think there's excellent points. And I think there are a lot of ethical

01:07:30issues that as lawyers advising clients, you need to be aware of and mindful of because these are

01:07:38it's only going to get more complicated before it gets any bit easier. So I thank you guys so much

01:07:42for your time. Again, if you do have any questions, please feel free to use the contact details that

01:07:48are found here. Any questions, they're more than qualified and I'm sure we welcome the conversation

01:07:55and dialogue. So thank you all very much for your time and I wish everybody a good rest of their

01:08:00day. Take care. Thanks very much, Jim. Thank you.

Automatically generated; it may contain errors.

Lesson notes

Bias, reliability, privacy, IP, and attorneys' ethical obligations.

Study materials

Notes, flashcards and practice questions for this lesson, reviewed before publication.

Study notes

Locating bias in an AI system across its life cycle

AI ethics covers a long list of concerns: fairness, inclusivity, transparency, accountability, reliability, safety, privacy, security, intellectual property and legal ethics. Fairness and bias come first because they arise everywhere, from dividing dessert to hiring decisions.

Bias is not only a property of the model. Look at the whole system: the humans, the data and the technology.

  • The objective. Was the business objective fair, reasonable and appropriate to the task, and was it articulated without introducing new bias? How a question is framed strongly affects the outputs.
  • The data. If training data are not representative of the data the system is used on, results will be biased. Facial recognition systems trained on too few people of color work well on white male faces and poorly on Black women's faces. An image generator trained on the internet, asked for pictures of a lawyer, may return only white men, and for a felon, only Black men, because the training data contain that bias. Employment screening tools trained on historical hiring may favor the kinds of candidates hired in the past.
  • The algorithm and its designers. Developers can make choices that build in bias. Using last year's health care costs to predict health care needs produces a biased algorithm, because people in underserved neighborhoods use services less than they need them, whether from travel distance, inability to get time off or lack of childcare. Proxy variables pull protected traits in indirectly: a model may never ask about race but use ZIP code, which is highly correlated with race.
  • The person interpreting the output. Some people over-trust algorithms ("death by GPS," following directions off a cliff). Others suffer from algorithm aversion: many lawyers distrust technology-assisted review even though the science shows the algorithms find responsive documents better than lawyers do.

Rather than saying "the algorithm was biased," examine what preceded the algorithm and how its output was used.

Unlawful bias versus business fairness

There is broad agreement that if you cannot discriminate without AI, you cannot discriminate with it. But unlawful bias is a fairly narrow set of concepts tied to protected classes. A broader set of fairness questions affects business performance: a lender does not want an underwriting model keying on whether homes have red roofs.

Bias is a function of how a data point is used, not just the data point. ZIP code may contribute to biased outcomes in underwriting, yet it is essential for delivering a package. Regulators in financial services are examining whether AI credit decisions are explainable and based on appropriate data, and whether factors such as phone battery level or social media connections perpetuate unfair bias. That raises transparency questions: should consumers be told what data are used, and can they object? Data reflect society, and AI applied at scale compounds and amplifies what is in them. Clients need to evaluate that risk and its downside.

Validity, reliability, privacy and security

Two concepts are often confused:

  • Validity: does the system measure or predict what it is supposed to? A good scale is not a valid measure of height.
  • Reliability: does it measure consistently over time under similar circumstances? A person stepping on a scale three times should see roughly the same weight.

An AI system must be both before it can be trusted. Reliability also depends on the human process around the tool: a reliable algorithm inserted into an unreliable process will perform unreliably. Models that use reinforcement learning, where outputs feed back into training, tend to drift from their original purpose and need monitoring that a locked model may not.

On privacy and security:

  • Do not put client confidential information into public generative AI tools; it may be given away or used for further training.
  • Generative systems are vulnerable to adversarial attacks. In one reported case, asking ChatGPT to repeat the word "poem" indefinitely eventually caused it to output personal information such as phone numbers and addresses.
  • Context matters. Lawyers must worry not only about client names but about disclosing strategy a tool could tie to a client. Financial institutions must protect their own trades from informing other models, and maintain internal walls such as those between investment banking and research, just as law firms use ethics walls.

AI is most dependable when it automates a well-understood process where we know what good looks like and the tolerances for variation. It is useful for brainstorming when we do not. Governance should be tailored to the use case.

Inclusiveness, transparency and accountability

  • Inclusiveness. Development teams should be diverse and involve the people who will use the system. Crash test dummies designed by male teams were modeled on male bodies; women's and pregnant bodies came many years later.
  • Transparency. The question is what level of transparency, and for whom. A computer scientist and an end user need different explanations, and someone denied a loan wants more than someone approved. Transparency matters less once validity and reliability are established, much as people take FDA-validated drugs without knowing their mechanism. There can be a trade-off: a simpler, more explainable system may be less accurate than deep learning. The debate distinguishes explainability of outcomes from explainability of process.
  • Accountability. You cannot blame the AI. A lender whose approvals show a discriminatory pattern is accountable whether the cause was an algorithm or a biased loan officer. Users are responsible for evaluating their tools and for outcomes; relying on an HR vendor's assurance that a tool is not discriminatory will not be enough. Lawyers will be held accountable as they are for eDiscovery and other vendors.

An open contract question is whether AI providers will accept liability through representations and warranties. Providers of narrow, use-specific services may; for black-box general tools, customers may argue they cannot determine regulatory compliance themselves.

Intellectual property, candor to the court and the lawyer's role

If a client's AI video tool inserts material the client had no right to use, the copyright owner is unlikely to accept "the AI did it," and allocation of that risk will be a matter of contract. Courts hold lawyers responsible for submitted words, and judges strongly object to briefs citing made-up case law.

AI will demand more cognitive work from humans, not less, for a long time: critical thinking, restraint and discretion. Lawyers should be the "department of how," not the department of no: understand the technology, weigh and reduce the risks, get hands-on experience, use the tools for what they are good at such as summarizing and creative work, and remember that generative AI hallucinates.

Flashcards

14 cards. Test your recall, then flip to check.

Select the card, or press Space, to flip it. Left and right arrow keys move between cards.

1 of 14

Practice quiz

Practice — not graded, not the final exam

Question 1 of 7A hiring model never asks about race but heavily weights applicants' ZIP codes. What risk does this create?
Choose an answer to check it.