MCLL Symposium - Q&A with Paul Yachnin and Andrew Piper

Yes. We talked a little bit about kind of the downside of some of this machine-driven prediction. And I guess maybe to play on this app idea a little bit, I mean, you can imagine going a step further and having the app based on your past conversations and writings and readings already know what you want to hear, you know, and just put the rest to a background hum. You know, certain presidential folk are speaking and just sort of goes to static to save you the trouble of listening to stuff that you're already gonna dismiss. You know, which I'm saying a bit facetiously, but not exactly, because even Amazon deciding with its machine learning, you know, that I like this book and I'll probably like these other books is already— it's already kind of channeling me in a way that I'm not sure I want and that might restrict, you know, all of our sort of transformative or creative or tangential experiences. Going to a library was interesting because you could look for the book you were looking for, and then you'd see the book next to it and go down a totally different direction. And we're kind of losing that with this AI stuff.

Could you imagine the app that is growing and growing also feeding back to the teacher saying, the app would say, you're losing them now. And I think that would be a valuable feature, to have the kind of feedback loop.

Excuse me.

Every now and then you bring up the topic of ethics and putting ethics into computers. Can you expand on that?

A little bit.

And whose ethics would it be?

That's a great and difficult question. Where are you? I didn't— I beg your pardon. That's a great question. I think that my sense of it is that it should never be the machine itself. And that is a task that I think the university can take on. that the way we use computational AI is modeled on the way we've used literary AI. And that's ethical in 2 senses of the word. It's ethical because it also contains the ethos of a particular culture that the human participant comes from. And it's ethical also because it's always based in dialogue. And the project that I completed my remarks by describing is, seems to me, an extraordinarily good ethical use of new technology to bring silenced voices from the past to the 21st century to help us understand where our ideas about ability and disability come from. So, of course, I'm coming at this from a humanities point of view and thinking also always in terms of social justice and the ways in which the humanities can contribute to the advancement of social justice. But I think that the ethical requires that dialogue between human and machine.

And just really fast on top of that is, I think one of the key things of an exercise like this shows you is one way of thinking about ethics is they're consensual norms that we agree to. And right now, what I think the frustration around AI is, is there are no consensual norms. behind the algorithms. The Facebook or Amazon algorithm really serves Amazon's needs, not necessarily my needs. If you were in charge of your hearing app, that's you determining what its inputs and outputs are and its values that are encoded in it. That's very different from Amazon deciding what kind of hearing app you should get. And I think that really is one of the fundamental distinctions, is community ownership. over the algorithm, over the decision-making process, that's where the ethics is built in, right? The participation in design reflects your norms and your expectations. That would be how I think about ethics entering into the process that's really been missing. These are black boxes held away from communities of use, and if you flip that around, you would get a very different landscape, I think.

I think I have a question on a related topic. You mentioned at the beginning that— or someone mentioned that machines can think, but they don't have thoughtfulness. And I'm wondering about that. If you look at wisdom as an algorithm, I mean, what is wisdom? I mean, you can define wisdom, but it is a decision-making process, a decision that is judged by communities, by individuals to meet certain criteria. I would suggest that what is wise is what is true and what is important, but that's mine. But I think wisdom could be programmed. I'd like your comments.

I think that's a really— I picked up the use of the word wisdom earlier in the day, and it seems to me a word that we need to actually think more about. It tends to be bracketed. We use it, but we put it in brackets. I just wrote an essay called How the Merchant of Venice and We Grew Wise Together, which tracks the history of the play's performance and criticism from the beginning to the 21st century. And the wisdom that I'm now saying is kind of attached, is accumulated around the play, is this complex, long-term, Experience of the language, experience of the action when the play is performed, and very, very different views that have been in dialogue and contest over a period of hundreds of years. So it seems to me that wisdom is always going to be based in experience. How can— and I don't know what the answer to this question is, I just put it on the table— how can the new technologies contribute to the development of wisdom. And I think it's certainly something that we might consider as an important goal of a university education, the pathways that lead to this thing that we call wisdom. We don't really know exactly what that means, what wisdom means.

Good morning.

There's a question here.

Are the AI researchers at McGill going to get all this information from our program?

Definitely. I know who to talk to. Yeah, no, I mean, I was so excited to participate in today's event for precisely the reason of listening, that it really— the field is driven by some intrinsic questions, intrinsic goals, intrinsic needs. And I think really the next stage is about listening better to different kinds of communities and finding out what their needs are and beginning to engineer solutions that are different across groups. I just— I think We've, the process has been led by certain actors with certain needs and focusing on certain communities, and I think it's really about widening up that space and also having ownership over that process. People really worry about algorithms making decisions independently of human beings. Teacher performance, for example, right? We potentially know the story about you can design an algorithm to assess how well we teach. Now the problem is, what exactly are the criteria that that judgment should be based on? Well, that's a very complicated question. It certainly should involve all of the different stakeholders involved in the process to agree on that, and it certainly shouldn't be made once. You would never want to design a form of artificial intelligence that finishes at a certain date. Wisdom is accumulated knowledge. It's something that evolves over time. So if you're thinking about something like teacher performance, you want to make sure teachers are involved in that discussion.

Yeah.

You also want to make sure students are involved in that discussion. You want to make sure administrators or the public at large that is paying the public salaries are involved in that discussion. And everybody is constantly arguing about it, right? That that debate is encoded in our technologies, which it completely is not right now. And that's the biggest problem. I mean, I think we can make AI a public good when we think about it as a public good.

So. I'd like to ask the group that has the orange card up on the board there, and you, Andrew. I don't know exactly— to expand a little bit on using AI to bring people with complementary interests and coming from different generations together. That seems very abstract to me. I don't know how you do that. Can that group tell us what they had in mind a bit? Are they saying that? Or I see 2 asterisks next to the— a footnote.

Yeah, there's a footnote.

We weren't actually sure that it was meriting AI because we figured it was a sort of model of a dating sim. But we had also explored the idea of making it a compulsory course at McGill that every first-year student had to enter into this kind of Well, not a database, but community, friendship community. And then we said for the beginning of it, we could take all the MCLL people and they would input all their interests and their cares and so on. And then everyone would be required to, for credit in that case, to find a friend in the course of the year and develop that friendship. Kathleen, you know. But it would be the matching would be done by the AI, we think. There's And then the footnote is, what would be the sort of morality of matching with an AI system? Because that's important, and we don't know.

It has the same idea, but it's not quite the same way. But our idea was really—

Wait, wait, wait, wait, wait.

Our idea was that we would be comparing our student records They've got all our information online, and if our photographs were there too, aside from the fact of what we're studying, which kind of study groups and so on, they would have probably enough information. They have our— since we're on the McGill student database, no reason why they can't compare and bring people together from similar backgrounds and interests.

Yeah, the— one of the really interesting things that kept coming up was how to use this for, I think we mentioned at the beginning, the real interaction. So we think about book recommendations, it's sort of putting virtual objects together seems to be what computers like to do right now. And the really big challenge is how can you make real people meet other real people in successful environments? And, you know, the artificial intelligence dimension of that is simply, Being good at predicting positive outcomes. So when you have a positive social interaction, you let the machine know, and then it'll say, oh, okay, I took this person with these qualities and I had them go to this event and they were happy. And then I took this person with these qualities and had them go to this event and they were unhappy. So it's learning to partition the happy and unhappy events. But something that makes— facilitates you going somewhere and seeing somebody, even randomness. I'm also, one of the things I love about computation is randomness. So we really often think we want these things to be good at making guesses, but sometimes the best thing about human interaction is it's random, it's unpredictable, and it gets us out of our shells and our grooves. And so the nice thing about computation is to say, well, just don't make an informed choice, make a random choice and see what happens. So.

Yeah. Our— oh, my voice is very loud. Our group actually came up with some ideas, but we actually spent a long time talking about the ethical aspects, but also the limits to AI and to, you know, should one limit AI? And maybe there are areas where AI shouldn't go. And any thoughts about that? Because, you know, we talked about just winging it, just saying, okay, AI can be done for this and that and so on. But, you know, where should we stop?

Where—

what are the boundaries where maybe AI shouldn't go?

I'll just say, and this is coming from, you know, the recent thinking about this, is that It's not a matter of where it shouldn't go. It's a matter of being able to turn it off. And so one of the projects that I think we're going to start in the next month or so is a system that allows us to study different forms of radical transformation in early modern Europe. And the system will allow us to track different forms of radical transformation across space and through time. But you'll be able to move away from that at any point and go into the archive, which is digitized, and read the texts so that you can move between distant reading and close reading. And you could simply— you could just close off the AI function at that point and spend time with the text. And this has been a real interesting concern in the use of digital technology in the humanities, is how do you move between these 2 modes of reading? But for me, the basic principle for me is dialogue between the machine and humans, which means that You don't have to shut it out. You could if you want to, because it's always going to be something that you're working with. And what Andrew just said about you don't want an algorithm that just says, OK, that's it. It's not changing. And it's going to work this way forever. And we're going to be able to say who's teaching well and who's not teaching well. It always has to be dialogical.

Yeah, and I think that makes it very hard to know when you wouldn't want it. For me, the question is always making sure we're always talking about it, I guess, almost. Always revising, always engaging with the decisions it's making. I mean, it's such a problem now with social networks and polarization of news and belief systems. It's always impossible for me to imagine going backwards. So it's— I can't imagine saying, oh, well, if only we just go back to newspapers, that would solve the problem. But my sense is there's partially an engineering problem in there, that the way this has been set up is generating these partitioning effects, and that there might be a way now, if we have consensus that we do not like this outcome, is there a better way? And that's where the stakeholder issues are different. Readers' goals are different than social media distributors' goals, right? It's very good business for Facebook to have as polarized an audience as possible. We know as political beings in a polity that's very bad. And so figuring out a way to get control of that process is a big part of the future.

Hi.

So at the same table as Richard over here, and I thought I might be able to elaborate on some of the concerns we actually had more specifically to be able to ask another question. One of the things we're really talking about is how AI and technology in general is often used to increase our productivity, and in a society which is always looking for more and more productivity, a lot of the tools that we have are oriented towards achieving that. And so, just to maybe use a McGill-specific example, there's very limited resources, as we know, for our universities. I'm from a provincial student association, so something I work on a lot. But in mental health services, for example, You're seeing an allocation that's not enough to be able to really meet all the needs of students. Students are being put towards websites with information on mental health instead of being able to see a counselor, for example, and there's always these difficulties in resources and resource distribution. So there's always this idea that technology, for example, AI might be able to replace people so that those resources could be reallocated to other places. So maybe in the case of mental health services, you can have automated appointments and matching with counselors, but then more More services or more resources directed towards actually having counselors who can have that face-to-face interaction. And so that would be the ideal, you would think, if we're gonna implement AI or any sort of technological support to make this system better. But in reality, a lot of the times we're seeing any increases in efficiency is just gonna mean those resources get reallocated to another area and you're still seeing those same problems. So I'm wondering, a question for the room, but especially our speakers, And maybe also speakers who are coming up, what we can do to make sure that as we're implementing these sorts of solutions, we're not just reallocating resources somewhere else and we're actually making sure that people are having more time to, you know, have that face-to-face interaction, more invested in human resources, meeting people's needs, and being able to have more like collective and individual self-care and interactions as well.

That's great. That's a great question. Now, have you seen or read about the computational counseling, computer counseling?

Not yet.

We had a meeting here a couple of years ago about the future of the university, and we had some people come up from Stanford and show us work they were doing on computer avatar counseling for war vets. It was quite remarkable. It was really remarkable. And in fact, for the vets, it's often the case that they prefer the counseling of a non-human because they have— they bring a lot of shame home with them because they've survived and their friends have died. And so, the fellows from Stanford showed us a video of a counseling session between a woman that was not a woman, she was a machine. And a man who was trying to deal with post-traumatic stress disorder.

Stress disorder.

And it was really remarkably effective. And the computer was very knowledgeable and very sensitive. And the person who was being counseled, he was— his skin tension was being monitored, his eyes, his blood pressure, in non-invasive ways, but in ways that allowed the system to track his responses. In some ways, it's absolutely terrifying to me. In some ways, it suggests a way in which we could not replace human counselors, but we could add additional resources that can be very effective for helping people work through problems. And we certainly— we need those additional resources. So, and I understand what you're saying about how resources move around, and there's a constant competition for resources within a university. I know that very well.

I think we may have to put an end to the questions, unless you have nothing on for the evening. My head's spinning. This is marvelous. We could go on for hours. This is absolutely amazing. Did you actually just put this summary thing together?

He did.

Andrew did.

Oh, that's good. That is very good. This has been very provocative.

Thank you.

Thank you.

Thank you.

That was marvelous.

That was really cool. Please do it again.