Cadmore media player playing video PS24 Closing Keynote: Reid Blackman
Video Player
Accessibility
Using forms mode, you can utilize keyboard shortcuts to make usability more efficient.
Space bar will play and pause the video. Pressing escape will close all displays and pressing it again will jump to the end of the player.
You can press A to open the Keyboard Shortcut Dialog which also provides an audio description of the player and all of its features.
Language
1x
Play Speed
Adjust Volume
Transcript
Search
Result Count
of
Click to jump to result
Search
Re-center
Segments
Result Count
of
Click to jump to result
Search
Name:
PS24 Closing Keynote: Reid Blackman
Description:
PS24 Closing Keynote: Reid Blackman
Thumbnail URL:
https://cadmoremediastorage.blob.core.windows.net/5bef3d0f-a209-495c-b6e4-f1f80c0f41a8/thumbnails/5bef3d0f-a209-495c-b6e4-f1f80c0f41a8.png
Duration:
T00H51M50S
Embed URL:
https://stream.cadmore.media/player/5bef3d0f-a209-495c-b6e4-f1f80c0f41a8
Content URL:
https://cadmoreoriginalmedia.blob.core.windows.net/5bef3d0f-a209-495c-b6e4-f1f80c0f41a8/PS24 Closing Keynote.mov?sv=2019-02-02&sr=c&sig=NmYhq9NH4c5hVHPfSqSesRqvkFFdQS4zPVnjmF%2BUb%2FE%3D&st=2025-01-22T04%3A36%3A22Z&se=2025-01-22T06%3A41%3A22Z&sp=r
Upload Date:
2025-01-22T04:41:22.3171977Z
Transcript:
Language: EN.
Segment:0 .
NIALL LITTLE: Hi, everyone. We're going to get ready for our keynote speaker here. My name is Neil Little. I am the CTO and technical co-founder at Hum. And I'm also a proud Silverchair alumni. This has always been one of my favorite conferences. So I'd like to thank Will and Stephanie and their entire team for putting on such a great program today. Let's give them a round of applause. [APPLAUSE] As an AI company, at Hum, and stewards of your data, ethical AI is a topic we take very seriously, which is why I'm super excited to be introducing our closing keynote speaker today.
NIALL LITTLE: As artificial intelligence reshapes the landscape of digital experiences, it's also crucial that tech providers and organizations consider the ethical implications of these powerful technologies. Dr. Reid Blackman is an accomplished author, thought leader and an expert in AI ethics and digital risk management. As the founder and CEO of Virtue, a digital ethical risk consultancy, Dr. Blackman has been at the forefront of helping Fortune 500 and Global 1000 companies scale AI and digital ethical risk management programs.
NIALL LITTLE: His expertise has been sought after by industry giants, and he has served as a senior advisor to the Deloitte AI Institute and a founding member of the Ernst & Young's AI advisory board. He also shares his ethical AI insights in his book Ethical Machines, Your Concise Guide to Totally Unbiased, Transparent and Respectful AI. And his work has been featured in The Wall Street Journal, The Harvard Business Review and TechCrunch.
NIALL LITTLE: Please join me in welcoming Dr. Reid Blackman. [APPLAUSE]
REID BLACKMAN: Hey, everyone. How you doing? You've made it to the end of the day. Now you have to listen to a philosophy lecture. You draw the short straw there. OK. So let's see. What should I talk about. I guess I'll talk about AI ethics. This is not my slide, though.
REID BLACKMAN: Does this work? That's me, blah, blah. OK. So here's where I'm going to take you today. At least, this is the plan. First, I want to show you how I conceive of the AI ethics landscape. And I want to talk about, of course, AI ethical risks. And then I want to talk about how you might think about those risks in the context of scholarly values, scholarly publication values.
REID BLACKMAN: And then, finally, I'm going to talk a little bit about, pragmatically, what are my suggestions for what you ought to do, going forward. So that's the plan here. First thing, I want to talk about what AI is. AI is just software that learns by example. People talk. It's supposed to be this really fancy, complicate-- it's software that learns by example.
REID BLACKMAN: That's it. So when you hear me say, AI, or you hear anyone say, AI, for that matter, just think, "software that learns by example." That's all it is. You know what it is? You know what software is, right? Everything on your phone, every website software got it covered. You know what learning by example is.
REID BLACKMAN: I have a six-year-old. At some point, I told him, this is what pigs look like. And now he can spot pigs in new books. He's very talented. To take a more complicated version, I could give you a bunch of impressionist paintings. At some point, you've probably seen a bunch of impressionist paintings.
REID BLACKMAN: And now you've learned what impressionist paintings look like. So you go to the National Art Gallery, and you see some impressionist paintings. And you go, oh, yeah, that's an impressionist painting. I know that from all the examples that I saw. That's all AI is doing. It's software that learns by example.
REID BLACKMAN: If you want your AI to learn what Pepe, your dog, looks like, give it a bunch of examples of what your dog looks like, photos. It'll, quote, unquote, "learn" what Pepe looks like. And when you snap a new photo of Pepe, it'll put it in the Pepe folder. Because it recognizes, oh, yeah, that's Pepe. I know that guy. I know this is not your area, but if you wanted to approve or deny various mortgage applications, give it a bunch of examples of, oh, here are all the good mortgage applications that we approved in the past.
REID BLACKMAN: Here are all the ones that we denied. And it's going to learn, from those examples, what the good mortgage applications look like, what the bad ones look like, so on and so forth. Of course, there are complications. It messes up, blah, blah, blah. But at the core, it's just software that learns by example. If you want your AI to output text in a way that's coherent, then give it a bunch of examples of words cobbled together coherently.
REID BLACKMAN: And it'll learn, by those examples, how to do exactly that. That's all large language models are, LLMs. They're just cobbling words together based on all the other examples it has of words being cobbled together. That's it. So I don't care if you're talking about generative AI or large language models or you're talking about so-called narrow AI.
REID BLACKMAN: It's all software that learns by example. If you want a fancy word, for example, you can use the word data. That's the fancy word, for example. That's it. So instead of saying, AI is software that learns by example, you can say, AI is software that learns by data. It's a little bit more high brow than saying, learns by example. And if you want an even more highbrow, really fancy, call it "training data." AI is software that learns by training data.
REID BLACKMAN: You give training data to the software, and "training data" is just another fancy word for "example." And I think it's important that we just recognize that we're just talking about software that learns by example because data scientists, computer scientists, people who are trying to sell AI usually talk in this opaque way that intimidates us. I'm not a technologist by background, right? I said I'm a-- by background, I'm a philosopher.
REID BLACKMAN: An odd thing to be, but that's what I am. And you don't need to let people say things like, training data, convolutional neural networks. Software that learns by example, relax yourself. So let's make sure that they don't push us around with all their fancy words. OK. Once we've got that-- now, here's the nice thing about learning that AI is just software that learns by example.
REID BLACKMAN: It allows you to see the AI ethics landscape really easily. So when I think about AI ethics, I distinguish between what I call "AI for good," on the one hand, and "AI for not bad," on the other. The people in the "AI for good camp," they ask questions like, how can we do a bunch of good in the world using the powerful tool that is AI? So for those of you who are familiar with the United Nations Sustainability Development Goals, the SDGs, lowering poverty worldwide, increasing education, increasing women's rights, et cetera, et cetera, you might ask, OK, how can we use the powerful tool that is AI to pursue those SDGs?
REID BLACKMAN: So it's, if you like, wearing the ethics on its sleeve. The objective, the goal is to do something ethical using the powerful tool that is AI. Great stuff, noble people, people should do that. I'm not. Not what I'm going to talk about, either. I focus on "AI for not bad." And as you can imagine, that means you're going to do the things you're going to do.
REID BLACKMAN: Let's assume they're not ethically evil. Let's assume they're either ethically neutral or ethically good. And then the question is, how can we pursue those normal business objectives, ethically acceptable, but not particularly ethically amazing goals? How can we do that without ethically messing things up with AI?
REID BLACKMAN: How do we not go sideways on this? So that's "AI for not bad." You can call it "AI risk mitigation" or "AI ethical risk mitigation" or "AI ethical management." You can call it "AI governance." You can call it "responsible AI." All these words are floating around in this space, it doesn't really matter. I think "AI for not bad" is the one we should go with.
REID BLACKMAN: Nobody else listens to me, but that's what I think we should do. Once we got "AI for not bad" in our heads and what we're doing is we're just-- we want to identify and manage the ethical, reputational, regulatory, legal risks of AI. There's another distinction that I like to make between content and structure, content and structure. So the content question asks, what are the ethical risks of AI?
REID BLACKMAN: And what are their sources? Why are we even talking about-- what's the sources of these ethical risks? Why is it such a topic of conversation? Why am I up here right now? Why are you listening to me? That's the content side. The structure side is in, what are we going to do about it? As an organization, what are we supposed to do about that sort of thing?
REID BLACKMAN: For the most part, people look at the content stuff very superficially. You've probably heard about biased AI or black box AI or privacy violations. People sometimes freak out, and they say, OK, well, what are we going to do about it? We got to do something? So they treat the content side very superficially, sort of at the headline level.
REID BLACKMAN: And then they jump over to the structure side and ask, what should we do? And then they say, I don't know what to do. Usually, they come up with some piece of paper with values on it, like, we're for justice and privacy and respect for all. And it doesn't mean anything because they have no idea how to actually put that stuff into practice. So people don't really do much.
REID BLACKMAN: And one of the things that I contend is that we would do a lot better, on the structure side, knowing how to manage those risks, if we actually went deeper on the content side. And in one way, what I'm saying is utterly mundane because all I'm saying is, if you don't really understand the nature and the source of your problem, you're not going to be able to develop a good solution for it. That's all I'm really saying.
REID BLACKMAN: I'm very unimpressive. [LAUGHTER] So I want to dig deeper on the content side. You'll get a better sense of where these ethical risks are coming from. And then we can talk about what can you do about it. OK, content side, blah. There are three main ones. And I don't care if you're talking about generative AI or more narrow AI.
REID BLACKMAN: Can you see that? It's a little bit low. Privacy, black box and bias, so let's take these one at a time, more or less. So let's go back to Pepe. You're taking pictures of your dog, Pepe. Now, when you look at Pepe, you're looking at two eyes, a nose, a mouth, ears. And so you learn what Pepe looks like.
REID BLACKMAN: And then you can say, oh, yeah, that's Pepe, right? AI doesn't do that. Software that learns by example doesn't do that. It's not looking at ears and eyes and a nose and a mouth. It's looking at pixels. It's analyzing these pictures on the pixel level. And it's looking at the thousands of pixels and the thousands of mathematical relations among those pixels.
REID BLACKMAN: It's really complicated. So when it says, "Yeah, yeah, that's Pepe" or, "No, it's not Pepe," it's because it's identifying a phenomenally complex mathematical pattern in that photo and saying, "Yeah, that really complex mathematical pattern fits the same mathematical pattern that was in all those previous photos of Pepe" or No, it doesn't." Let's just say this mathematical pattern is 50 pages long, right, just 50 pages of equation.
REID BLACKMAN: You and I can't keep that. At least, I mean, I certainly can't. I'm betting that most of you can. Keep that in your head. Right, it's too long. You speak the language, but the sentence is way too long for you to follow. And so when it says, "Yes, this is Pepe," "No, it's not," we can't actually understand why it's saying "yes" or "no." We know it's based on some mathematical pattern, but the pattern itself is one that we cannot comprehend because our minds are so finite.
REID BLACKMAN: So why does it say, "Yes, Pepe" or "No, Pepe"? We're not really sure. It's something about this mathematical pattern. But we can't really explain to you how it results in or how it outputted, "Yes, that's Pepe," "No, it's not." Now, we're talking about photos of Pepe. Who cares? But if you're talking about mortgage applications, job applications, applications to a university, sentencing guidelines, sentencing judgments for criminal defendants, now you start thinking, maybe we need to know why this thing is saying "yes" or "no" or "high risk," "low risk" or something along those lines.
REID BLACKMAN: So that is, in essence, the black box problem. And so some people think-- and we can talk about this in Q&A if you'd like-- that we really need to solve the black box problem if we're ultimately going to trust and use AI safely and responsibly. I'm not persuaded of that totally. Again, we can talk about that. But that is at least a concern on people's mind, the black box problem.
REID BLACKMAN: OK. Now, all those examples that the software is learning by, all that data, to use the fancy word, is often not always, but often about people. Which means that, in order to get this thing working well-- because, all else equal, the more examples you give it to learn by, the better it learns, the better it does. The more data that you have, the better your AI is going to be. Which means that the data scientists who want to train their AI to make it as awesome as possible are incentivized to hoover up as much data as they can because that is the most obvious pathway to getting more awesome AI.
REID BLACKMAN: But if that data is often about people or, let's say, it's the [? IP ?] of people, then we have privacy violations on the horizon, right? Because the fuel of machine learning or AI-- I don't know if I said, machine learning, yet. AI, machine learning, software that learns by example, treat it all the same. The fuel of it is data. And if they want to go far, they need to get a lot of fuel.
REID BLACKMAN: So that's one way. There's more, but that's one way in which we risk privacy violations, by virtue of the nature of the beast. Now, actually, I want to say something about the nature of the beast. Notice that, so far, both the black box issue and the privacy issue are not incidental to how AI works. It's not like, oh, yeah, we happened to use the screwdriver to kill someone.
REID BLACKMAN: No, no, it's not coincidental to the tool. It's built into the tool that these kinds of things are not merely likely. Sorry. They're not merely possible. They are likely. They're not merely possible. They are probable, just given the way the technology works. Let's talk about the bias issue now, which, probably, a lot of you are familiar with.
REID BLACKMAN: You give it a bunch of examples. If the examples are, in some way or other, skewed, you're going to wind up with skewed or biased and potentially ethically problematic results. So for example, if you give it a bunch of, let's say, medical data that predicts whether someone's going to develop diabetes in the next two years and you only give it data related to-- a very small percent of all that data is, let's say, about Black men.
REID BLACKMAN: You haven't given enough examples for it to learn the probability that Black men will develop diabetes in the next two years, given their medical profile. To take a simpler example, a famous or infamous one at this point, when they first started developing facial recognition technology, it was really bad at recognizing Black women especially, Black people, darker-skinned people, especially Black women, for the simple reason that they just did not have a tremendous amount of photos of Black women in various lighting conditions, taken from different angles.
REID BLACKMAN: And so it just didn't learn very well. So what you wound up with was facial recognition software that was good at recognizing white men in photos, but bad at recognizing Black women. So that's the bias problem. There's, of course, more to it. This is 30,000-foot stuff, but you should get a sense of where this is coming from. Again, this is not incidental.
REID BLACKMAN: This is built into the technology itself. And another thing to note is that we're not talking about bad behavior, right? I'm talking about ethics, but I'm not talking about the bad behavior of people. I'm talking about things that often accidentally, even in some cases, blamelessly-- although, there's more to say on that-- get introduced into the AI without anyone intending to do bad things.
REID BLACKMAN: So with the screwdriver, if you're going to do something really evil with a screwdriver, you really got to try hard, right? You got to intend to do something bad. But you can create privacy-violating AI that's a black box, that systematically discriminates against women without any intention to do so. This is one of the reasons why we're talking about this stuff. Again, it's not merely possible.
REID BLACKMAN: It's probable, given the nature of the beast. You can also have use-case-specific risks, in addition to these. So this is just built into the technology. They can be mitigated. They can be managed. But it's built into the technology, as opposed to, say, using AI to power or create a self-driving car.
REID BLACKMAN: We're now maiming and killing pedestrians, is an ethical risk. That's a high priority. That's not because of the nature of the beast that is AI, that is software that learns by example. It's because the purpose is to which you put the beast. OK. Everything I said, so far, applies to generative AI, large language models, just as much as to narrow AI.
REID BLACKMAN: Is this person a good candidate for the job? Does this person deserve a mortgage, whatever it is. When we come to large language models in particular, we get two other kinds of ethical risks that are special. One, I'm sure that you've heard this phrase, hallucinations. You get sentences that are just false. You get things that are just false. So rather famously, infamously, there was a lawyer who went to the courts citing various court cases in his defense.
REID BLACKMAN: But the judge looked into it, and it turns out those court cases weren't real. And the reason was that the lawyer went to ChatGPT. He said, hey, I'm looking for some cases that do this kind of thing, that demonstrate this kind of thing. And it spat out some things. The lawyer didn't have the wherewithal to check or to verify those things, presented it in court. They were completely made up, completely fictional.
REID BLACKMAN: And he got in a lot of trouble, a lot of trouble. So why does it do that? Why does it hallucinate? In short, remember what AI is. It's software that learns by example. When it learned to create outputs of a textual variety, how is it doing it? It's just cobbling words together coherently, based on the examples of other things it's seen, other words it's seen cobbled together coherently.
REID BLACKMAN: It's not doing it with an eye towards the truth. It doesn't know the truth. It's not tracking the world. That's not what it's in the business of doing. It's in the business of putting words together coherently. And so since something can be coherent, but plainly false, you get hallucinations. It's built into it. There's various kinds of attempted solutions to this thing.
REID BLACKMAN: Some of you have heard of RAG, for instance. Nothing is settled, yet. No one has solved the hallucination problem. People might say, oh, we solved it. We have a start up. We're looking for funding. We solved it. OK, we'll see. So the hallucination problem is one problem.
REID BLACKMAN: And then there's something that I call the "deliberation problem," which is a slightly more subtle problem. It's super fascinating. It scares me a bit because people don't get this one, by and large. The deliberation problem is this. You might think that the large language model at ChatGPT or Anthropic's Claude or whatever it is is giving you an explanation for why it told you what it did.
REID BLACKMAN: Why did you tell me that this is the court case? Why did you tell me that this is the appropriate diagnosis? Why did you-- whatever it is. And it'll come up with something. And it might even appear reasonable. But the crucial thing is that, whatever it outputs, it's not telling the truth. And it's not telling the truth because it doesn't give outputs on the basis of evidence.
REID BLACKMAN: It doesn't give it on the basis of reasons. It's not a reasoning machine. It is a next-word-prediction machine. And so when it says, the reason that I told you that this person probably has cancer is that, blah, blah, blah, blah, blah, that blah, blah, blah, blah, blah, blah, blah has nothing to do with why it told you that person is going to develop cancer.
REID BLACKMAN: Because it's just not that kind of thing. It's not reasoning, in the first place. So what is it doing when it gives you that explanation? It's just doing its normal thing of putting together words, coherently, that make sense with what came before, even though it bears no actual relation to the stuff that came before. So the deliberation problem is a problem of, in some sense-- I'm not sure that I want to use this word.
REID BLACKMAN: I'm going to use it anyway, deception, because it's very easy to get fooled into thinking that you're dealing with a reasoning machine. Because you can ask it for recommendations and then reasons why it gave those recommendations, and it will say things that look like evidence or justification for why it said what it did. But it's making it up.
REID BLACKMAN: It's just cobbling words together that it thinks you're going to be like, oh, yeah, that makes sense. So that's the deliberation problem. OK. So now we're beginning to see the landscape a bit more, right? We've got AI ethics, more generally. You want to do some great things. That's "AI for good." You want to not ethically screw up pursuing the ordinary business objectives that you have.
REID BLACKMAN: That's "AI for not bad." Then we distinguish between the content and structure. Where are these risks coming from? Talked about that. It's because it's software that learns by example, making these things not merely possible, but probable. And then we've seen five main ones. There are others.
REID BLACKMAN: I just didn't include them here because one can only say so much in this time. But there are others. These are the biggest ones, I think. So now let's talk about the structure stuff. Now that we have a grip on what the risks are, where they come from, why we're talking about them, what are we supposed to do about this? Here are two suggestions.
REID BLACKMAN: And this is tailored for this audience. This is not what I would say if I were talking to chief information officers that have 100,000 employees. It's a different beast. That's a much bigger lift. But for smaller organizations that are trying to figure out where does AI fit here, you need something like an AI ethics board or an AI ethics committee or an ethical risk committee or a brand risk committee or whatever you want to call it.
REID BLACKMAN: I don't care. I'm going to call it an AI ethics board. You need that, and then you need some tool for risk assessment. So I'm going to run through these kinds of things, about what they should, more or less, look like. Because I think that, all else equal-- I don't know your position, of course. But you might want to go back and create this kind of thing as you think about integrating AI into your organization.
REID BLACKMAN: So the ethics board, it should be cross-functional. It can't just be data scientists. You're not going to math your way out of these problems. You need a cross-functional, cross-departmental ethics board who can weigh in, in various ways, about whether this thing is ethically safe, reputationally safe, et cetera. Legally safe, you'll want a lawyer, regulatorily safe. You'll want to include external stakeholders, especially, for instance, authors or researchers, since they're going to have lots to say about what's getting done with their work.
REID BLACKMAN: And this board has to have the authority to issue go or no-go decisions. They have to be able to say, mm, this is just way too risky for us, just no. That's the advice anyway, that they should have that authority to issue go, no-go decisions, as opposed to mere advice. But maybe that advice gets downplayed when there's a lot of money on the line. And you think, thank you, ethics board, but you're being overly conservative.
REID BLACKMAN: No, thank you. There's many millions to be made here, et cetera, et cetera. That's when we see ethical breaches. But if it has the authority to say, sorry, red light, then it should be a really, really big deal if that board is going to get overruled, as opposed to, they didn't follow the advice. That's a bit about the ethics board. And then we have to talk about what's a risk assessment going to look like.
REID BLACKMAN: Now, this has to get customized to your particular organization. So I've just taken some values that I cobbled together-- I guess I'm like an LLM-- from the Society for Scholarly Publishing and the Association of University Presses. Stewardship is-- who is that? I think that's the Association of University Presses. Community is Scholarly Publishing. Integrity is on both people's lists-- both organizations' lists.
REID BLACKMAN: Intellectual freedom and adaptability and equity and inclusion-- They can be different. I'm just using these as examples. If your organization has different values that are close, but not identical, use the ones you've got. OK. So here's how it goes, roughly.
REID BLACKMAN: This is not exhaustive, but if your ethics board is convened and you've got a solution that you're thinking about, this is a rough-and-ready tool for guiding that discussion. So you can ask things like, do we have buy-in from authors for the solution? Do we know what happens to the data that a user inputs into the AI? Do I know what happens to the AI outputs? Can we clearly communicate the risks or best practices for use of this AI?
REID BLACKMAN: Did we provide readers or clients with the appropriate resources for using the solution responsibly? Again, not exhaustive, but instructive, I hope. Gives the ethics board something to do, something to talk about. Are readers, clients able to verify the outputs? Have we considered whether using a black box AI in this context is inappropriate? By the way, all large language models, ChatGPT, Anthropic, Gemini, et cetera, they're all black boxes.
REID BLACKMAN: [LAUGHS] They have kind of a clue, but it's more of a suspicion. They understand how 1% of it works. They don't know. The outputs are all unexplainable, which is understandable when you think about the vast quantity of data that these things are trained on. Trying to figure out and understand the mathematical patterns that underlie the predictions, we're not going to do it.
REID BLACKMAN: Do we have the expertise to assess whether the AI is fit to the task or it does a good enough job? I'm stressing-- I put "enough" in parentheses. It's not going to do a perfect job. If your goal is perfection, forget it. Don't use AI. But one thing to always keep in mind is that there are three kinds of systems that we're actually talking about.
REID BLACKMAN: There's the human alone, making decisions, recommendations or doing the work. There's AI alone, making decisions, doing the work, whatever it is. And then there's AI plus human. So those three different systems, if you like, are accurate or inaccurate to varying degrees. And one question is going to be, when you ask, "Is this sufficiently safe to deploy," is, "Well, what are you comparing it against, human alone, human plus AI, AI alone, et cetera?" OK.
REID BLACKMAN: Where are we now? Is there a way this AI can operate to inappropriately stifle legitimate views? So think about the bias issue, right? So suppose that you're trying to unearth summaries of texts or something along those lines. And you have very little text or very little content coming from, say, a certain region of the world. You're probably going to get not that view bubbling up to the top, not in the AI's outputs, because it's such a small part of its training data.
REID BLACKMAN: Is the AI designed in a way that encourages inappropriate deference to the AI outputs? Deference to the AI outputs is a big thing. You may have heard stories about people literally driving into lakes because Google Maps told them to. This doesn't feel right, but I don't know. The computer says so. Woop. [LAUGHTER] And that's really like, what are you doing?
REID BLACKMAN: I didn't do this. I almost did this recently. But I did turn into a harbor when I knew that I couldn't possibly-- you're right. [LAUGHTER] But imagine if it's a lot lower stakes, and the lake is not literally in front of your face. And it says, yeah, you should go do X. Or, yeah, the report says Y. Yeah, OK.
REID BLACKMAN: Well, sure yeah. Easy to defer, very easy, especially to computers. But we ought not to. Do I know whether or how this AI solution may be biased or discriminatory, which stakeholders may be at risk, and why? So this is just a summary. I'll actually share this with you. I didn't talk about this with Silverchair.
REID BLACKMAN: But I'll find a way of distributing this slide to you, so you can all use it. It's yours. So you can take a picture if you'd like, but I can also just send you a version. So let's look at a quick and dirty example. Here's the opportunity. Notice, by the way, I use the word "opportunity." I'm not a skeptic.
REID BLACKMAN: Maybe. I'm a skeptic. I'm not a doomsday person. I don't think, oh, this stuff is terrible, and it's never going to-- I don't know. It could work. Sometimes I've seen it work. It's amazing technology.
REID BLACKMAN: So don't take anything I say to indicate, forget the whole thing. It's never any good. I don't think that at all. I think there's a tremendous amount of opportunity. I also just happen to think that we should probably watch out for the risks as well. So let's take this opportunity and run through an example of what would the questions look like.
REID BLACKMAN: AI can read articles and books, including charts, infographics, et cetera and create audience-appropriate summaries, analysis and recommendations for various stakeholders. So you can imagine a doctor wanting to get recommendations on how to treat this patient, or a nurse or somewhere in the healthcare system wanting guidance on this sort of thing, or a lawyer wanting guidance on "What kind of cases should I be looking at," so on and so forth. So that's an opportunity.
REID BLACKMAN: And given the insane mountain of data, the insane mountain of text that everyone in this room is stewards of, we could come up with some amazing things. So there's a great opportunity. But as we already made our way towards, are the readers, clients able to verify the AI outputs? So you've got all this text, and then it summarizes. It says, well, essentially, this documents or these 10,000 pages, the upshot is X. And the major recommendation is Y.
REID BLACKMAN: Can you verify those outputs? Yeah, in principle, they should be able to verify it. They better not be too differential, but in principle, they can verify it. That's not to say they necessarily will, by the way, because they might not have the time or the bandwidth. Or they have due dates. Or they have a boss that says, just get it done. But in principle, they should be able to verify it.
REID BLACKMAN: It also might be the case, though, that it's going to take too much time. It's 10,000 documents. I don't have the time to read those 10,000 documents and then verify. That was the whole point of using the AI. So in principle, they can verify it. But can they really verify it? Can they verify it enough?
REID BLACKMAN: And remember, can they verify it in a way that results in an overall outcome that is better than the alternative? Not is it perfect, not perfect, but better than a human poring through those 10,000 documents, yada, yada, yada. Because the human's going to miss stuff, for sure. They're also going to include stuff that they shouldn't have included in the summary. Same with an AI.
REID BLACKMAN: Have you considered whether using a black box is inappropriate? You may decide that. So long as you can independently verify the outputs, the fact that it's a black box doesn't matter. So you might just care about reliability. The black box issue is an issue. It's something that deserves thought. Attention, but it's not a nail in the coffin of AI, of LLMs.
REID BLACKMAN: It just means, hey, is this use case-- is it OK for us to use a black box? You might think, yeah, I don't really care how it gets the outputs, as long as we can show that it's efficiently, reliably giving out the right answers. That's a view. Do we have the expertise to determine whether the AI is good enough?
REID BLACKMAN: And this is going to potentially vary by context. Third-party subject matter experts may be needed, so you might need to bring in other people, externally, to vet this thing. So you can see you're starting to get some sense of can we use this thing. I think, right now, so far, the biggest worry is about can they verify it. But what's the context?
REID BLACKMAN: How much time do they have to verify it? That's going to matter, to whether you give a yes or no decision. Let's do one more. Do we have buy-in from authors for this solution? And that's going to vary by context. It's going to require stakeholder interviews, focus groups, et cetera. It's worth talking to them.
REID BLACKMAN: And so this is, again, not a no. This is not a red flag, stop all things. Oh, you know what? Right, we should talk to some representative authors. Not all of them, but we should form some focus groups and see what our community thinks. Do we know what happens to the inputs or outputs of the data? And this might vary by provider. And so you might have heard things like, is the AI in the cloud?
REID BLACKMAN: Is it on prem? Is it on premise? Is it on your property, so that no one has access to it? Are you sending it to a third party, like an OpenAI or a Microsoft or a Google? And if so, what do they- do they see the inputs? Do they see the outputs? What are they doing with it? What are they allowed to do with it?
REID BLACKMAN: What are they not allowed to do with it? Again, red flag, but we can change it to green if we get the right kinds of answers. And can we clearly communicate the risks or best practices for use of this AI? Probably. [LAUGHS] In principle, yes, you can say the things. The question is, of course, whether you can get the people who need to hear the things actually hear them.
REID BLACKMAN: Can we get them in a room? Can we give them the right instruction? Can they know what they can and cannot do? Do they have the bandwidth to learn the things that we want to communicate to them? Do they have the workflows, the processes in place, so they can actually incorporate those learnings into how they actually use the AI? Things are getting a lot more complicated now.
REID BLACKMAN: But at least this should give you-- begin to give you, at least, some grip on the risks. They're not, in all cases, insurmountable. They're surmountable, on the condition that you give it the right kind of thought, which means that the bottom line is whether using a particular AI is commensurate with your values. It's not a black and white issue. And it requires your due diligence and professional judgment.
REID BLACKMAN: That's the point of the board. The board should be exercising its professional judgment. And you can see those series of questions as a way to assist the board in seeing what kinds of questions it needs to ask. So in conclusion, wrapping up, I think you should create-- probably create, if you're seriously considering integrating AI into your organization, an AI ethics board.
REID BLACKMAN: Number two, create your AI ethics assessment. It could be the exact one that I gave. It could be more tailored to your organization. That would be preferable. But create one. It doesn't have to be super complicated, as you can see. And finally, don't let the technologists push you around. It's just software that learns by example. Thank you.
REID BLACKMAN: [APPLAUSE]
SPEAKER: [INAUDIBLE] for questions.
SPEAKER: All right. And we do have time for questions, great. We have 16 minutes, yeah, minutes, good.
SPEAKER: Hi. Thanks for coming today. How often have you discovered entities who are deliberately skewing the data in order to get results that they want?
REID BLACKMAN: How often are they skewing the data to get the results they want? So if you don't use the word "skewing" and you mean how often are they tinkering or introducing new data, a lot. I mean, what data scientists are doing is they are introducing different data, different examples to get the AI to be more accurate. So that sort of thing happens all the time. And so this is one reason, incidentally, why, if I were talking to a group of data scientists, I would say things like, you need to check for these risks at every stage of the AI life cycle.
REID BLACKMAN: Some organizations will do it when they're just thinking about it. We're just trying to think about, what's the solution here? What are the risks here? Let's note those risks. And now let's start building and testing. And when they're testing, they're testing for accuracy. It's not accurate enough, so they introduce more data. And that's when privacy violations occur.
REID BLACKMAN: Bias occurs and that sort of thing. Does that answer your question?
SPEAKER: I think you're-- I think you're suggest-- sorry. It seems like you're suggesting that they're skewing the data to introduce more and more data that would provide accurate results, truthful results. I'm asking you have you discovered any entities who are doing the opposite.
REID BLACKMAN: Where they're decreasing the data?
SPEAKER: No, they're probably increasing-- it's possible-- the data to get certain results that they want beforehand, that are not truthful.
REID BLACKMAN: Ah, no. I can't think of any particular cases where I've seen such things. Usually, the problems are just, we didn't realize it was going to do that kind of thing. It's usually negligence of some kind, not-- the data scientists are in this little bubble. [LAUGHS] They just want the number of-- the accuracy percentage to go up.
REID BLACKMAN: That's all they want. They're not thinking about marketing and sales and bottom-line stuff and satisfying the client. They don't even know the client. They're tinkering with their computers. So they're not trying to get-- it's not politically motivated or something like that, to get a certain output, typically. I mean, I see you-- "smirk" is not quite the right word, but something next to a smirk.
REID BLACKMAN: So let me go on. [LAUGHTER] There are ways in which you can tinker with AI because you think it's inappropriate, right? So rather famously or infamously, depending upon your view, OpenAI, Google, they've all tinkered with the data because they don't think that the outputs are appropriate. They're not safe.
REID BLACKMAN: And so we need to restrict it. And then you are going to see things that you might think are politically influenced in the wrong way. So that's possible. And even, there are things that go sideways, even despite the best intentions. So relatively recently, Gemini was creating pictures of Black Nazis because they didn't have enough Black people in the training data generally.
REID BLACKMAN: So when you ask for an image of, say, a CEO, you always got a white CEO. And people were like, what's up with this? How come every time I search for a CEO or a titan of industry, I always get an elderly white man? And so Google tried to fix this. And then it turns out they did fix that problem, but they went too far. And so now we have pictures of Black Nazis.
REID BLACKMAN: Is that of some help?
SPEAKER: Yeah, we can talk about that [INAUDIBLE].
REID BLACKMAN: Sure, sounds good. [LAUGHTER] Not really? OK.
SPEAKER: Thanks for this talk. It was great. You're talking largely here about ethical use of AI within your own organization. The organizations here are-- you could say that they create data. Or you could say they create content or articles or what have you, human intervention. And how do you think the publishers and those who are creating the content should be looking at ethics, not from an internal use, but from a licensure, external-use perspective?
SPEAKER: Because while a lot of our content in our industry is used without our consent, increasingly there are requests for consent from LLMs and others. So how should we be thinking about the ethics in the outgoing direction?
REID BLACKMAN: That's a great question. So if you're talking about selling the data, so that other parties can use it to train their AI to blah, blah, blah, well, it's going to vary, the answer. First of all, what do your authors think about that. The people who generated all the-- who did all the writing, how do they feel about that? My guess is that they're going to say, you can do that, on the condition that I get compensated.
REID BLACKMAN: And then if they say, yeah, you could do that, on the condition I get compensated, then you have these negotiations. And then they say, yeah, sure. Then you're probably mostly fine, I would say. It's going to be, in some cases, out of your control, what ultimately gets done with it and how it's going to get used. Your responsibility gets attenuated the more the data gets messed with. And it's going to get messed with a lot.
REID BLACKMAN: And it's going to be combined with other people's data a lot. And so at the end of the day, if someone uses it to, let's say, make a recommender for doctors for how they should treat patients, my suspicion is that-- And I have to think more about this. And I have to see the actual case. But my suspicion is that you're so far removed from what the solution actually turned out to be. I don't see you as being ethically responsible.
REID BLACKMAN: Does that kind of case speak to your question?
SPEAKER: Yeah, that certainly answers the question.
REID BLACKMAN: Yeah, there's lots of parties who are going to be response-- because there's going to be better and worse ways to deal with that data. And it's going to be the people who use that data. it's going to be primarily their responsibility to make sure that they use it, that they tweak it, that they skew it and whatever they do with it, they do it in an ethically responsible way.
SPEAKER: Thank you.
REID BLACKMAN: Yep.
SPEAKER: I think I know the answer to this question, but I'm going to ask it anyway.
REID BLACKMAN: It makes my job easy.
SPEAKER: [LAUGHS] The example you used earlier, about the black box, I think you had said there's a 1% chance they know what's going on. And in essence, what happens is, when the results come out, I guess you do a gut check and say, does this look like these are valid results? And if so, theoretically, the process works. Going back a little bit to what I think this gentleman was alluding to, with regards to people perhaps wanting less ethical outcomes to be the results of AI, the fact that the black box issue stands, is there a way to delineate between something that is correct and on the mark versus something that is manufactured and maybe not at all on the mark?
SPEAKER: Does the black box issue make that impossible, then, to determine whether something is something that you can rely on and act on responsibly or something that you have to watch out for and maybe avoid? How does that get resolved?
REID BLACKMAN: So there's a few things to say here. So the first thing is that I wouldn't rely on your gut. Not your gut, in particular, anyone's gut.
SPEAKER: Yeah.
REID BLACKMAN: [LAUGHS] Because gut instincts are remarkably bad, including in ethical situations. I mean, I don't know. People thought being gay was disgusting, and that was what their gut told them. And so that was not an ethically great situation. So I wouldn't rely on people's guts or moral intuition or something to test it. You probably want to test it against certain benchmarks. So let me just take Pepe the dog, the photo.
REID BLACKMAN: We don't know how the thing works, but we don't do a gut check. We look to see how many mistakes does it make. And we gave it 100 examples of Pepe. Did it do a good job? It got 50 right and 50 wrong. That's not good. That's not my gut. That's just, this is really bad.
REID BLACKMAN: It got one wrong out of 10,000. All right. That's really good. So the question is, what's the appropriate benchmark that we're comparing it to, that we want to achieve here? Summary, we don't have those same kinds of quantitative metrics for measuring whether the summary was good. Unfortunately, there, the only way to verify is to actually go do the work.
REID BLACKMAN: There's different ways of doing this. It doesn't mean that you have to test every single output. So there's different ways of testing. So you can say, all right, we're going to have it summarize these 50 documents. And we're going to do a random audit on three of them or five of them and see if it did a good job. And if it did a good job, if it passed with flying colors, the risk context or risk appetite is such that the other five that we didn't look at, we think it's probably good enough, given that it did so well on those five.
REID BLACKMAN: If, on the other hand, you do that random audit and they're a disaster or if it's 50/50, some were great, some were bad, you're not going to trust it for those other ones. So when it comes for a thing like summarization, large language models, you actually have to do the work to verify it. Unfortunately, there's no other alternative, which potentially makes it potentially useless, if you have to do all the research or verifying it in the first place.
REID BLACKMAN: So you've got to figure out what are the risk con-- what are the contexts in which we're comfortable using it? What does testing it look like? What's the comparison? What's the relevant comparison? Is it just a human alone? What's the training look like, so we can do these summaries? But one thing that you might do is you might train the end user in a certain way, so that they can use those answers anyway.
REID BLACKMAN: So let me give you a an analogous example. First, I'll give you an interesting example. It turns out some research shows that-- not that I performed-- chess-- oh, sorry. AI beats humans at chess, all day, every day. But apparently, AI plus human beats AI alone at chess. Odd result and shocking, empirical finding.
REID BLACKMAN: Turns out to be true. And so, relatedly, when it comes to bias issues, you could have a biased AI, a discriminatory AI. But if you combine it with a human and the right way, you can get a less biased output overall. I'm going to give some coarse-grained numbers, that sort of thing. Let's say that a human is biased at a 6 out of 10. An AI is biased 4 out of 10.
REID BLACKMAN: But it turns out that when you introduce the AI to the human, you say, look, the AI is biased in certain ways. And so you just counteract for that. As you use it, you might come up with a lower-- sorry, a lower bias score. I should have set the parameters properly. So it might be that judges are biased when doing sentencing. It might be that this AI is biased when sentencing. But if you tell the judge, hey, listen, this AI is biased against Black people, as an example, take that into consideration-- and when it gives the risk rating, you might wind up with an overall less biased system.
REID BLACKMAN: Same thing with this verification thing. So we have these summaries. Maybe if we tell people, listen, these are summaries generated by a large language model. We've done such and such to try to mitigate hallucinations or false negatives, not including something in the summary that they ought to or false positives, including something that it ought not to have.
REID BLACKMAN: Take that into consideration as you use it for whatever your intended use case is. And it might be the case that that, the AI testing that you've done, plus educating of the end user, is sufficient to make it useful, usable and, more or less, safe or, at least, as safe, if not safer, than the alternative. Does that answer the question.
SPEAKER: Yeah, it does.
REID BLACKMAN: Cool. It's a very long-winded explanation.
SPEAKER: I actually have a question.
REID BLACKMAN: I just really like the chess thing.
SPEAKER: Yeah. So, obviously, this is a space where things are changing very quickly. And I think we've seen a lot of people are in the rush to get to market, are OK with the ethical sacrifices that they've made along the way, for example, hoovering up the content of everyone in this room without their consent. Do you feel hopeful at all? Have you seen us moving in a direction, whether through structures or policies or lawsuits or whatever it is, toward having more ethical AI?
SPEAKER: What do you think is that trajectory?
REID BLACKMAN: Yeah, so-- [LAUGHTER] Yes, but it's going to be a real bumpy ride. I'm not so sure that's an answer. I mean, this might be more of a reflection of my personality or something like that than it is a well-considered view. But I think, yeah, there's going to be some bad stuff. I don't see how we're going to avoid the bad stuff. I work with companies. But I work with-- I have a small boutique consultancy.
REID BLACKMAN: There's a small subset of the set of companies developing and using this stuff that actually has an AI ethics or responsible AI program in place. We're going to see Cambridge Analytica-type scandals. I'd be shocked if we didn't. We've seen little things, so far. Some of you may have heard of-- This is not so scandalous as it is almost stupid. But Canada Air had a customer-facing chatbot powered by an LLM.
REID BLACKMAN: The person, the customer asked about certain-- what's the deal with bereavement fares? Gave him some information. He said, great. He bought the ticket. He was supposed to get reimbursed. Then he tried to get reimbursed. And Canada Air said, oh, no, that's not true. You're not going to get reimbursed for that.
REID BLACKMAN: The AI said that, not us. It was a very weird move. [LAUGHTER] And he sued, and he won. That's a tiny blip. But we're going to see-- we've seen biased AI in hospitals. So there's been an AI released by a major health care company. This was covered both in The Wall Street Journal and The Washington Post, where AI recommended to doctors and nurses to pay more attention to white patients than to sicker Black patients.
REID BLACKMAN: They weren't intending to do bad. They weren't skewing the data to get a certain result. They just mistakenly thought that how much money you spent in the past on health care was a good indicator of how much you need health care in the future. And so they used that. Since they thought that was an indicator, they used that to train their AI, not thinking, oh, right, if you don't have the money to spend, then you can't spend it.
REID BLACKMAN: That doesn't mean you don't need health care. It just means that you don't get it. So that was an unintentional introduction of bias. So I think we'll see either the accumulation of-- I don't want to say, little, but we're not talking about atom bomb-level stuff, ethical disaster. But we're talking about something that's pretty ethically bad. We'll see lots of those things, of a smaller variety.
REID BLACKMAN: I think we'll probably see some bigger things. And that's the kind of thing that will catalyze people taking it responsibly. Yeah, so yes, I think we're going to get there in the long term. It'll be a very bumpy road. That's when companies spend budget. [LAUGHTER]
SPEAKER: Probably have time for one more question.
REID BLACKMAN: I know there's a hand here. Yep.
MICHAEL MUSKAT: Hi, I'm Mike Muscat. I'm with AACN. And I'm just wondering if you think there is a role to be played by explanatory AI. Are there different ethical questions that it raises? You mentioned healthcare applications. And I recently interviewed a nurse who had written an article about how AI models were better predictors of risk in a critical care unit than were traditional models.
MICHAEL MUSKAT: It was all about trust. It was all about ex-- I'm just wondering if explanatory AI is-- it's not a panacea. But is that something that's going to raise other ethical issues? Or do you think it's going to obviate some of the problems with traditional AI?
REID BLACKMAN: Yeah, I think explainable AI is, all else equal, a good thing. If you can understand the rationale behind the output or the decision, you're in a place to assess it. Whereas, if you're not, you can't assess if you like the reasoning or the rationale. So, all else equal, yeah, it's a good thing. We should have it. It's also good for other non-ethical things, like squashing bugs, technical bugs.
REID BLACKMAN: If you understand how it works and you're getting weird outputs, and you understand how the thing works, now you can say, oh, that's why we're getting that weird thing. And then they can squash that bug. So it's useful for a variety of reasons. It could potentially increase trust, if the explanation is of the right sort. It can also decrease trust, if the explanation is of the wrong sort.
REID BLACKMAN: I'm not inclined to think that it's a necessary condition for trust or reliability or something like that. Yes, it would be nice. Yes, it would be good. Are there other things? Here's a quick and dirty example that I like to use. Suppose you've got this magic box. It's got two buttons on it, A and B.
REID BLACKMAN: And you find that every time you hit button A, some random person dies. And every time you hit button B, some random person with cancer is cured of cancer. Now, you've done this 1,000 times. I'm not sure that you should have engaged in that kind of testing so haphazardly. OK. But put that point to the side.
REID BLACKMAN: You now know, I think, with pretty good reliability, "I don't know how it works. Its magic, but let's keep hitting button B. And everyone, don't hit button A." And so it's completely unexplainable, but it's acted sufficiently reliably, such that you know how you ought and ought not to use it. So explainability is nice, but I don't think it's necessary.
SPEAKER 1: All right. Well, with that, join me in thanking Reid Blackman. Thank you so much.
REID BLACKMAN: Thanks, everyone. [APPLAUSE]
SPEAKER 1: We do have copies of Reid's book back here. And with that, we'll hand it over to Will to close us out.
WILL: Thank you very much, Reid, and to all of our speakers today. A really big thanks to all of our sponsors, to the entire Silverchair team that made today possible and to you all, for the gift of your time. It means a lot. And I think nearly all of us assumed positive intent, maybe with everything except the restaurant service today. And for contributing generously. If we've achieved one thing today, it is that I hope we left you with a sense of optimism, whether that is about your mission, whether it is about our industry, whether it is about how AI can help all of us in our jobs to be done.
WILL: It is a trite note, but the future is on our platforms, whether it is the content we publish or it is our users. And that is consequential. We do very meaningful work in this market. And we should all be very proud of it. So like I started, earlier today, I really do believe that technology can fulfill the promise of science and scholarship. And I hope, today, you've felt that, too.
WILL: So with that, we're getting ready for cocktails on the roof, to the very, very top floor. I have two asks for all of you, while there. One is please speak with someone new. There's a lot of great people here. This community needs knowledge-sharing to solve so many of the problems or to see so many of the opportunities that we talked today. So connect with someone new.
WILL: I think that'll be very helpful. And plan to join us next year. So thank you all very much. And again, to our speakers, we really appreciate you all coming. Thank you. [APPLAUSE]
Segment:0 .
NIALL LITTLE: Hi, everyone. We're going to get ready for our keynote speaker here. My name is Neil Little. I am the CTO and technical co-founder at Hum. And I'm also a proud Silverchair alumni. This has always been one of my favorite conferences. So I'd like to thank Will and Stephanie and their entire team for putting on such a great program today. Let's give them a round of applause. [APPLAUSE] As an AI company, at Hum, and stewards of your data, ethical AI is a topic we take very seriously, which is why I'm super excited to be introducing our closing keynote speaker today.
NIALL LITTLE: As artificial intelligence reshapes the landscape of digital experiences, it's also crucial that tech providers and organizations consider the ethical implications of these powerful technologies. Dr. Reid Blackman is an accomplished author, thought leader and an expert in AI ethics and digital risk management. As the founder and CEO of Virtue, a digital ethical risk consultancy, Dr. Blackman has been at the forefront of helping Fortune 500 and Global 1000 companies scale AI and digital ethical risk management programs.
NIALL LITTLE: His expertise has been sought after by industry giants, and he has served as a senior advisor to the Deloitte AI Institute and a founding member of the Ernst & Young's AI advisory board. He also shares his ethical AI insights in his book Ethical Machines, Your Concise Guide to Totally Unbiased, Transparent and Respectful AI. And his work has been featured in The Wall Street Journal, The Harvard Business Review and TechCrunch.
NIALL LITTLE: Please join me in welcoming Dr. Reid Blackman. [APPLAUSE]
REID BLACKMAN: Hey, everyone. How you doing? You've made it to the end of the day. Now you have to listen to a philosophy lecture. You draw the short straw there. OK. So let's see. What should I talk about. I guess I'll talk about AI ethics. This is not my slide, though.
REID BLACKMAN: Does this work? That's me, blah, blah. OK. So here's where I'm going to take you today. At least, this is the plan. First, I want to show you how I conceive of the AI ethics landscape. And I want to talk about, of course, AI ethical risks. And then I want to talk about how you might think about those risks in the context of scholarly values, scholarly publication values.
REID BLACKMAN: And then, finally, I'm going to talk a little bit about, pragmatically, what are my suggestions for what you ought to do, going forward. So that's the plan here. First thing, I want to talk about what AI is. AI is just software that learns by example. People talk. It's supposed to be this really fancy, complicate-- it's software that learns by example.
REID BLACKMAN: That's it. So when you hear me say, AI, or you hear anyone say, AI, for that matter, just think, "software that learns by example." That's all it is. You know what it is? You know what software is, right? Everything on your phone, every website software got it covered. You know what learning by example is.
REID BLACKMAN: I have a six-year-old. At some point, I told him, this is what pigs look like. And now he can spot pigs in new books. He's very talented. To take a more complicated version, I could give you a bunch of impressionist paintings. At some point, you've probably seen a bunch of impressionist paintings.
REID BLACKMAN: And now you've learned what impressionist paintings look like. So you go to the National Art Gallery, and you see some impressionist paintings. And you go, oh, yeah, that's an impressionist painting. I know that from all the examples that I saw. That's all AI is doing. It's software that learns by example.
REID BLACKMAN: If you want your AI to learn what Pepe, your dog, looks like, give it a bunch of examples of what your dog looks like, photos. It'll, quote, unquote, "learn" what Pepe looks like. And when you snap a new photo of Pepe, it'll put it in the Pepe folder. Because it recognizes, oh, yeah, that's Pepe. I know that guy. I know this is not your area, but if you wanted to approve or deny various mortgage applications, give it a bunch of examples of, oh, here are all the good mortgage applications that we approved in the past.
REID BLACKMAN: Here are all the ones that we denied. And it's going to learn, from those examples, what the good mortgage applications look like, what the bad ones look like, so on and so forth. Of course, there are complications. It messes up, blah, blah, blah. But at the core, it's just software that learns by example. If you want your AI to output text in a way that's coherent, then give it a bunch of examples of words cobbled together coherently.
REID BLACKMAN: And it'll learn, by those examples, how to do exactly that. That's all large language models are, LLMs. They're just cobbling words together based on all the other examples it has of words being cobbled together. That's it. So I don't care if you're talking about generative AI or large language models or you're talking about so-called narrow AI.
REID BLACKMAN: It's all software that learns by example. If you want a fancy word, for example, you can use the word data. That's the fancy word, for example. That's it. So instead of saying, AI is software that learns by example, you can say, AI is software that learns by data. It's a little bit more high brow than saying, learns by example. And if you want an even more highbrow, really fancy, call it "training data." AI is software that learns by training data.
REID BLACKMAN: You give training data to the software, and "training data" is just another fancy word for "example." And I think it's important that we just recognize that we're just talking about software that learns by example because data scientists, computer scientists, people who are trying to sell AI usually talk in this opaque way that intimidates us. I'm not a technologist by background, right? I said I'm a-- by background, I'm a philosopher.
REID BLACKMAN: An odd thing to be, but that's what I am. And you don't need to let people say things like, training data, convolutional neural networks. Software that learns by example, relax yourself. So let's make sure that they don't push us around with all their fancy words. OK. Once we've got that-- now, here's the nice thing about learning that AI is just software that learns by example.
REID BLACKMAN: It allows you to see the AI ethics landscape really easily. So when I think about AI ethics, I distinguish between what I call "AI for good," on the one hand, and "AI for not bad," on the other. The people in the "AI for good camp," they ask questions like, how can we do a bunch of good in the world using the powerful tool that is AI? So for those of you who are familiar with the United Nations Sustainability Development Goals, the SDGs, lowering poverty worldwide, increasing education, increasing women's rights, et cetera, et cetera, you might ask, OK, how can we use the powerful tool that is AI to pursue those SDGs?
REID BLACKMAN: So it's, if you like, wearing the ethics on its sleeve. The objective, the goal is to do something ethical using the powerful tool that is AI. Great stuff, noble people, people should do that. I'm not. Not what I'm going to talk about, either. I focus on "AI for not bad." And as you can imagine, that means you're going to do the things you're going to do.
REID BLACKMAN: Let's assume they're not ethically evil. Let's assume they're either ethically neutral or ethically good. And then the question is, how can we pursue those normal business objectives, ethically acceptable, but not particularly ethically amazing goals? How can we do that without ethically messing things up with AI?
REID BLACKMAN: How do we not go sideways on this? So that's "AI for not bad." You can call it "AI risk mitigation" or "AI ethical risk mitigation" or "AI ethical management." You can call it "AI governance." You can call it "responsible AI." All these words are floating around in this space, it doesn't really matter. I think "AI for not bad" is the one we should go with.
REID BLACKMAN: Nobody else listens to me, but that's what I think we should do. Once we got "AI for not bad" in our heads and what we're doing is we're just-- we want to identify and manage the ethical, reputational, regulatory, legal risks of AI. There's another distinction that I like to make between content and structure, content and structure. So the content question asks, what are the ethical risks of AI?
REID BLACKMAN: And what are their sources? Why are we even talking about-- what's the sources of these ethical risks? Why is it such a topic of conversation? Why am I up here right now? Why are you listening to me? That's the content side. The structure side is in, what are we going to do about it? As an organization, what are we supposed to do about that sort of thing?
REID BLACKMAN: For the most part, people look at the content stuff very superficially. You've probably heard about biased AI or black box AI or privacy violations. People sometimes freak out, and they say, OK, well, what are we going to do about it? We got to do something? So they treat the content side very superficially, sort of at the headline level.
REID BLACKMAN: And then they jump over to the structure side and ask, what should we do? And then they say, I don't know what to do. Usually, they come up with some piece of paper with values on it, like, we're for justice and privacy and respect for all. And it doesn't mean anything because they have no idea how to actually put that stuff into practice. So people don't really do much.
REID BLACKMAN: And one of the things that I contend is that we would do a lot better, on the structure side, knowing how to manage those risks, if we actually went deeper on the content side. And in one way, what I'm saying is utterly mundane because all I'm saying is, if you don't really understand the nature and the source of your problem, you're not going to be able to develop a good solution for it. That's all I'm really saying.
REID BLACKMAN: I'm very unimpressive. [LAUGHTER] So I want to dig deeper on the content side. You'll get a better sense of where these ethical risks are coming from. And then we can talk about what can you do about it. OK, content side, blah. There are three main ones. And I don't care if you're talking about generative AI or more narrow AI.
REID BLACKMAN: Can you see that? It's a little bit low. Privacy, black box and bias, so let's take these one at a time, more or less. So let's go back to Pepe. You're taking pictures of your dog, Pepe. Now, when you look at Pepe, you're looking at two eyes, a nose, a mouth, ears. And so you learn what Pepe looks like.
REID BLACKMAN: And then you can say, oh, yeah, that's Pepe, right? AI doesn't do that. Software that learns by example doesn't do that. It's not looking at ears and eyes and a nose and a mouth. It's looking at pixels. It's analyzing these pictures on the pixel level. And it's looking at the thousands of pixels and the thousands of mathematical relations among those pixels.
REID BLACKMAN: It's really complicated. So when it says, "Yeah, yeah, that's Pepe" or, "No, it's not Pepe," it's because it's identifying a phenomenally complex mathematical pattern in that photo and saying, "Yeah, that really complex mathematical pattern fits the same mathematical pattern that was in all those previous photos of Pepe" or No, it doesn't." Let's just say this mathematical pattern is 50 pages long, right, just 50 pages of equation.
REID BLACKMAN: You and I can't keep that. At least, I mean, I certainly can't. I'm betting that most of you can. Keep that in your head. Right, it's too long. You speak the language, but the sentence is way too long for you to follow. And so when it says, "Yes, this is Pepe," "No, it's not," we can't actually understand why it's saying "yes" or "no." We know it's based on some mathematical pattern, but the pattern itself is one that we cannot comprehend because our minds are so finite.
REID BLACKMAN: So why does it say, "Yes, Pepe" or "No, Pepe"? We're not really sure. It's something about this mathematical pattern. But we can't really explain to you how it results in or how it outputted, "Yes, that's Pepe," "No, it's not." Now, we're talking about photos of Pepe. Who cares? But if you're talking about mortgage applications, job applications, applications to a university, sentencing guidelines, sentencing judgments for criminal defendants, now you start thinking, maybe we need to know why this thing is saying "yes" or "no" or "high risk," "low risk" or something along those lines.
REID BLACKMAN: So that is, in essence, the black box problem. And so some people think-- and we can talk about this in Q&A if you'd like-- that we really need to solve the black box problem if we're ultimately going to trust and use AI safely and responsibly. I'm not persuaded of that totally. Again, we can talk about that. But that is at least a concern on people's mind, the black box problem.
REID BLACKMAN: OK. Now, all those examples that the software is learning by, all that data, to use the fancy word, is often not always, but often about people. Which means that, in order to get this thing working well-- because, all else equal, the more examples you give it to learn by, the better it learns, the better it does. The more data that you have, the better your AI is going to be. Which means that the data scientists who want to train their AI to make it as awesome as possible are incentivized to hoover up as much data as they can because that is the most obvious pathway to getting more awesome AI.
REID BLACKMAN: But if that data is often about people or, let's say, it's the [? IP ?] of people, then we have privacy violations on the horizon, right? Because the fuel of machine learning or AI-- I don't know if I said, machine learning, yet. AI, machine learning, software that learns by example, treat it all the same. The fuel of it is data. And if they want to go far, they need to get a lot of fuel.
REID BLACKMAN: So that's one way. There's more, but that's one way in which we risk privacy violations, by virtue of the nature of the beast. Now, actually, I want to say something about the nature of the beast. Notice that, so far, both the black box issue and the privacy issue are not incidental to how AI works. It's not like, oh, yeah, we happened to use the screwdriver to kill someone.
REID BLACKMAN: No, no, it's not coincidental to the tool. It's built into the tool that these kinds of things are not merely likely. Sorry. They're not merely possible. They are likely. They're not merely possible. They are probable, just given the way the technology works. Let's talk about the bias issue now, which, probably, a lot of you are familiar with.
REID BLACKMAN: You give it a bunch of examples. If the examples are, in some way or other, skewed, you're going to wind up with skewed or biased and potentially ethically problematic results. So for example, if you give it a bunch of, let's say, medical data that predicts whether someone's going to develop diabetes in the next two years and you only give it data related to-- a very small percent of all that data is, let's say, about Black men.
REID BLACKMAN: You haven't given enough examples for it to learn the probability that Black men will develop diabetes in the next two years, given their medical profile. To take a simpler example, a famous or infamous one at this point, when they first started developing facial recognition technology, it was really bad at recognizing Black women especially, Black people, darker-skinned people, especially Black women, for the simple reason that they just did not have a tremendous amount of photos of Black women in various lighting conditions, taken from different angles.
REID BLACKMAN: And so it just didn't learn very well. So what you wound up with was facial recognition software that was good at recognizing white men in photos, but bad at recognizing Black women. So that's the bias problem. There's, of course, more to it. This is 30,000-foot stuff, but you should get a sense of where this is coming from. Again, this is not incidental.
REID BLACKMAN: This is built into the technology itself. And another thing to note is that we're not talking about bad behavior, right? I'm talking about ethics, but I'm not talking about the bad behavior of people. I'm talking about things that often accidentally, even in some cases, blamelessly-- although, there's more to say on that-- get introduced into the AI without anyone intending to do bad things.
REID BLACKMAN: So with the screwdriver, if you're going to do something really evil with a screwdriver, you really got to try hard, right? You got to intend to do something bad. But you can create privacy-violating AI that's a black box, that systematically discriminates against women without any intention to do so. This is one of the reasons why we're talking about this stuff. Again, it's not merely possible.
REID BLACKMAN: It's probable, given the nature of the beast. You can also have use-case-specific risks, in addition to these. So this is just built into the technology. They can be mitigated. They can be managed. But it's built into the technology, as opposed to, say, using AI to power or create a self-driving car.
REID BLACKMAN: We're now maiming and killing pedestrians, is an ethical risk. That's a high priority. That's not because of the nature of the beast that is AI, that is software that learns by example. It's because the purpose is to which you put the beast. OK. Everything I said, so far, applies to generative AI, large language models, just as much as to narrow AI.
REID BLACKMAN: Is this person a good candidate for the job? Does this person deserve a mortgage, whatever it is. When we come to large language models in particular, we get two other kinds of ethical risks that are special. One, I'm sure that you've heard this phrase, hallucinations. You get sentences that are just false. You get things that are just false. So rather famously, infamously, there was a lawyer who went to the courts citing various court cases in his defense.
REID BLACKMAN: But the judge looked into it, and it turns out those court cases weren't real. And the reason was that the lawyer went to ChatGPT. He said, hey, I'm looking for some cases that do this kind of thing, that demonstrate this kind of thing. And it spat out some things. The lawyer didn't have the wherewithal to check or to verify those things, presented it in court. They were completely made up, completely fictional.
REID BLACKMAN: And he got in a lot of trouble, a lot of trouble. So why does it do that? Why does it hallucinate? In short, remember what AI is. It's software that learns by example. When it learned to create outputs of a textual variety, how is it doing it? It's just cobbling words together coherently, based on the examples of other things it's seen, other words it's seen cobbled together coherently.
REID BLACKMAN: It's not doing it with an eye towards the truth. It doesn't know the truth. It's not tracking the world. That's not what it's in the business of doing. It's in the business of putting words together coherently. And so since something can be coherent, but plainly false, you get hallucinations. It's built into it. There's various kinds of attempted solutions to this thing.
REID BLACKMAN: Some of you have heard of RAG, for instance. Nothing is settled, yet. No one has solved the hallucination problem. People might say, oh, we solved it. We have a start up. We're looking for funding. We solved it. OK, we'll see. So the hallucination problem is one problem.
REID BLACKMAN: And then there's something that I call the "deliberation problem," which is a slightly more subtle problem. It's super fascinating. It scares me a bit because people don't get this one, by and large. The deliberation problem is this. You might think that the large language model at ChatGPT or Anthropic's Claude or whatever it is is giving you an explanation for why it told you what it did.
REID BLACKMAN: Why did you tell me that this is the court case? Why did you tell me that this is the appropriate diagnosis? Why did you-- whatever it is. And it'll come up with something. And it might even appear reasonable. But the crucial thing is that, whatever it outputs, it's not telling the truth. And it's not telling the truth because it doesn't give outputs on the basis of evidence.
REID BLACKMAN: It doesn't give it on the basis of reasons. It's not a reasoning machine. It is a next-word-prediction machine. And so when it says, the reason that I told you that this person probably has cancer is that, blah, blah, blah, blah, blah, that blah, blah, blah, blah, blah, blah, blah has nothing to do with why it told you that person is going to develop cancer.
REID BLACKMAN: Because it's just not that kind of thing. It's not reasoning, in the first place. So what is it doing when it gives you that explanation? It's just doing its normal thing of putting together words, coherently, that make sense with what came before, even though it bears no actual relation to the stuff that came before. So the deliberation problem is a problem of, in some sense-- I'm not sure that I want to use this word.
REID BLACKMAN: I'm going to use it anyway, deception, because it's very easy to get fooled into thinking that you're dealing with a reasoning machine. Because you can ask it for recommendations and then reasons why it gave those recommendations, and it will say things that look like evidence or justification for why it said what it did. But it's making it up.
REID BLACKMAN: It's just cobbling words together that it thinks you're going to be like, oh, yeah, that makes sense. So that's the deliberation problem. OK. So now we're beginning to see the landscape a bit more, right? We've got AI ethics, more generally. You want to do some great things. That's "AI for good." You want to not ethically screw up pursuing the ordinary business objectives that you have.
REID BLACKMAN: That's "AI for not bad." Then we distinguish between the content and structure. Where are these risks coming from? Talked about that. It's because it's software that learns by example, making these things not merely possible, but probable. And then we've seen five main ones. There are others.
REID BLACKMAN: I just didn't include them here because one can only say so much in this time. But there are others. These are the biggest ones, I think. So now let's talk about the structure stuff. Now that we have a grip on what the risks are, where they come from, why we're talking about them, what are we supposed to do about this? Here are two suggestions.
REID BLACKMAN: And this is tailored for this audience. This is not what I would say if I were talking to chief information officers that have 100,000 employees. It's a different beast. That's a much bigger lift. But for smaller organizations that are trying to figure out where does AI fit here, you need something like an AI ethics board or an AI ethics committee or an ethical risk committee or a brand risk committee or whatever you want to call it.
REID BLACKMAN: I don't care. I'm going to call it an AI ethics board. You need that, and then you need some tool for risk assessment. So I'm going to run through these kinds of things, about what they should, more or less, look like. Because I think that, all else equal-- I don't know your position, of course. But you might want to go back and create this kind of thing as you think about integrating AI into your organization.
REID BLACKMAN: So the ethics board, it should be cross-functional. It can't just be data scientists. You're not going to math your way out of these problems. You need a cross-functional, cross-departmental ethics board who can weigh in, in various ways, about whether this thing is ethically safe, reputationally safe, et cetera. Legally safe, you'll want a lawyer, regulatorily safe. You'll want to include external stakeholders, especially, for instance, authors or researchers, since they're going to have lots to say about what's getting done with their work.
REID BLACKMAN: And this board has to have the authority to issue go or no-go decisions. They have to be able to say, mm, this is just way too risky for us, just no. That's the advice anyway, that they should have that authority to issue go, no-go decisions, as opposed to mere advice. But maybe that advice gets downplayed when there's a lot of money on the line. And you think, thank you, ethics board, but you're being overly conservative.
REID BLACKMAN: No, thank you. There's many millions to be made here, et cetera, et cetera. That's when we see ethical breaches. But if it has the authority to say, sorry, red light, then it should be a really, really big deal if that board is going to get overruled, as opposed to, they didn't follow the advice. That's a bit about the ethics board. And then we have to talk about what's a risk assessment going to look like.
REID BLACKMAN: Now, this has to get customized to your particular organization. So I've just taken some values that I cobbled together-- I guess I'm like an LLM-- from the Society for Scholarly Publishing and the Association of University Presses. Stewardship is-- who is that? I think that's the Association of University Presses. Community is Scholarly Publishing. Integrity is on both people's lists-- both organizations' lists.
REID BLACKMAN: Intellectual freedom and adaptability and equity and inclusion-- They can be different. I'm just using these as examples. If your organization has different values that are close, but not identical, use the ones you've got. OK. So here's how it goes, roughly.
REID BLACKMAN: This is not exhaustive, but if your ethics board is convened and you've got a solution that you're thinking about, this is a rough-and-ready tool for guiding that discussion. So you can ask things like, do we have buy-in from authors for the solution? Do we know what happens to the data that a user inputs into the AI? Do I know what happens to the AI outputs? Can we clearly communicate the risks or best practices for use of this AI?
REID BLACKMAN: Did we provide readers or clients with the appropriate resources for using the solution responsibly? Again, not exhaustive, but instructive, I hope. Gives the ethics board something to do, something to talk about. Are readers, clients able to verify the outputs? Have we considered whether using a black box AI in this context is inappropriate? By the way, all large language models, ChatGPT, Anthropic, Gemini, et cetera, they're all black boxes.
REID BLACKMAN: [LAUGHS] They have kind of a clue, but it's more of a suspicion. They understand how 1% of it works. They don't know. The outputs are all unexplainable, which is understandable when you think about the vast quantity of data that these things are trained on. Trying to figure out and understand the mathematical patterns that underlie the predictions, we're not going to do it.
REID BLACKMAN: Do we have the expertise to assess whether the AI is fit to the task or it does a good enough job? I'm stressing-- I put "enough" in parentheses. It's not going to do a perfect job. If your goal is perfection, forget it. Don't use AI. But one thing to always keep in mind is that there are three kinds of systems that we're actually talking about.
REID BLACKMAN: There's the human alone, making decisions, recommendations or doing the work. There's AI alone, making decisions, doing the work, whatever it is. And then there's AI plus human. So those three different systems, if you like, are accurate or inaccurate to varying degrees. And one question is going to be, when you ask, "Is this sufficiently safe to deploy," is, "Well, what are you comparing it against, human alone, human plus AI, AI alone, et cetera?" OK.
REID BLACKMAN: Where are we now? Is there a way this AI can operate to inappropriately stifle legitimate views? So think about the bias issue, right? So suppose that you're trying to unearth summaries of texts or something along those lines. And you have very little text or very little content coming from, say, a certain region of the world. You're probably going to get not that view bubbling up to the top, not in the AI's outputs, because it's such a small part of its training data.
REID BLACKMAN: Is the AI designed in a way that encourages inappropriate deference to the AI outputs? Deference to the AI outputs is a big thing. You may have heard stories about people literally driving into lakes because Google Maps told them to. This doesn't feel right, but I don't know. The computer says so. Woop. [LAUGHTER] And that's really like, what are you doing?
REID BLACKMAN: I didn't do this. I almost did this recently. But I did turn into a harbor when I knew that I couldn't possibly-- you're right. [LAUGHTER] But imagine if it's a lot lower stakes, and the lake is not literally in front of your face. And it says, yeah, you should go do X. Or, yeah, the report says Y. Yeah, OK.
REID BLACKMAN: Well, sure yeah. Easy to defer, very easy, especially to computers. But we ought not to. Do I know whether or how this AI solution may be biased or discriminatory, which stakeholders may be at risk, and why? So this is just a summary. I'll actually share this with you. I didn't talk about this with Silverchair.
REID BLACKMAN: But I'll find a way of distributing this slide to you, so you can all use it. It's yours. So you can take a picture if you'd like, but I can also just send you a version. So let's look at a quick and dirty example. Here's the opportunity. Notice, by the way, I use the word "opportunity." I'm not a skeptic.
REID BLACKMAN: Maybe. I'm a skeptic. I'm not a doomsday person. I don't think, oh, this stuff is terrible, and it's never going to-- I don't know. It could work. Sometimes I've seen it work. It's amazing technology.
REID BLACKMAN: So don't take anything I say to indicate, forget the whole thing. It's never any good. I don't think that at all. I think there's a tremendous amount of opportunity. I also just happen to think that we should probably watch out for the risks as well. So let's take this opportunity and run through an example of what would the questions look like.
REID BLACKMAN: AI can read articles and books, including charts, infographics, et cetera and create audience-appropriate summaries, analysis and recommendations for various stakeholders. So you can imagine a doctor wanting to get recommendations on how to treat this patient, or a nurse or somewhere in the healthcare system wanting guidance on this sort of thing, or a lawyer wanting guidance on "What kind of cases should I be looking at," so on and so forth. So that's an opportunity.
REID BLACKMAN: And given the insane mountain of data, the insane mountain of text that everyone in this room is stewards of, we could come up with some amazing things. So there's a great opportunity. But as we already made our way towards, are the readers, clients able to verify the AI outputs? So you've got all this text, and then it summarizes. It says, well, essentially, this documents or these 10,000 pages, the upshot is X. And the major recommendation is Y.
REID BLACKMAN: Can you verify those outputs? Yeah, in principle, they should be able to verify it. They better not be too differential, but in principle, they can verify it. That's not to say they necessarily will, by the way, because they might not have the time or the bandwidth. Or they have due dates. Or they have a boss that says, just get it done. But in principle, they should be able to verify it.
REID BLACKMAN: It also might be the case, though, that it's going to take too much time. It's 10,000 documents. I don't have the time to read those 10,000 documents and then verify. That was the whole point of using the AI. So in principle, they can verify it. But can they really verify it? Can they verify it enough?
REID BLACKMAN: And remember, can they verify it in a way that results in an overall outcome that is better than the alternative? Not is it perfect, not perfect, but better than a human poring through those 10,000 documents, yada, yada, yada. Because the human's going to miss stuff, for sure. They're also going to include stuff that they shouldn't have included in the summary. Same with an AI.
REID BLACKMAN: Have you considered whether using a black box is inappropriate? You may decide that. So long as you can independently verify the outputs, the fact that it's a black box doesn't matter. So you might just care about reliability. The black box issue is an issue. It's something that deserves thought. Attention, but it's not a nail in the coffin of AI, of LLMs.
REID BLACKMAN: It just means, hey, is this use case-- is it OK for us to use a black box? You might think, yeah, I don't really care how it gets the outputs, as long as we can show that it's efficiently, reliably giving out the right answers. That's a view. Do we have the expertise to determine whether the AI is good enough?
REID BLACKMAN: And this is going to potentially vary by context. Third-party subject matter experts may be needed, so you might need to bring in other people, externally, to vet this thing. So you can see you're starting to get some sense of can we use this thing. I think, right now, so far, the biggest worry is about can they verify it. But what's the context?
REID BLACKMAN: How much time do they have to verify it? That's going to matter, to whether you give a yes or no decision. Let's do one more. Do we have buy-in from authors for this solution? And that's going to vary by context. It's going to require stakeholder interviews, focus groups, et cetera. It's worth talking to them.
REID BLACKMAN: And so this is, again, not a no. This is not a red flag, stop all things. Oh, you know what? Right, we should talk to some representative authors. Not all of them, but we should form some focus groups and see what our community thinks. Do we know what happens to the inputs or outputs of the data? And this might vary by provider. And so you might have heard things like, is the AI in the cloud?
REID BLACKMAN: Is it on prem? Is it on premise? Is it on your property, so that no one has access to it? Are you sending it to a third party, like an OpenAI or a Microsoft or a Google? And if so, what do they- do they see the inputs? Do they see the outputs? What are they doing with it? What are they allowed to do with it?
REID BLACKMAN: What are they not allowed to do with it? Again, red flag, but we can change it to green if we get the right kinds of answers. And can we clearly communicate the risks or best practices for use of this AI? Probably. [LAUGHS] In principle, yes, you can say the things. The question is, of course, whether you can get the people who need to hear the things actually hear them.
REID BLACKMAN: Can we get them in a room? Can we give them the right instruction? Can they know what they can and cannot do? Do they have the bandwidth to learn the things that we want to communicate to them? Do they have the workflows, the processes in place, so they can actually incorporate those learnings into how they actually use the AI? Things are getting a lot more complicated now.
REID BLACKMAN: But at least this should give you-- begin to give you, at least, some grip on the risks. They're not, in all cases, insurmountable. They're surmountable, on the condition that you give it the right kind of thought, which means that the bottom line is whether using a particular AI is commensurate with your values. It's not a black and white issue. And it requires your due diligence and professional judgment.
REID BLACKMAN: That's the point of the board. The board should be exercising its professional judgment. And you can see those series of questions as a way to assist the board in seeing what kinds of questions it needs to ask. So in conclusion, wrapping up, I think you should create-- probably create, if you're seriously considering integrating AI into your organization, an AI ethics board.
REID BLACKMAN: Number two, create your AI ethics assessment. It could be the exact one that I gave. It could be more tailored to your organization. That would be preferable. But create one. It doesn't have to be super complicated, as you can see. And finally, don't let the technologists push you around. It's just software that learns by example. Thank you.
REID BLACKMAN: [APPLAUSE]
SPEAKER: [INAUDIBLE] for questions.
SPEAKER: All right. And we do have time for questions, great. We have 16 minutes, yeah, minutes, good.
SPEAKER: Hi. Thanks for coming today. How often have you discovered entities who are deliberately skewing the data in order to get results that they want?
REID BLACKMAN: How often are they skewing the data to get the results they want? So if you don't use the word "skewing" and you mean how often are they tinkering or introducing new data, a lot. I mean, what data scientists are doing is they are introducing different data, different examples to get the AI to be more accurate. So that sort of thing happens all the time. And so this is one reason, incidentally, why, if I were talking to a group of data scientists, I would say things like, you need to check for these risks at every stage of the AI life cycle.
REID BLACKMAN: Some organizations will do it when they're just thinking about it. We're just trying to think about, what's the solution here? What are the risks here? Let's note those risks. And now let's start building and testing. And when they're testing, they're testing for accuracy. It's not accurate enough, so they introduce more data. And that's when privacy violations occur.
REID BLACKMAN: Bias occurs and that sort of thing. Does that answer your question?
SPEAKER: I think you're-- I think you're suggest-- sorry. It seems like you're suggesting that they're skewing the data to introduce more and more data that would provide accurate results, truthful results. I'm asking you have you discovered any entities who are doing the opposite.
REID BLACKMAN: Where they're decreasing the data?
SPEAKER: No, they're probably increasing-- it's possible-- the data to get certain results that they want beforehand, that are not truthful.
REID BLACKMAN: Ah, no. I can't think of any particular cases where I've seen such things. Usually, the problems are just, we didn't realize it was going to do that kind of thing. It's usually negligence of some kind, not-- the data scientists are in this little bubble. [LAUGHS] They just want the number of-- the accuracy percentage to go up.
REID BLACKMAN: That's all they want. They're not thinking about marketing and sales and bottom-line stuff and satisfying the client. They don't even know the client. They're tinkering with their computers. So they're not trying to get-- it's not politically motivated or something like that, to get a certain output, typically. I mean, I see you-- "smirk" is not quite the right word, but something next to a smirk.
REID BLACKMAN: So let me go on. [LAUGHTER] There are ways in which you can tinker with AI because you think it's inappropriate, right? So rather famously or infamously, depending upon your view, OpenAI, Google, they've all tinkered with the data because they don't think that the outputs are appropriate. They're not safe.
REID BLACKMAN: And so we need to restrict it. And then you are going to see things that you might think are politically influenced in the wrong way. So that's possible. And even, there are things that go sideways, even despite the best intentions. So relatively recently, Gemini was creating pictures of Black Nazis because they didn't have enough Black people in the training data generally.
REID BLACKMAN: So when you ask for an image of, say, a CEO, you always got a white CEO. And people were like, what's up with this? How come every time I search for a CEO or a titan of industry, I always get an elderly white man? And so Google tried to fix this. And then it turns out they did fix that problem, but they went too far. And so now we have pictures of Black Nazis.
REID BLACKMAN: Is that of some help?
SPEAKER: Yeah, we can talk about that [INAUDIBLE].
REID BLACKMAN: Sure, sounds good. [LAUGHTER] Not really? OK.
SPEAKER: Thanks for this talk. It was great. You're talking largely here about ethical use of AI within your own organization. The organizations here are-- you could say that they create data. Or you could say they create content or articles or what have you, human intervention. And how do you think the publishers and those who are creating the content should be looking at ethics, not from an internal use, but from a licensure, external-use perspective?
SPEAKER: Because while a lot of our content in our industry is used without our consent, increasingly there are requests for consent from LLMs and others. So how should we be thinking about the ethics in the outgoing direction?
REID BLACKMAN: That's a great question. So if you're talking about selling the data, so that other parties can use it to train their AI to blah, blah, blah, well, it's going to vary, the answer. First of all, what do your authors think about that. The people who generated all the-- who did all the writing, how do they feel about that? My guess is that they're going to say, you can do that, on the condition that I get compensated.
REID BLACKMAN: And then if they say, yeah, you could do that, on the condition I get compensated, then you have these negotiations. And then they say, yeah, sure. Then you're probably mostly fine, I would say. It's going to be, in some cases, out of your control, what ultimately gets done with it and how it's going to get used. Your responsibility gets attenuated the more the data gets messed with. And it's going to get messed with a lot.
REID BLACKMAN: And it's going to be combined with other people's data a lot. And so at the end of the day, if someone uses it to, let's say, make a recommender for doctors for how they should treat patients, my suspicion is that-- And I have to think more about this. And I have to see the actual case. But my suspicion is that you're so far removed from what the solution actually turned out to be. I don't see you as being ethically responsible.
REID BLACKMAN: Does that kind of case speak to your question?
SPEAKER: Yeah, that certainly answers the question.
REID BLACKMAN: Yeah, there's lots of parties who are going to be response-- because there's going to be better and worse ways to deal with that data. And it's going to be the people who use that data. it's going to be primarily their responsibility to make sure that they use it, that they tweak it, that they skew it and whatever they do with it, they do it in an ethically responsible way.
SPEAKER: Thank you.
REID BLACKMAN: Yep.
SPEAKER: I think I know the answer to this question, but I'm going to ask it anyway.
REID BLACKMAN: It makes my job easy.
SPEAKER: [LAUGHS] The example you used earlier, about the black box, I think you had said there's a 1% chance they know what's going on. And in essence, what happens is, when the results come out, I guess you do a gut check and say, does this look like these are valid results? And if so, theoretically, the process works. Going back a little bit to what I think this gentleman was alluding to, with regards to people perhaps wanting less ethical outcomes to be the results of AI, the fact that the black box issue stands, is there a way to delineate between something that is correct and on the mark versus something that is manufactured and maybe not at all on the mark?
SPEAKER: Does the black box issue make that impossible, then, to determine whether something is something that you can rely on and act on responsibly or something that you have to watch out for and maybe avoid? How does that get resolved?
REID BLACKMAN: So there's a few things to say here. So the first thing is that I wouldn't rely on your gut. Not your gut, in particular, anyone's gut.
SPEAKER: Yeah.
REID BLACKMAN: [LAUGHS] Because gut instincts are remarkably bad, including in ethical situations. I mean, I don't know. People thought being gay was disgusting, and that was what their gut told them. And so that was not an ethically great situation. So I wouldn't rely on people's guts or moral intuition or something to test it. You probably want to test it against certain benchmarks. So let me just take Pepe the dog, the photo.
REID BLACKMAN: We don't know how the thing works, but we don't do a gut check. We look to see how many mistakes does it make. And we gave it 100 examples of Pepe. Did it do a good job? It got 50 right and 50 wrong. That's not good. That's not my gut. That's just, this is really bad.
REID BLACKMAN: It got one wrong out of 10,000. All right. That's really good. So the question is, what's the appropriate benchmark that we're comparing it to, that we want to achieve here? Summary, we don't have those same kinds of quantitative metrics for measuring whether the summary was good. Unfortunately, there, the only way to verify is to actually go do the work.
REID BLACKMAN: There's different ways of doing this. It doesn't mean that you have to test every single output. So there's different ways of testing. So you can say, all right, we're going to have it summarize these 50 documents. And we're going to do a random audit on three of them or five of them and see if it did a good job. And if it did a good job, if it passed with flying colors, the risk context or risk appetite is such that the other five that we didn't look at, we think it's probably good enough, given that it did so well on those five.
REID BLACKMAN: If, on the other hand, you do that random audit and they're a disaster or if it's 50/50, some were great, some were bad, you're not going to trust it for those other ones. So when it comes for a thing like summarization, large language models, you actually have to do the work to verify it. Unfortunately, there's no other alternative, which potentially makes it potentially useless, if you have to do all the research or verifying it in the first place.
REID BLACKMAN: So you've got to figure out what are the risk con-- what are the contexts in which we're comfortable using it? What does testing it look like? What's the comparison? What's the relevant comparison? Is it just a human alone? What's the training look like, so we can do these summaries? But one thing that you might do is you might train the end user in a certain way, so that they can use those answers anyway.
REID BLACKMAN: So let me give you a an analogous example. First, I'll give you an interesting example. It turns out some research shows that-- not that I performed-- chess-- oh, sorry. AI beats humans at chess, all day, every day. But apparently, AI plus human beats AI alone at chess. Odd result and shocking, empirical finding.
REID BLACKMAN: Turns out to be true. And so, relatedly, when it comes to bias issues, you could have a biased AI, a discriminatory AI. But if you combine it with a human and the right way, you can get a less biased output overall. I'm going to give some coarse-grained numbers, that sort of thing. Let's say that a human is biased at a 6 out of 10. An AI is biased 4 out of 10.
REID BLACKMAN: But it turns out that when you introduce the AI to the human, you say, look, the AI is biased in certain ways. And so you just counteract for that. As you use it, you might come up with a lower-- sorry, a lower bias score. I should have set the parameters properly. So it might be that judges are biased when doing sentencing. It might be that this AI is biased when sentencing. But if you tell the judge, hey, listen, this AI is biased against Black people, as an example, take that into consideration-- and when it gives the risk rating, you might wind up with an overall less biased system.
REID BLACKMAN: Same thing with this verification thing. So we have these summaries. Maybe if we tell people, listen, these are summaries generated by a large language model. We've done such and such to try to mitigate hallucinations or false negatives, not including something in the summary that they ought to or false positives, including something that it ought not to have.
REID BLACKMAN: Take that into consideration as you use it for whatever your intended use case is. And it might be the case that that, the AI testing that you've done, plus educating of the end user, is sufficient to make it useful, usable and, more or less, safe or, at least, as safe, if not safer, than the alternative. Does that answer the question.
SPEAKER: Yeah, it does.
REID BLACKMAN: Cool. It's a very long-winded explanation.
SPEAKER: I actually have a question.
REID BLACKMAN: I just really like the chess thing.
SPEAKER: Yeah. So, obviously, this is a space where things are changing very quickly. And I think we've seen a lot of people are in the rush to get to market, are OK with the ethical sacrifices that they've made along the way, for example, hoovering up the content of everyone in this room without their consent. Do you feel hopeful at all? Have you seen us moving in a direction, whether through structures or policies or lawsuits or whatever it is, toward having more ethical AI?
SPEAKER: What do you think is that trajectory?
REID BLACKMAN: Yeah, so-- [LAUGHTER] Yes, but it's going to be a real bumpy ride. I'm not so sure that's an answer. I mean, this might be more of a reflection of my personality or something like that than it is a well-considered view. But I think, yeah, there's going to be some bad stuff. I don't see how we're going to avoid the bad stuff. I work with companies. But I work with-- I have a small boutique consultancy.
REID BLACKMAN: There's a small subset of the set of companies developing and using this stuff that actually has an AI ethics or responsible AI program in place. We're going to see Cambridge Analytica-type scandals. I'd be shocked if we didn't. We've seen little things, so far. Some of you may have heard of-- This is not so scandalous as it is almost stupid. But Canada Air had a customer-facing chatbot powered by an LLM.
REID BLACKMAN: The person, the customer asked about certain-- what's the deal with bereavement fares? Gave him some information. He said, great. He bought the ticket. He was supposed to get reimbursed. Then he tried to get reimbursed. And Canada Air said, oh, no, that's not true. You're not going to get reimbursed for that.
REID BLACKMAN: The AI said that, not us. It was a very weird move. [LAUGHTER] And he sued, and he won. That's a tiny blip. But we're going to see-- we've seen biased AI in hospitals. So there's been an AI released by a major health care company. This was covered both in The Wall Street Journal and The Washington Post, where AI recommended to doctors and nurses to pay more attention to white patients than to sicker Black patients.
REID BLACKMAN: They weren't intending to do bad. They weren't skewing the data to get a certain result. They just mistakenly thought that how much money you spent in the past on health care was a good indicator of how much you need health care in the future. And so they used that. Since they thought that was an indicator, they used that to train their AI, not thinking, oh, right, if you don't have the money to spend, then you can't spend it.
REID BLACKMAN: That doesn't mean you don't need health care. It just means that you don't get it. So that was an unintentional introduction of bias. So I think we'll see either the accumulation of-- I don't want to say, little, but we're not talking about atom bomb-level stuff, ethical disaster. But we're talking about something that's pretty ethically bad. We'll see lots of those things, of a smaller variety.
REID BLACKMAN: I think we'll probably see some bigger things. And that's the kind of thing that will catalyze people taking it responsibly. Yeah, so yes, I think we're going to get there in the long term. It'll be a very bumpy road. That's when companies spend budget. [LAUGHTER]
SPEAKER: Probably have time for one more question.
REID BLACKMAN: I know there's a hand here. Yep.
MICHAEL MUSKAT: Hi, I'm Mike Muscat. I'm with AACN. And I'm just wondering if you think there is a role to be played by explanatory AI. Are there different ethical questions that it raises? You mentioned healthcare applications. And I recently interviewed a nurse who had written an article about how AI models were better predictors of risk in a critical care unit than were traditional models.
MICHAEL MUSKAT: It was all about trust. It was all about ex-- I'm just wondering if explanatory AI is-- it's not a panacea. But is that something that's going to raise other ethical issues? Or do you think it's going to obviate some of the problems with traditional AI?
REID BLACKMAN: Yeah, I think explainable AI is, all else equal, a good thing. If you can understand the rationale behind the output or the decision, you're in a place to assess it. Whereas, if you're not, you can't assess if you like the reasoning or the rationale. So, all else equal, yeah, it's a good thing. We should have it. It's also good for other non-ethical things, like squashing bugs, technical bugs.
REID BLACKMAN: If you understand how it works and you're getting weird outputs, and you understand how the thing works, now you can say, oh, that's why we're getting that weird thing. And then they can squash that bug. So it's useful for a variety of reasons. It could potentially increase trust, if the explanation is of the right sort. It can also decrease trust, if the explanation is of the wrong sort.
REID BLACKMAN: I'm not inclined to think that it's a necessary condition for trust or reliability or something like that. Yes, it would be nice. Yes, it would be good. Are there other things? Here's a quick and dirty example that I like to use. Suppose you've got this magic box. It's got two buttons on it, A and B.
REID BLACKMAN: And you find that every time you hit button A, some random person dies. And every time you hit button B, some random person with cancer is cured of cancer. Now, you've done this 1,000 times. I'm not sure that you should have engaged in that kind of testing so haphazardly. OK. But put that point to the side.
REID BLACKMAN: You now know, I think, with pretty good reliability, "I don't know how it works. Its magic, but let's keep hitting button B. And everyone, don't hit button A." And so it's completely unexplainable, but it's acted sufficiently reliably, such that you know how you ought and ought not to use it. So explainability is nice, but I don't think it's necessary.
SPEAKER 1: All right. Well, with that, join me in thanking Reid Blackman. Thank you so much.
REID BLACKMAN: Thanks, everyone. [APPLAUSE]
SPEAKER 1: We do have copies of Reid's book back here. And with that, we'll hand it over to Will to close us out.
WILL: Thank you very much, Reid, and to all of our speakers today. A really big thanks to all of our sponsors, to the entire Silverchair team that made today possible and to you all, for the gift of your time. It means a lot. And I think nearly all of us assumed positive intent, maybe with everything except the restaurant service today. And for contributing generously. If we've achieved one thing today, it is that I hope we left you with a sense of optimism, whether that is about your mission, whether it is about our industry, whether it is about how AI can help all of us in our jobs to be done.
WILL: It is a trite note, but the future is on our platforms, whether it is the content we publish or it is our users. And that is consequential. We do very meaningful work in this market. And we should all be very proud of it. So like I started, earlier today, I really do believe that technology can fulfill the promise of science and scholarship. And I hope, today, you've felt that, too.
WILL: So with that, we're getting ready for cocktails on the roof, to the very, very top floor. I have two asks for all of you, while there. One is please speak with someone new. There's a lot of great people here. This community needs knowledge-sharing to solve so many of the problems or to see so many of the opportunities that we talked today. So connect with someone new.
WILL: I think that'll be very helpful. And plan to join us next year. So thank you all very much. And again, to our speakers, we really appreciate you all coming. Thank you. [APPLAUSE]
Form Title Download
Transcript
Form Title Bookmark
Entire Media
Segment(s)
Custom Clip
Start At:
End At:
Form Title Share
Entire Media
Segment(s)
Custom Clip
Start At:
End At:
Form Title Accessibility and Keyboard Shortcuts
The cadmore media player is a professional video and audio player experience designed to deliver media as well as relevant transcript, segmentation, and metadata. On the screen is a traditional html video player with buttons such as play, pause, forward, captioning, language and play speed selection and more. Adjacent to the video player are generally tabs, one of which shows a scrolling transcript of the video and the other shows any segmentation of the media, sometimes referred to as chapters. At the top of the video is a menu bar containing an Explore button and Tools button. The Explore button allows you to dig deeper into the video by opening up an explore dialog which contains tabs for a visual navigation of the video using thumbnails, an about tab which contains rich metadata about the media, a segments tab which contains further information about the video, and a related tab which contains links to other media related to this media. The tools button opens a menu with options for sharing the media on social media, creating bookmarking links, creating embed codes for your website, generating citations, and more.
Keyboard Shortcuts
Keyboard shortcuts only work while you are in forms mode. By entering forms mode, you will be able to quickly access portions of the media player with single key commands.
Spacebar: Play/Pause Toggle
I: Info Menu,
O: Tools Menu
T: Transcript Tab,
S: Segments Tab
F: Forward 15,
R: Rewind 15
L: Languages,
P: Playspeed
M: Mute Toggle,
V: Volume Bar
N: Video Only Toggle,
B: Full Screen Toggle
C: Captions Toggle
Escape: Closes dialogs and menus / Exits Player
Z: Next Transcript Chapter
X: Previous Transcript Chapter
Q: Next Transcript Paragraph
W: Previous Transcript Paragraph
Entire Media
Segment(s)
Custom Clip
Start At:
End At:
Embed Size
-
You Size - Responsive
-
We Size - Responsive
-
Picture Overlay Popup
-
400 x 225
-
512 x 288
-
400 x 700
-
768 x 1024
-
1024 x 512
-
1280 x 608
Form Title Cite
No citation provided.
Info
Segments
Related
Thumbnails
Exit Clip Mode
You have requested to see a portion of the video outside of the clip.
Click OK to switch to the full video or click if you would like to stay inside the clip. You can always go back to the clip with Tools and Clip Mode