Transcript
Transcript: Ian D. Shugart Visiting Scholar Lecture – AI and the Public Service: Moving at the Speed of Responsibility
[00:00:00 Text appears on screen: Welcome.]
[00:00:09 Text appears on screen: We would like to begin by acknowledging that this event is filmed on the traditional and unceded territory of the Algonquin Anishnabeg people. We encourage you to take a moment to reflect on the traditional Indigenous territory you occupy.]
[00:00:24 Text appears on screen: Ian D. Shugart Visiting Scholar Lecture 2026 – AI and the Public Service: Moving at the Speed of Responsibility.]
[00:00:30 Text appears on screen: Visiting Scholar Abdi Aidid Discussion with Taki Sarantakis.]
[00:00:43 The screen fades to Abdi Aidid.]
Abdi Aidid (Assistant Professor at the University of Toronto's Henry N.R. Jackman Faculty of Law): Hello, everyone. My name is Abdi Aidid. I'm an assistant professor at the University of Toronto's Henry N.R. Jackman Faculty of Law, as well as the Canada Research Chair in Artificial Intelligence and AI. This year, I've had the great privilege of serving as the Ian D. Shugart Visiting Scholar here at the Canada School of Public Service, which has meant that I've been, among other things, a bit of an in-house resource on all things AI.
And so, what I wanted to do today, is share a few thoughts about what I've learned over the past year, working with the federal public service and a bit of guidance from my position as to how we navigate this very difficult set of forthcoming changes safely. So, a couple of things before I get into it, which I thought maybe I could talk a bit about how I come by my interest in the topic. So, I usually wear two hats in this kind of conversation. So, on one hand, I'm a lawyer and a law professor, which means that I have commitments to things like justice and the rule of law, and I'm careful and risk-averse.
But as a legal educator, I'm also somebody who's thinking about the future of my particular profession and planning for my students to have an active place in it. I want my students to enter careers where they can bring their whole selves and be dynamic and creative and meet their needs, etc. And so, I'm thinking about AI not only as something that can help them, something that they can use to supercharge their ability, but also as a bit of a threat, if not properly managed.
I'm hoping that my students can enter a profession where they can sort of be professionally satisfied, which also means constraining AI to the extent that it might be labour displacing. But on the whole other side of things, I'm also a legal technologist. I've spent some years developing AI models, which among other things, were predicting how future courts were likely to rule in new legal situations. So, it'd be easy to look at that and say, Abdi, you are a bit of a contradiction.
And I actually resist that framing, in part because I have a particular view of what AI is and what it ought to be doing. So, the way I like to think about AI is that it's a technology that provides us with the capacity to do more good than we've ever done before, but also prospectively, a lot more harm than we've ever done before. In that way, it really depends on us to be interlocutors, to intervene where necessary, to constrain where necessary, and the concern today is that we haven't built up an ecosystem of constraints that could properly subordinate AI to our interests.
And so, part of what we're all doing in the federal public service and thinking about is how to build some of those constraints for ourselves to ensure that we're not allowing AI to run roughshod over us. That said, I remain convinced that AI is massively useful. And so, what I'd like to do today is talk a bit about ways that we can use AI effectively, subordinated to our interests, make it work for us without shortchanging our essential commitments to the public, because one thing that's different about the federal public servants and government workers in general is that we are supposed to have a lower error tolerance.
The things that we do and the things that we're involved in implicate people's life, liberty, property, security, and a lot of things that are very intimate but also potentially socially fraught. And so, it's really important to move carefully. There are plenty of institutions in the marketplace that don't have the same compunction that we do or that we should around adopting tools that could be speculative, and I think that's a positive thing and creates a real opportunity for the federal public service to lead on what responsible AI adoption might look like.
So, the first thing that I want to say is something that I think bears repeating, even though it might be obvious to anybody watching this, which is AI is absolutely here to stay, and what do I mean by that? Well, one, the economic case for AI is just too compelling. There are institutions the world over that are finding ways to economize and use technology efficiently to better achieve their goals. And once they've had a taste of that, there's no way to sort of un-ring that bell.
And I think that's an important starting point for people to recognize, that if you're having the conversation about whether AI is good or bad, you're probably having yesterday's conversation. The fact is that it's here to stay. And the question now is, how do we absorb it and use it responsibly? An important part of the story about AI being here to stay also is not misreading the goings on in the world of artificial intelligence, including things like problems of hallucination, errors, issues even that are frankly more serious, like bias and discrimination, as indications that AI is not progressing the way that we anticipate it's going to progress.
As a matter of fact, technology's trajectory is never linear. And as an example of this, just think about your experience with something as simple as Internet browsing. If you're somebody who was using the Internet to browse, maybe do some online shopping in the late 1990s, you were navigating a website that, at a minimum, probably cost a quarter million dollars for the developers to build. You had to hire an HTML expert, a designer. You had to pay exorbitant web hosting fees. These are all things that we don't have to do today because for a $9.99 subscription or if using AI tool possibly for free, you can have a serviceable website today.
Now, it took a long time to get here because a lot of people will remember that there was a messy interregnum in between. We were going on websites; I'm thinking of free blog hosting sites where you'd log on and they would play music, it would only play on one browser, you go to another browser, say Netscape Navigator, and suddenly the website was broken. But today, whether you're a talented computer scientist or a kid using a free public terminal in a library, you have pretty good browsing experience, but we had to navigate those choppy waters before the marketplace became coherent, and before we could guarantee quality at a low enough price that it started to permeate everybody's experience.
Think of AI a little bit like that. Right now, we have these fantastic models that are impressing a lot of people. We have at least 10 companies in the world that are building frontier models that look like they're approximating human intelligence. The question now is, how long before we get to a place where it's like our browsing experience where everybody can kind of understand how it works, where high quality can be guaranteed at a low enough price and where quality is consistent across the board?
Now, the thing that's certain is that in the meantime, we're going to be navigating choppy waters, which means some things that might look like us regressing, and it's important to remember that because the marketplace has to work out its contradictions. We have to decide what we want AI for, for example. If maybe some of you remember things like Pets.com, that was an exciting prospect, being able to buy pets on the Internet, until the public said, we actually don't want to do that. There was actually Internet crash. There was a boom-and-bust cycle, there was a bust, but would you deny today that the Internet suffuses basically every element of modern society? And that's even as many companies failed, even as the bubble bursts, so to speak.
And so, a bubble could burst. There could be a boom and bust. A bunch of the AI companies that you see today might not make it into our future. A lot of the tools that we're seeing proliferate today might not be the ones that we are stuck with when the dust settles. That doesn't mean we're not trending towards a place where there's consistently high-quality and widespread proliferation. So, if it's true that AI is here to stay, and if it's also true that in the long run, there's probably going to be even more widespread proliferation of AI, and that proliferation is going to mean good quality, low enough prices that the public can be using AI tools consistently, then what does that mean for you as a public servant? It means that the nature of your concerns, the nature of your objections even, to AI use, have to change. And so, my best advice to people is to invest in yourself as a credible objector.
There are plenty of reasons to resist the application of AI. I mentioned at the beginning that there are some areas that are very socially fraught, and if we point AI in that direction, it might not necessarily lead to good outcomes. And even taking it off of socially fraught circumstances, there are some areas for which we don't have good enough data to power the effective application of artificial intelligence. Well, have strong bases for resisting that don't take you out of the conversation altogether.
What do I mean by that? For example, you might resist the use of AI because you think that it could lead to outcomes that are socially undesirable in a specific context, but that's not a good basis for resisting the use of AI as back-office automation or using something as innocuous as a summarization tool. The challenge is that people are failing to disambiguate between types of AI and saying the entire enterprise is harmful. I think the position that we're in today demand different, more focused, more granular evaluations from us than the kind of objection which is categorical.
And the moment that you engage in a categorical objection, then what you've done is you've relegated yourself to a place outside the conversation where we're developing the constraints, where we're thinking about the effective use, where we're finding what the appropriate uses are in the public service. And so, even if you're somebody who is skeptical, concerned about the particular application of AI, even in your domain, the key, I think, is to get in the conversation in a way where people understand you not to be simply objecting to something that already seems overdetermined. Like I said, AI is not going anywhere. In fact, we're probably trending towards somewhere where it's even more widespread. And so, if that's the case, think about your resistance as having to contend with that social reality.
I think a related point to ensuring that your objections are credible and ensuring that you're not treating AI as a categorical thing to resist, is also recognizing that you can use AI differently in different contexts. And right now, the nature of the conversation around AI use in public service can be sometimes impoverished because people are failing to recognize that they can deploy AI much more strategically than a critical debate might suggest. The best example of this is chatbots.
A lot of people think of chatbots when they think about AI because a chatbot is kind of everyone's first touchpoint with artificial intelligence. Let's say, imagine you're a federal government agency that's using AI to respond to incoming questions. There's a couple of things you could do. One is, you can have the chatbot that just provides the first response to everything, or you could say, I understand AI well enough to know that this tool is particularly good for, say, wayfinding, it's really good at connecting people to information, but if I'm relying on it to extemporize or to provide on-the-fly advice, that could get me into a couple of different problems.
One, I might be backing into a situation where I'm relying on the AI not to hallucinate, even though the statistical incidents of hallucinations, as low as they might be, are something real that we have to take seriously, but the other problem is that you might be inadvertently limiting your own critical involvements in that kind of decision-making, and the point here, I think, is a crucial one. You can exist with artificial intelligence in a professional and intelligent division of labour. So, imagine a tool that directs people to the appropriate information, the AI doing the wayfinding, but when it identifies a situation where there's a sort of impermissible amount of discretion involved, it can kick that query up to somebody who's available, who's responsive, and who's benefited from having the upfront information articulated to them by the AI that has done the critical triage already. That's an example of an effective division of labour that you can't get to if you're thinking about the chatbot as the categorical device that is good or bad in the context of being responsive.
More to the point about the different uses of AI in particular contexts, I think it's important also to engage in the debate around AI with a real recognition of what you want human beings to do. Right now, there's an important conversation happening here about whether or not AI takes discretion away from us. And so, you've probably heard people say things like, we could all benefit from having a human in the loop. But in the same way that I'm encouraging everybody to recognize how we've departed from the conversation about whether AI is good or bad, we also, in a lot of ways, have departed from the conversation about whether humans are good or bad as they work with AI. A human in the loop, that idea is really our proxy for having non-technological discretion, some decision that could go either way, that we want to infuse with our principles and our best judgement. If that's the case, then there are some circumstances where the very fact of a human is not going to do it.
So, for example, if you're using an algorithm that is effectively sorting people, and you're asking an individual to do the quality control at the end of it and make sure it's accurate, there's a human in the loop but that human is not invested necessarily with the authority to intervene. In fact, all they're really doing is validating. In that kind of system, what happened is the discretion got moved upstream. The technology was making the discretionary decisions and the individuals really only empowered to resist in circumstance where the technology did not perform as to its specifications.
That's not true and honest discretion, but that has a human in the loop. And so, I think what we want to get to is a place where we're not just talking about humans or technology, but we're talking about what's best for the kind of decision set that we have, and what the right level of authority, judgement, discretion are to effectuate the kind of outcomes that the public expects from us. So, as you become more of a credible interlocutor, somebody who has facility with AI and recognizes that there are certain ways it can be used effectively and there are ways that you can, in practice, constrain some of its downside risk, you also need to become curious about the state of play in the world of AI.
Why? Because it's incumbent on us to anticipate where the changes are going. For example, when we first started thinking about using AI in federal public service, it was a lot of algorithmic tools, typically premised on machine-learning. And so, the way to think about machine-learning is it resolves a different set of problems than maybe today's AI tools do. So, think about machine-learning the way you think about your credit card company setting your credit limit. It'll ingest all this data from your history, your spending, your income, and will reconcile that against people who have similar spending habits or similar income and try to predict what the most amount of money they can front you is without you running a significant risk of default. They use machine-learning for that because they have a set of historical data that can help them identify and make predictions.
So, if you think back to 12th grade math, for example, that's like finding a line of best fit in a scatterplot, just to make it more boring. Now, the thing about that is that it's limited by the data that you supply it. So, I like to liken it to, if you give a sculptor a block of ice, they can carve a beautiful swan, but it'll never be larger than the block of ice that you give them. So, that can mean a few things. If there's flaws in the block of ice, there may well be flaws in the ultimate sculpture. So, if there's things like a history of inequity in the underlying data, then using machine-learning tools, you run the risk of projecting that forward. Compare that to where we got to once you started seeing ChatGPT proliferate, which is generative AI.
Generative AI is not like providing a block of ice to a sculptor and them only being able to create a sculpture that's the size of the block of ice at most. It's like giving a sculptor a block of ice, them observing the block of ice, learning about the properties of freezing water, and then making another block of ice and then making two ice sculptures, the idea of taking data to create net new information. In government, we were initially doing the machine-learning, which meant a different risk profile. It meant that the universe of what was possible was circumscribed by what existed in the data. That's not the only world we're in today.
Today, we can take that information, learn things from it, glean insights from it to create new work product, to do new higher-order tasks because of the generative revolution. Now, that one caught us unaware, but it turns out the next one is giving us a lot of notice, and that's the agentic AI revolution. So, imagine taking all those capabilities from machine-learning, all the capabilities from generative AI, and now investing them in an AI-enabled persona. That's agentic AI, that's the agent, and the thing that's unique about the agent is it can engage in what you call multi-step transactions without supervision at any one constituent step. So, the less boring way to think about that is, imagine calling tech support 10 years ago and the person on the other side of the line says, hey, is it okay if I take over your desktop? And when you see a cursor move around your computer, imagine an artificial intelligence tool being able to do that, being able to perform tasks on your behalf.
That's a different risk profile. Now, that's also where most of the R&D and most of the creative energy is in the world of AI. So, you're effectively as a public servant being put on notice that we're shifting from a world where AI was limited and circumscribed by the data, a world where AI could generate net new information to one now where it's performing specific tasks with light touch supervision. The question for you is, how do I use that and harness that effectively?
Now, at some point, as we go down the path of agentic AI, we're going to be put on notice about what the next newfangled AI revolution is going to be. And so, it's important for you to engage in that conversation early. That way, we're not caught unaware. Now, if you don't get in the game with AI, then you might be, as I mentioned, precluded from participating in the public service of the future. But I think even more importantly, what's at risk is not being able to participate in the upside of what a technology as public service might look like. A couple of examples here are worth thinking about. Think about all the areas of our social and public life where we generate a lot of data and a lot of information.
So, think about traffic, for example. Every day, there are cars going up and down the street, people making left turns and right turns. One way to think about that is a bunch of activity happening out in the world. Another way to think about that is it's important data for us to better understand how we might want to route streets, how we might want to think through what the effect of public transit allotment might be, how we might think about environmental harm and risk, and how we might think about public and pedestrian safety, for example.
All of those are data points that can inform effective decision-making that today would involve a little more guesswork than we would like to admit. And so, even if you're somebody who says, I have good judgement, I'm an effective policy professional, I'm good at developing effective social programs, then why wouldn't you want to start off on a higher floor with more inputs, more data, more information by which to make those decisions? That's what AI can do at its best if, as I mentioned, you exist with it in an effective and intelligent division of labour.
As I conclude, I want to leave you with a couple of parting thoughts about how you might find your place in tomorrow's AI future. So, the first thing I want to do is acknowledge that some of this does sound scary. It sounds scary because often the technology is positioned as a thing that's here to replace you. Other times, it sounds scary because you suddenly feel like it's out of your hands. You used to understand very well how things would go. Today, it sounds like there's a bunch of zeros and ones that are being crunched behind some technology that you don't fully understand. I want to say that's not going to be the case forever. Part of the reason that we're obsessed with the underlying technology of AI is because we don't trust it. Compare that to technologies that you're very familiar with in your life.
Think about your computer, your laptop and the keyboard. I would suggest that most people are fundamentally incurious about how their keyboard works. You know that somehow there's a signal being sent to a motherboard and it's leading to the cursor moving a millimeter on your word processor, whatever, you don't really care. Part of why you don't care is because you've absorbed it socially. You understand that you can call a hotline and get tech support if something goes wrong with your keyboard. You also know that you can replace your keyboard in the event that it really is broken. But critically, you have already been able to sort of score what risk of broken keyboard poses to you in your life, and you decided it's not that significant.
That kind of future is coming in the context of AI when we talk less about the underlying technology and more about what we do when something does go awry, how we evaluate the risk that it does pose for us, and you'll see that we'll be talking and speaking in a way that's much more outcome-oriented. Today, we're obsessed with the process, but we're obsessed the process because we don't trust it. We want to look under the hood because we don't believe what's happening there is good for us. If we get to a place where we've seen enough examples of the technology being used effectively and we've seen survivable examples of the technology misfiring, my suggestion is that we're actually not going to be talking about the technology of it all for that much longer. Instead, we'll be talking about the things that are already in your wheelhouse.
So, for example, how do you get people to use AI effectively? That's a question of personnel management. Today, it's a question of understanding the underlying technical capabilities. But tomorrow, it's a question about how to get the best out of colleagues. If we're talking about ways in which you might develop a social or public program that uses AI effectively, today, that's about how do I coexist with AI? How do I subordinate it to my interests? Tomorrow, that conversation is just going to be about, was the program good or not? In part because the technology is going to become more and more invisible to us, so long as we identify the right constraints, and the key thing is, so long as we step forward in a way that's unintimidated.
As a final word, I want to thank everybody in the federal public service this year for welcoming me and for giving me the opportunity to share some of my insights. I know it's a challenging time. Not only are there technological uncertainties, but in many ways, the world seems to be upside down, and I think there's no place more critical to be than at the forefront of protecting people's rights, ensuring that their needs are met, and finding ways for them to live safely and securely, and that's what the federal public service does, and I continue to believe that the Canadian federal public service can model what good AI stewardship looks like. And so, I look forward to continuing that conversation with everybody. Thank you.
[00:24:42 Taki Sarantakis is shown sitting next to Abdi Aidid.]
Taki Sarantakis (President of the Canada School of Public Service): Welcome back through the magic of television, so to speak. The lecture has concluded. And now, we are doing our kind of a post-game wrap up, Abdi. So, Abdi, I want to start with telling you how much I enjoyed your lecture, mostly for the following reason. If I understood you correctly, I think you basically gave us an exhortation to the effect of, it's not if AI, it's kind of how AI. Is that a good way to summarize kind of what you've told us?
Abdi Aidid: I think that's exactly right. I think we have been sold the illusion of choice right now, as if we can meaningfully resist. Whether you like AI or not, it's a social fact you have to contend with. And if you think it's a bad thing, you have to figure out ways to use it in manners that are good. If you think it's a great thing, well, you need to figure out how to properly constrain any of its downsides while appreciating all of its upsides. And so, I think that the truth of the matter is that for a given individual, particularly a given public servant, very often what you're doing is now responding to a mandate as opposed to deciding whether or not to embrace. And I think the sooner that we can agree on that, the quicker we can move on to the second order of conversation, which is the how of it all.
Taki Sarantakis: Exactly. Now, you mentioned that there were kind of two limiting factors. One was the notion of losing control, which I think we've all felt, and the second is the notion of trust. I want to talk a little bit about control because we offload so much of our lives to technology already and technology that has, in most respects, absolutely nothing to do with AI. The example that I give constantly to my friends is, do you remember the very first time your ABS system kicked in? And you were approaching a stop light or a stop sign. And all of a sudden, you're pressing the brake and nothing happened because the machine was kind of saying, step aside, human, I got this, I'm better at this than you, and the first time that kind of happened, you went, how strange. However, you didn't hit something because the machine indeed was better than you. And then, it's not like the rest of us right after that moment went off and disabled our ABS. So, talk to us a little bit about the loss of control and why for many people, is AI a fundamentally different loss of control or is AI just the latest variant of things that we feel we might be losing control over?
Abdi Aidid: I hate to do this but it's kind of both, but I'll explain why. So, it's different because more than any other technology that we've encountered so far, it purports to do a greater share of the thing that we think of ourselves as doing. So, in that way, it's true, honest automation, and that is frightening because especially if you're someone who's so invested in your output as part of your self image, it can be particularly frightening in that way. And so, I want to acknowledge that it is different in those ways. The thing we need to, I think, acknowledge, is that there are certain kinds of control that are worth giving up. So, for example, the example you gave about ABS, it's actually a good one because the advantage that the technology has over us is that it's better at passive anomaly detection than we are. So, think about it like this. You can think about the pop cultural example of the security guard that falls asleep during the burglary. At the critical moment that you need to be awake, you're nowhere to be found.
Taki Sarantakis: Watching a video screen.
Abdi Aidid: Watching a video screen, but also maybe if you're trying to be attentive, you just have a harder time doing so, having sustained attention over a long period. Think of yourself trying to watch a lecture. After an hour, you start to trail off.
Taki Sarantakis: After an hour? Most humans, after nine seconds.
Abdi Aidid: After nine seconds. Yeah, well, technology has no such limitation in the sense that it doesn't need coffee. It doesn't need reassurance. It doesn't need a good night's sleep to be able to spot issues really quickly. And so, maybe that's the role it can play. The passive anomaly detector that provides a buffer against our occasional attention deficit. That's an example of the kind of control where if we were reasonably appraised of what technology could accomplish, we would say, that's a good example of how we should use it. The bigger concern I just want to say is when people start to see the technology get competitive with them, so less about the ABS example and more about, imagine taking your hands off the wheel and then doing all the steering inputs and you can see the wheel turning and making smooth turns and maybe even making judgements about the fastest route. All that is technically possible today, but it would be more disconcerting because that's the thing we imagine ourselves contributing to the economy.
Taki Sarantakis: Because none of us know a bad driver. You never hear somebody say, I'm a terrible driver. Everybody's the best driver in the world.
Abdi Aidid: Everyone is the best driver. It's like how you never know anybody… there's like a million comments on a YouTube video, but no one will admit to being a YouTube commenter, for example. You don't know anyone who's having arguments under videos. I think it's a cartoonish way of making a point that I think is important here, which is that we're not actually that good at gauging what we're expert at relative to technology.
Taki Sarantakis: Yeah. Second is trust. So, you talked about control and trust. Talk to us a little bit about why some of us would hesitate to trust AI.
Abdi Aidid: I mean, the first thing, which I'll take off the table because it's so obvious, is the newness. The unfamiliarity leads to some distrust. I think the key one right now is less the unfamiliarity because we've actually embraced unfamiliar technologies when we find them to be useful. I think it has to do with how it feels competitive with us right now.
Taki Sarantakis: So, is it something, and I think if you go back in history, fundamentally, everybody would say, the radio will fundamentally change who we are, the television, electricity, and all those things were true, but how is this particular one different from electricity or the radio or television or the Internet?
Abdi Aidid: So, all of those technologies were doing things that would, even to a skeptical person, enable human flourishing.
Taki Sarantakis: They were augmentations of the human.
Abdi Aidid: They were augmentations and they were like facilitative. So, brighter lights more consistently applied over the course of a day mean I can do more things in my home. But today, if we're saying that the things that you're going to do in your home can be automated, the coffee, the parenting, the cleaning, all of that stuff, then you start to say, well, what's the point of me now? I can't countenance, I can't stomach a world where there's no point of me, and I think it's the combination of catastrophic thinking where you imagine that all this can do is derail you, but the other piece of it is where we take a kind of a dim view of ourselves and what we have to contribute. So, think about professionals. There are a lot of professionals that are concerned. I'll take lawyers as an example just to make it about us.
A lot of lawyers are concerned about the prospect of AI-enabled research, because they say, well, what's that going to mean for me if the AI is doing the work of curating the information? Well, if you conceive of yourself as an information retrieval professional, then yes, AI is threatening to you. But if you imagine yourself as somebody that can do things with that information, creative judgement, client service, but that you can actually provide additional information, reconcile the information that technology is giving you with your understanding of the local context or the judge's disposition or the weird contours of the law, then you really have something. And so, I think very often, we take a bit of a dim view of ourselves, but it masquerades as us thinking we're better than the AI. We are afraid of AI because we think that it does a growing share of what we're capable of. As a matter of fact, it does a bunch of things which, if those were just a list of your qualities, you wouldn't be that interesting.
Taki Sarantakis: So, I'm hearing two things from you, and I'd like you to comment on each of them in turn. The first thing that I'm hearing is that in some way, AI threatens our, let's call it, identity. Let's not go all the way to humanity because it's an even bigger word, but it threatens our identity in a way that kind of electricity and radio and television and the Internet didn't threaten our identity. And the second thing I'm hearing from you is that it's not so much about the thing that's scary is the AI per se, it's that the AI is taking away a value add that we associate with us.
Abdi Aidid: Yeah.
Taki Sarantakis: Identity.
Abdi Aidid: I think that's exactly right, but I think professional identities need to shift. I think we need to think less about the tasks that we do and more about the value that we deliver. So, I gave you the example in our conversation some time ago about graphic designers right now, as an example of a professional that's learning a bit of a hard lesson, a cruel lesson, actually. So, if you're a graphic designer, you're an artist, you're somebody who has knowledge of design, aesthetics, you're able to hear the motivations of your client, translate them into this visual identity, and that's a creative act of brilliance, and the best of the graphic designers also have good business judgement and all these things. And so, the value of their services are priced accordingly. And then, technology enters the picture. And now, suddenly, people can self-remedy their need for a logo. And it turns out that all along, what the public wanted were serviceable, unpixellated logos. Now, there are some people who want to procure expensive graphic design services.
Taki Sarantakis: If you're Starbucks, if you're Costco.
Abdi Aidid: For a multimillion-dollar brand, if you're a public institution, if you're a university. But in terms of the sort of consumer-facing kind of graphic design services, it turns out people have lower standards than the profession has for themselves, and that disjuncture between what it turns out that people want and how you imagine your services being packaged and bundled is teaching people a cruel lesson. Lawyers, for example, are a great example of this. Lawyers think of themselves as philosopher kings or something. It's like every contract I draft is pushing the boundaries of human knowledge, but all people want in most circumstances, and especially garden-variety legal representation, is an enforceable agreement.
Taki Sarantakis: I want my title transfer, I want my will to be executable, and generally speaking, I want those things because a lot of them are commodities, I don't want them to be an artistic expression, I want them to be a functional utilitarian thing out the door.
Abdi Aidid: That's predictable, for example. And the thing is, compare that to, say, just to keep the professional theme going, compare that to doctors. So, if you're an oncologist, or better yet, a radiologist, you've been competing with AI for a long time, a generation at least, and it turns out that AI is really effective at identifying cancers, lesions, and radiological scans, and on some studies, as effective or more effective than the oncologists, but there isn't a single person that wants a cancer diagnosis from an AI. So, they see an independent value of the medical professional that legal professionals, for example, haven't been able to articulate. And so, the challenge that we need to embrace is, can we convince the public of our value at this stage, and not an invented value, but an actual value that was always there that maybe we didn't sufficiently attend to?
Taki Sarantakis: Yeah. Now, the first visiting scholar that we had at the Canada School was Dr. Rachel Zellars, and one of the things that she taught us, that on the face of it, kind of you go, that's bad. But actually, if you dig into it, it's actually a profound insight. She said that the human brain is a bias machine. We are instantly conveying biases constantly because in a way, that's kind of how we survive. We have a bias towards safety, we have a bias away from unfamiliar noises, etc., etc., etc., yet one of kind of the big criticisms of AI from a lot of people is that it's biased. It seems to me that it's easier to un-bias an algorithm than it is to un-bias a human being. Talk to us a little bit about that paradigm.
Abdi Aidid: Yeah. So, I think it's right that we're biased machines. Dr. Zellars focused on the kinds of snap judgements and decisions that we make, but we have an overarching bias, which is the bias against exhaustive information analysis. So, we take certain cognitive shortcuts because we can't hope to process all the information involved in making a decision.
Taki Sarantakis: It's a lot of energy to go through.
Abdi Aidid: Totally, totally. And so, if you're, say, a policy analyst and you're trying to make a recommendation as to what a policy should be, you might start by looking at some of the empirical literature or the scholarly literature on the issue. What's an effective gun violence intervention? Well, go to Google Scholar and you'll find 10,000 peer-reviewed pieces about effective gun violence interventions. Now, the limiting factor on the quality of your advice is how much of that can you ingest in a day?
Taki Sarantakis: Right.
Abdi Aidid: And so, you already have some limitations.
Taki Sarantakis: So, you default to your bias.
Abdi Aidid: So, either you default to your bias, or you do things like sampling. But the problem is that when you make determinations about what to read, you're already then infusing it. That's being cut against your experience, your understanding of what's salient, your sense of how the rest of your day is going to look. And so, you're infusing bias even in the information sorting, and that's before we get to your childhood, your traumas, how you feel about other people.
Taki Sarantakis: So, I'm biased as to what of the 10,000 things, what am I going to look at in that day? What am I going to look at seriously? What am I going to glance at? What am I going to completely disregard?
Abdi Aidid: All of that are opportunities for bias to happen even before we get to the kinds of bias that we talk about socially, which relates to pre-determined thinking or prejudice.
Taki Sarantakis: Where you grew up, who your parents were, or what your friend network is.
Abdi Aidid: Yeah, and so, AI might share some of those negative social biases with us to extend a sense reflective of ours, but what it doesn't share is that information-condensing bias that we have.
Taki Sarantakis: So, the only thing I really worry about with AI vis-à-vis the bias question, is that we might lock in some current biases because right now, we are actually training the AI on available data, and a lot of that because of humans is biased.
Abdi Aidid: Yeah.
Taki Sarantakis: So, other than that, I don't think that AI will be more or less biased. In fact, I think long term, AI will be less biased than humans.
Abdi Aidid: I think that's a fair thing to say, but I want to flag a different way to think about this. So, if we're using AI to do things like criminal sentencing, yes, that's going to be reflective of the underlying data, the history of criminal sentencing. We know sociologically that criminal sentencing is inequitable, particularly cross-racial categories. So, if we're throwing that in the hopper for the algorithm, it's going to project that negative history forward. We know that.
Taki Sarantakis: But I just want to make a quick distinction here because this is important. It seems to me; that's not a machine problem. That's a human problem that's been put into a different domain, which is the machine domain.
Abdi Aidid: Totally, totally. So, the reason I say we should think about this a little bit differently is, yes, that's holding up the predictive mirror to us. We know that, but there are other kind of biases that that creates room for, one you might call scientism, people thinking that because it's data intensive or it's analytically rigorous…
Taki Sarantakis: You got to trust the machine.
Abdi Aidid: Even if you don't trust the machine, the machines do something more sophisticated than I'm doing so I'd rather default to that kind of inequitable prediction.
Taki Sarantakis: There's a wonderful kind of cartoon where there's two people standing in the pouring rain and somebody looks at their phone and says, the machine says it's not raining, it has to be right, it's not raining even though we're both wet.
Abdi Aidid: Exactly. And so, that I think is a real risk for us, and part of that has to do with our evolving relationship to technology. We distrust it, but we don't distrust it because we think it's less capable. We distrust it in part because we wonder if we're as capable. And so, what happens in that? It's like any insecurity. You basically are like well, it's powerful, it's powerful in a way that I'm not. So, that's one risk. The other, I think, is maybe even more critical, how we think about bias with these kinds of tools. There is, of course, the bias of what happens in the tool, what they end up producing. But then, there's the bias that's a deeply human enterprise of choosing where to point the AI gun, so to speak. If AI is this powerful tool, our selection of where to direct the AI sometimes determines whether we're going to have good outcomes.
Taki Sarantakis: Exactly. What's this called? Domain deployment. Are you going to deploy it here or here or here?
Abdi Aidid: And sometimes, the negative outcomes are less a function of stuff lurking in the data, and maybe it's sometimes more of a function of, this area is not ready. This area is not ready. And so, I'm thinking about the public service here. Having to do things like a real reckoning with the quality of your data before you layer a bunch of AI on top of it, that's like building a house on quicksand or something.
Taki Sarantakis: Yeah, it's actually one of the things that will slow down adoptions in many, not just the Government of Canada but in many government domains, forever, which is bizarre. The governments, they're created on the notion of data registries. Who are you?
Abdi Aidid: Totally.
Taki Sarantakis: What piece of land do you own? How old are you before I can give you your pension? How old are you before I can give you a driver's licence? The very state was built on data, yet we haven't taken data as seriously as the private sector has.
Abdi Aidid: I think part of it is a failure to anticipate that that data would be useful in a different way.
Taki Sarantakis: Yeah.
Abdi Aidid: We imagined the data being the thing we had to retrieve. We never imagined we'd have to compute with it or perform functions on top of it. And so, now that we know that, again, as a theme of the lecture, we're being put on notice that from now on, it's not just about going and fixing the existing data so that it's readable by machine. It's about all future data collection ought to be amenable to computation and function, which means things like revisiting forms, doing some of that upfront work that ends up being a prerequisite for the complicated AI that is going to be increasingly part of our future. And as it relates to the data quality and data maturity, I think there's an important thing for us to remember here about data. Performing a functional type of data is about one use of AI, and we need to disambiguate too. That's not going to stop you from using AI to summarize the literature that is informing your policy memo or AI for drafting a component part (inaudible).
Taki Sarantakis: People cut and paste from Google, which we've been doing since 1998.
Abdi Aidid: And so, what you can't do is be like, the data architecture is insufficient, to stop you from your little office AI corner solution. Still use AI in the context in which the data is not the problem.
Taki Sarantakis: Exactly.
Abdi Aidid: That's the key.
Taki Sarantakis: Now, we're kind of fumbling and stumbling with this in the public sector and we will for a long time. It's kind of natural and normal, although with the rapidity of change, it's actually scary how long it takes us to fumble and stumble on a technology, but we've done it in the past with television and radio and electricity and cars and the Internet and social media. And in some of those cases, we're still fumbling and stumbling 30 years later. Another area or another domain that is living this in real time is your domain, education.
Abdi Aidid: Yeah.
Taki Sarantakis: How do you deal with AI in your classroom? How does your law faculty deal with it? How does the University of Toronto deal with it?
Abdi Aidid: Yeah, so the University of Toronto, like every institution, is grappling with the reality of AI and is coming up with brokered solutions and negotiated solutions to the problem. I think the most critical thing we're not doing is hiding our head in the sand and saying that AI is not…
Taki Sarantakis: AI bad or AI good.
Abdi Aidid: Or that we can categorically ban AI use of any kind. I think, again, the theme is disentangling use of AI. I think what we don't want is for people to use AI for functions for which we need to evaluate their retention of information, their judgement, their analysis, but using AI to prepare for all those things, for example, for that ultimate evaluation might not be as odious as using it on the evaluation. And so, even the maturity of even separating those things out, you're starting to see right now, my personal philosophy about this is that we need to think about taking stock of what skills people need to learn. And if those skills are best learned in an analog way, then we should jealously guard the analog method.
I have constantly been giving people an example of the calculator. It's the easiest example to think about. My daughter is seven. She's never used a calculator. The reason why is because we said we are absolutely positively certain that young kids will be better off if they first have conceptual arithmetic knowledge. In fact, they'll be better able to manipulate the technology if they have conceptual arithmetic knowledge. The question for us is, how many things in our education system are we as sure about as conceptual arithmetic knowledge? And once we come up with that list, for those, let's restrict AI. Then, let's think about thresholding.
So, my Juris Doctorates, my JD students, it's a three-year program. Maybe the case is that you throttle AI use early enough, but you allow them to use it at a later stage once they've done an initial showing of the conceptual foundation. The reason those conversations get shied away from, not us but others shy away from those conversations, is because that work of identifying what the conceptual foundation is, is a hard task and there's not a ton of agreement on it.
Taki Sarantakis: Now, this is our penultimate question. I hate doing this, but I'm going to go philosophical, because the reason why I hate doing this, going philosophical, is it seems to me that we have to get granular, we have to get on the ground, and a lot of us vis-à-vis AI, we're acting like we're Socrates and Plato, and it's like, yes, we can think about that. But often, that's being used as an excuse for us actually not grappling with the thing. So, I'm going to be philosophical in our penultimate question. You and I are of a certain age. It seems to me that AI is coming in at a perfect time for us. We already have jobs. We already have kind of grown-up learning how to read. We already have grown up learning how to research, how to synthesize, how to understand the basics of mathematics. So, it seems to me that for kind of our rough age cohort, this is almost a panacea. This is a tool that you take the little Internet, and it just blows away the little Internet as a tool, but you and I are also professionals who deal with the generation behind us. Then, you and I are also parents, as you just talked about your seven-year-old daughter. What are we philosophically going to be doing vis-à-vis raising our children, vis-à-vis the next articling student, and vis-à-vis the next person to enter into your law class as kind of an incoming law student?
Abdi Aidid: I'm so excited to answer this question because I get to use another analogy. And as an educator, I love using analogies. So, first of all, just because technology is capable doesn't mean we're going to accept using it in a bunch of different contexts, nor does it mean that we're going to trend towards the most technologized version of something. I think very often about, you go to a public bathroom and sometimes you're still turning knobs, and you're like, ugh, didn't we figure out the automated technology for this 40 years ago? We did. Still, most of the time, we're turning knobs, and part of that has to do with there are other things that are dispositive or help us to determine how technology is going to proliferate. And sometimes, they're unrelated to our technical capacity. So, I don't expect AI to be deployed everywhere. I simply don't. Part of it is going to be, there could be cost barriers for it to be deployed in certain settings, others are, we're going to decide that there's not a particular context in which we want it, and there's a whole other category of things for which we have a positive, affirmative vision of what we want to accomplish and AI would undermine it.
So, think about this. We have devices to lift heavy things. Would you use those in the gym if you're weightlifting? No, because the social, physical, economic, biological value of lifting the weight yourself is so clearly articulated to you, and to the point you feel it in your bones. You'd be like, this would be a waste of my time, and it wouldn't get me where I'm going. And so, I think in the world of education, we need to make the same kind of arguments they made in the world of fitness and in certain health contexts where it's so understood that we forged a consensus around what constitutes the right way to do this, and that's a public conversation we all need to participate in, but the thing that we can't do is be skittish about the AI overall because that conversation can't be had unless we actually reckon with what AI's capabilities are and then also people's incentives to skirt the process.
Taki Sarantakis: Absolutely. So, our last question. So, as part of your time as a visiting scholar at the Canada School of Public Service, you spent a lot of time interacting with other arms of the government, different management tables, different departments here at the School. Given that you have kind of visited the zoo, so to speak, as a visitor, tell us a little bit about what you're taking away from this experience in terms of insights vis-à-vis what the animals in the zoo, and I'm one of them, look like and think like and act like. And then, maybe close us off with a bit of kind of scholarly wisdom and practitioner wisdom vis-à-vis near-term practical. What should we be doing vis-à-vis AI?
Abdi Aidid: Yeah. So, the public service, I think, is rightly concerned about what AI is going to mean both in terms of getting the work done, so there's concern about error and hallucination and bias and discrimination, then there's a concern about labour displacement, which is what it's all going to mean for me. I think we need to embrace those concerns and lean into them. Why? Because it gives us an opportunity to model effective AI use. There are folks in the private sector that have none of the compunction that the public service has about using and embracing AI. If you're somebody who used to make online catalogs or something or you're an e-commerce website, and now you can use AI to write the product descriptions, you're leaping at that opportunity. Compare that to your mode of public servants, which is like the drafting is core to what I do.
And we have to ask ourselves, why is there that disjuncture? Why do we react so differently in the public service to your model private sector context? And I suspect that it's a mishmash of reasons. Some of it is that we think that what we do has kind of this gravity to it, that we can't over-delegate to speculative technologies. Some of it is the self-preservation. I think the hard work right now is going to be pulling those apart. For example, rather than have that be this globular mass of anxiety, you could actually use that to guide you as to how you should apply AI. For example, what are the things that actually are about me not wanting to delegate something because there's a public interest in a human being doing it? That's the one that you don't deploy the AI in the context of.
Taki Sarantakis: I'm going to guard that jealously.
Abdi Aidid: But pull it apart. If it's just the equivalent of a product description in the catalog, then you have to ask a different question, not can AI do it, but if I really care about the public, am I shortchanging the public by not having AI do it? I think that sort of cultural and cognitive shift is what you need. It's about pulling this stuff apart right now. It's been presented to you as this thing bundled, this bundled basket of goods, pull it apart, and that's what I think we need to be doing in terms of some scholarly wisdom…
Taki Sarantakis: And practical too.
Abdi Aidid: And practical. To end, to share with people, I want to say, every conversation about AI feels too late. But in the grand scheme, we're not at day zero. Maybe we're at day one, and here's how you know that. If we tomorrow just said, let's automate everything we do, I'm going to be the automated professor, the automated lawyer, there actually wouldn't be enough tools for me to fully automate what I do end-to-end, let alone good ones. And so, we're not at the place yet where there's replacement-level technology, nor are we at the place yet where we know exactly how the tech is going to settle. The market's going to declare some losers. They're going to say there's certain things we don't want, there are certain things that are an unjustifiable use of resources.
Taki Sarantakis: The Pets.com.
Abdi Aidid: The Pets.com example. And then, Pets.com might come back around many years later. Today, you can buy a pet online because we said that we have now the mechanisms, the ecosystem, the laws, the regulations, the social awareness to be able to transact for pets that way online now. 30 years ago, we said we weren't ready. We don't know exactly where this is going to settle. So, be a keen observer. But also, don't overlearn today's lessons by saying, the way that the landscape looks right now is how it is going to look in 20 years. Actually, for all we know, there's going to be more AI, that's for certain, but there might be less of this kind of AI, maybe less silly consumer AI.
Taki Sarantakis: Right.
Abdi Aidid: There might be a lot less retail-focused AI. There might be less AI games. There might be less AI toys. There might be more AI infrastructure.
Taki Sarantakis: Right, back-office enterprise AI.
Abdi Aidid: Back-office enterprise. And so, hold open the possibility that we are actively in plain view working out the kinks.
Taki Sarantakis: Absolutely. Professor Abdi Aidid, thank you so much, not just for this conversation, for your lecture, but most specifically and most importantly, thank you for spending, I don't know what it was, six months with us?
Abdi Aidid: It's been longer than that. It's been more like nine so far, and I'll still be here for a couple more months.
Taki Sarantakis: Yeah, and you have done what I love doing, what I see as my job at the Canada School of Public Service, is you have disturbed mental sediment when you're discussing with people, and that's really what learning is all about. It's taking something that you thought was a given and understood and throwing it all up in the air and reconfiguring in a different way.
Abdi Aidid: Well, thank you. It's been one of the great honours of my life and I hope we can continue.
Taki Sarantakis: Thank you so much.
[00:59:13 The CSPS logo appears on screen, then the URL canada.ca/school.]
[00:59:18 The Government of Canada logo appears on screen.]
