Artificial (and Human) Intelligence - A Conversation with Prof. Rose Luckin

Welcome to the Future Learning Design Podcast.

So we really do need to get collaboration between those communities, and not only will we then end up with more appropriate technologies, we'll also end up with a community of more knowledgeable educators and developers who will constantly be looking at the data with a sort of researcher's mindset.

Welcome back. Thanks for joining us again on the Future Learning Design Podcast. It was fantastic to be able to speak with Rose Luckin, who is the Professor of Learner-Centered Design at the UCL Knowledge Lab in London. Her research involves the design and evaluation of educational technology using theories from the learning sciences and from artificial intelligence. Rose has a particular interest in using AI to open up the black box of learning to show teachers and students the detail of their progress intellectually, emotionally, and socially. Rose is also the Director of Educate, a London hub for educational technology startups, researchers, and educators. Rose was named on the Seldon List of 2017 as one of the 20 most influential people in education. She's taught in the state sector, in further education, in higher education, and was previously the Pro-Vice-Chancellor of Teaching and Learning at the University of Sussex. Rose is the author of Machine Learning and Human Intelligence: The Future of Education in the 21st Century, which she published in 2018. She's also the specialist advisor to the UK House of Commons Education Select Committee and the co-founder of the Institute of Ethical AI in Education. Hi, Rose.

Hi, Tim.

How are you this morning?

I'm all right, thank you very much.

Well, brilliant. Thank you so much. It's a real pleasure to be speaking to you today. And I've been really interested in the work you've been doing on, obviously, artificial intelligence, but the way that you've begun to articulate human intelligence as a way to differentiate, because I think obviously intelligence has been around for a long time and we're quite enamored by this concept, but there's different ways of seeing it. And I just thought perhaps that would be a good place to start. If you could just give us a bit of an overview of what you think about human intelligence before we get into the artificial intelligence stuff.

Thanks, Tim. Yes. And thanks for asking me to talk to you. Absolutely. You're so right. We have very quickly moved to thinking about artificial intelligence and whilst The subject of intelligence has been discussed since the days of Socrates. We need to really think about human intelligence in respect to artificial intelligence, which is something we haven't been doing for that long. Because I think the important thing is to differentiate between what you can do with artificial intelligence and what you can do with human intelligence. And quite understandably, given where modern artificial intelligence started, if we think of that as being the famous conference where a group of scientists got together and thought that they would be able to analyze every aspect of human intelligence to such an extent that you would be able to automate it. And that's just not the case. You know, 70 years later, we're still nowhere near being able to automate all of human intelligence. And I don't believe we ever will be able to automate all of it because I think there are some aspects of human intelligence that will always be beyond artificial intelligence. And that's what I think so important. But because of where we started, I think we've been very quick to believe artificial intelligence is more intelligent than perhaps it is, because it's very good at doing the things that we have valued as humans for a long time. You can get an AI system to perform well on IQ tests. So you can get AI systems to do well on exams. It's very good at dealing with well-defined problems, dealing with crunching lots of data. And so we've tended, I think, to overestimate the intelligence of our AI systems because we underestimate our human intelligence. And that's why it's so important. We don't misunderstand me. I think our AI systems are very useful and very smart, but it's not the same as human intelligence. So I think my approach to human intelligence was to try and understand it from the perspective of artificial intelligence and to look specifically at aspects of human intelligence that are not easy to automate. So the way that I've tried to unpack human intelligence is as obviously very complex, very interwoven. I'm not suggesting that we have separate intelligences, although, but you know, I love the work of Howard Gardner around multiple intelligences. I don't see it as separate intelligences. I see it as different forms of intelligence that are created from this interwoven complex set of ways in which we interact with the world. And yes, there is the academic intelligence that obviously we prize. And I think increasingly that must be interdisciplinary because the problems we face in the world, you look at coronavirus, that's a great example. You need loads of different disciplines in order to try and tackle the problems that that poses. And then you need social intelligence because of course because we need to be able to interact together effectively to solve big problems together. I think an aspect of intelligence that's underestimated to our detriment is what I call meta-knowing. But I know that frame is sometimes used to place metacognition. And what I mean by meta-knowing is more about epistemology, which is an old-fashioned, probably very unpopular word, but it's really about understanding what knowledge is. Where it comes from, what evidence is, and how you make good judgments about whether you should believe something to be true or not because there's good evidence. And I think in a world of fake news, that's increasingly important. And then absolutely there's metacognition, which we know is hugely important. We know as humans we can learn to be good. We can develop highly sophisticated metacognitive regulatory skills, metacognitive awareness, you know, skills that are hugely valuable That's something that's way beyond AI. And the same with the meta-subjective intelligence is though not just EQ or emotional intelligence, but actually that the kind of ability to reflect on our development of emotional intelligence and on the way that other people are or are not developing emotional intelligence. And that impacts on the way that we interact with them socially. And then, you know, meta-contextual intelligence, I feel, is a way of looking at intelligence that we really underestimate. And again, it's way beyond AI. So, you know, as humans, we're really good at moving around into different environments when we're not locked down, interacting with lots of different sorts of people. We're not fazed by this. You know, I know that before the days of lockdown, I could get on a plane and I could travel to a part of the world where I'd never been before, and I would be able to function. Now, I might struggle with the language, and I might struggle with some of the signage if it's a language where there's no translation into English, but actually, I know I can manage.

Yeah.

And that's amazing.

Yeah, it is. Yeah.

And it's incredibly difficult for AI. And then, you know, I think if you really get this right, then you can develop this, what I feel is highly prized, and what we really need is this accurate perceived self-efficacy where as an individual you can set your own goals, you can make judgments about how likely you are to be able to achieve them, what you need to be able to do in order to achieve them, who might help you do that, how you can develop that to know when you're moving towards them accurately, when you've got there, how how you then move to the next space. Do you know what I mean? It's a holistic view that I think it's important that we acknowledge and that we try and move our education systems more towards not just looking at the academic. And I know in a lot of schools they do way more than that, but I'm not sure that we cherish that enough.

I totally agree. I think that it's such an interesting idea that I think we underestimate the other aspects of all those things you've talked about. Which all, you know, sound in a way quite complex. And obviously in terms of processes, they are deeply complex, but we've always been around and about them, right? We've talked about reflection or we talk about growth mindsets. So we talk about, you know, it's all very similar and connected stuff, but I think we've underestimated or we haven't given it the label of intelligence. So we've underestimated the rigor of that. And therefore we kind of think, oh, well, the only really rigorous aspect is the academic intelligence. And of course, of course, that's nonsense.

Right.

But if we, if we get into this slightly more fuzzy, let's say, aspect of the meta dimensions that it becomes much more difficult to deal with, but it's no less important. In fact, I mean, I think in a way more important.

It is more important. Yeah. Because if you have really good metacognitive skills and you understand about what knowledge is and where it comes from and good social skills, you can actually much more easily gain those academic skills. I think it does, unfortunately, also come down to measurement. You know, we tend to treasure the things we measure and measure the things we treasure. And of course, now we are in a position where we can actually evaluate— I prefer as a word to measure, but we can start to really evaluate these meta-intelligences in a way that I hope will enable us to give them much more priority and, and to give them more respect.

Yeah. And I loved what you said as well about accurate self-efficacy, because of course, you know, there's all the confirmation bias or, you know, cognitive bias aspects in which we're aware of in terms of the fact that we're very subjective as individuals. But at the same time, that doesn't discount all of that stuff. Right. And the better we get at articulating what these kind of competencies and capacities are, we can become perhaps hopefully more accurate in the way that we can self-reflect.

Yes. I mean, you're so right. There's so much evidence. I think most of it is originated in medicine, but demonstrating that the more information you have about something, the more confident you are that you understand it and you know it, but in actual fact you're not. And that relates back to the meta-knowing, the epistemic cognition, the understanding of what the evidence is that you should demand of yourself to persuade yourself that you do actually know and understand this stuff.

Yeah, absolutely. And obviously humility just being a big part of it.

Exactly.

So then obviously that's— that paints, I think, a really useful picture of human intelligence. So then if we then move a step into artificial intelligence, I was really interested in what you've been saying in your writing about artificial general intelligence, and maybe you could just say a little bit more about what that means as opposed to more— a more specific artificial intelligence. And then I just wanted to ask you the question, how far away do you think we are in terms of that kind of futurist idea of artificial general intelligence?

Well, I think it's interesting, you know, at that original meeting back in the '50s, really, I think what those scientists were thinking in terms of was artificial general intelligence because they wanted to be able to understand every aspect of our human cognition. And years later, we're nowhere near that. And I think every time we get a bit nearer, we suddenly realize that, oh, it's way more complex than we thought it was. So it's, I personally don't think we're anywhere near, sorry, I've answered your second question first. In terms of artificial general intelligence, it really is that intelligence that can, like us humans, operate across the piece. So if I think about where we are currently with artificial intelligence, we're mainly in the state where we've got narrow artificial intelligence. So yes, we can have really good, say, diagnostic systems. We can have really precise surgical systems that can assist human surgeons to perform surgery. We can have excellent systems that can play world-class chess or Go or any of those things. We've got self-driving cars. We've got systems that can recognize faces and can interpret emotions. But no, my human surgeon will probably also be able to drive a car and recognize my face and work out when I'm feeling anxious. He may well be able to play chess or Go, I don't know. But our systems that are surgical experts can't play chess and our systems that can play Go can't drive cars. And, you know, so it's, that's the difference. It's being able to do all of those intelligent activities. And that's what we really mean by artificial general intelligence. And I just don't believe we're anywhere close. You know, every advance is a step closer, but we're still a long way from it.

Is the, is that where we kind of segue into machine learning? Because if we can build a learning capacity into the AI, then it can then begin to almost teach itself or learn itself.

Yes. And you see, now this is where the really interesting piece is, I think. So what really happened as a result of that conference in the US in the '50s wasn't machine learning. It was mainly rule-based. And so you had to write the program. It didn't learn. You know, you might write your expert system that could say, diagnose a cancer tumor. But if we learned something new about cancer tumors of that sort, we'd have to reprogram the system, write new rules because we were building a knowledge base, a fixed knowledge base within the AI system. We see machine learning comes along and because machine learning learns from data, from experience. And of course that is similar to the way humans learn. We learn in multiple ways, but one of our ways of learning is learning from experience. That means that we now have systems that can learn. And so it's very easy, I think, for people to fall into the trap and think, well, okay, that's okay. Now we've got systems to learn. Then we can have artificial general intelligence because humans are generally intelligent because they learn. But there is a problem there because it's still tough. And it's why there's such a big movement within artificial intelligence around explainable artificial intelligence is that whilst as humans, We cannot always justify the decisions that we've made. And sometimes, well, all of the time, we're really, really good at postdoc rationalization and all of those kinds of things, but we can improve at that self-reflective awareness, that ability to say, okay, I've made this decision because of A, B, C, D, E, and F. Machine learning can't do that because machine learning is black box in that If you think about the AlphaGo system that was built by Google DeepMind to play Lee Sedol as the world Go champion, it took DeepMind a long time to analyze where the precise moves were that made the difference. Now, of course, there's lots of work within AI that's trying to help that happen more quickly. In other words, help scientists be able to identify within machine learning systems precisely how they can explain the decision that's been made, justify the decision that's been made. But it's not easy because those systems don't naturally allow that. Of course, the old-fashioned rule-based systems, it was easy. That was one of the advantages of those. So in actually what's happening, as I understand it, in explainable AI is that some people are actually saying, okay, well, we need a combination of that old-fashioned approach with the machine learning so that we can then kind of use the best of both worlds so that we can then explain. And I think that's particularly important in education. I mean, another area where it's obviously extremely important is autonomous weapons, but we won't go into that. But you know, in education, imagine the situation, and I can say this as a parent and a grandparent, as well as an educator, you know, you've got an AI system, a bit like the Sorting Hat in Harry Potter that says, okay, there you go. This child here, 'No, they should go to this type of school and have that sort of an education, and that child over there should go to this sort of school and go to tennis.' That's never going to work. Obviously, I'm painting an extreme picture here, but, you know, at every step of the way, of course, as parents and indeed as students, we want to know, 'Well, why are you saying— why did I get that mark? Why are you saying you think I'd be good at this? Why do you think I need to strengthen that?' And so we're always going to need systems that can explain the decisions that they've made. So I think the most important thing and the thing that will actually perhaps lead to what might become what we accept as artificial general intelligence is the interface between the human and the artificial.

Yeah.

The place where that decision-making becomes joint. Yes, of course we want to hand over some of the decision-making to AI systems, but I think that's the low-level decision-making. Yeah. The high-level decision-making we want to be as humans, but there's always gonna be that gray bit in the middle. And I think that's, that's the bit that's really important to focus on. And so perhaps I would say I don't think we'll ever have artificial general intelligence that's purely artificial. I think what we might get is very smart AI systems that can really augment human intelligence to the extent that they can behave across multiple applications.

Yeah, and that's, I mean, that's fascinating, but that's a kind of perfect segue into what one of my little pet projects, I guess, is this idea of long-term memory. And so I'd love to just ask you specifically about this, if I can explain it clearly, but—

Sure.

So Alex Beard had this fantastic documentary on BBC about the future of learning and what it might look like. And that idea that, which he painted with, you know, a child being born in whatever, 2030 or something, had this kind of little pet AI that accompanied her everywhere and would essentially augment her long-term memory. So would log all of her experiences. And then she could then draw on those experiences in real time. So the reason I think I'm so interested in this is because I think there's still a big debate within education around the role of knowledge, right? And, and the way that we construct curriculum around, around knowledge. And therefore we should, some people would argue, we shouldn't disappear too quickly into the 21st century skills type arena because actually everything is disciplinary knowledge-based. You can't think critically about with things you don't understand, et cetera. And therefore, I think that's one of the things that keeps us in a particular place around the academic intelligence that you were talking about before. And if AI may augment human cognitive architecture to such an extent that in real time we would have an AI long-term memory bank that we can access and then process within our working memory, then immediately we then move away from the need to have a vast resource of long-term knowledge nuggets, if you like, in, in our long-term memory to neuron to think fluidly? And again, how far off into the future is that as an idea, would you say?

That's interesting. I think the idea of augmenting long-term memory is not that far away, but I think it's more complex than that because I really get upset about the division between knowledge and skills, because I think it's really unhelpful. Both are important. There's no question about it. You know, you need knowledgeable skills and skillfully applied knowledge because the core to knowledge being useful in the world is the ability to apply that knowledge, not just have that knowledge. And often the process of applying that knowledge, you'll know better than me, means that you have to adapt it, appropriate it, as perhaps Piaget would've said. You need to make it your own and know how you use it in different contexts. Yeah. And that's not just a case of memory in terms of that long-term memory of knowledge. That's the ability to meld your skills and your knowledge together to be able to apply it. So I think it's about more than that long-term memory knowledge base, as you might call it. I think it's, it's is the ability to understand how to apply knowledge. And that then comes back to the meta-level piece, because I think you are right. You can't develop those 21st skills without some disciplinary knowledge. Mm-hmm. But I think what you need more of is a deep understanding of an aspect of a discipline so that you understand how it works, if you know what I mean. Yeah. How that disciplinary knowledge has, has come about, what it's like, and then how you work with a small part of it. And that's actually the, it's the ability to understand how you learn about that and how you apply that. That's the thing that's the most important, more important than the actual factual knowledge itself. I mean, I know there's been a lot of talk recently about facts and factfulness, and facts are lovely. But actually, that's a bit of an easy way out because actually being— having a very good, factful knowledge base wouldn't have helped people with coronavirus very much, for example, because we're actually often having to apply thinking from other areas of understanding. So initially from understanding of influenza, for example, which wasn't entirely appropriate, but at least it was a start. And then you've got to know, well, why is that not entirely appropriate? Whereas if you're just taking that purely sort of very fact-based knowledge approach, oh dear, I have no facts about coronavirus. I can't do it. Of course I'm being extreme, but you see what I mean? So I think it's about more than long-term memory.

Yeah. I totally agree with you. Absolutely. I just, one of the things I definitely see within this debate is that many people keep coming back to the evidence around cognitive science, which is a great knowledge base for a applying into education, absolutely, but for me it's not, it's not sufficient. But I just hear resonances of all the things, meta-contextual intelligence, the meta-knowing, all the things you've talked about, like human intelligence. It's just all of those things which are again so important but don't necessarily appear within that kind of academic knowledge base.

And I think that's largely, I mean, there is a lot of work on metacognition, on self-regulation, and a growing body of work which comes under the umbrella of learning how to learn or learning to learn. But of course there's decades of evidence about memory, for example, and problem solving. And so there's a much larger evidence base and it tends to be a broad disciplinary approach. So it's not just— you've got experimental scientists working in that area as well as social scientists, uh, as well as developmental scientists. So you've got the kind of quantitative, hard measuring type approach, and you've got the more intersubjective approach. So you've got a, you've not just got many years of data, you've got a very broad approach to collecting that data and evidence for cognitive science, basically. And I think we're now growing a new body of evidence because I think It's really important what you're asking, Tim, because we often focus on the sort of birth of modern AI in the '50s and think, oh, you know, haven't we come a long way? We have, we've come an amazing— I mean, it is amazing. There's no question about it. What we often forget is we've also come an amazing way in understanding our own intelligence, our own processing. Yeah, there's still a lot that we don't understand, but we have come a long way and we do need to recognize that.

Related to that idea of growing that knowledge base, I loved hearing you talk about the signifiers where you were using some of the AI to identify particular signifiers in the way that, for example, students were working together on complex problem solving. Could you just describe that a little bit and just in the way that the body movements and the eye movements? I thought that was fascinating.

Actually, a really useful example for what I was saying earlier about decision-making and shared decision-making and how you might divide those up. So yes, we were very interested in looking at how students collaborate and it was very pragmatic, you know, looking around classrooms and not just in schools but in universities, you want students to do group problem solving, putting it bluntly, and you've usually got one teacher, maybe an assistant in a school, and you've got 30 kids or more, you know, how on earth do you manage that process? So is there a role that AI could play in at least doing some level of analysis of what's going on so that maybe it can feed back to the group of students themselves, but certainly it could feed back to the teacher so they know more what's happened when they weren't actually there. And so we thought, okay, we need to come up with data and we need to be able to analyze that data automatically. What does that mean? So we were working with a group of scientists from across Europe. And one of the partners was a furniture design company, and, you know, we had a technical design company too, and basically they built a sort of a test desk, if you like, where we had multiple ways of capturing data, not thinking that this would be the kind of thing that would appear in a classroom, but enable us to understand what was possible, and then you can think, okay, what's realistic for the class? And so we captured eye movement, we captured hand movements, we captured speech, of course we captured interactions with technology. Students were usually doing a task that involved some kind of physical computing or design. We videoed, they could respond through emotion buttons. They had their own tracking technology that if they wanted to, they could record how things were going.

So amazing.

There was a huge amount of data coming in. And initially, you know, when we looked at it, it's like, wow, yeah, we've got huge amounts of data and it's We don't know anything, you know. Yeah, because then you need to build onto that what we know from research literature about collaborative problem solving. So we then looked at collaborative problem solving and came up with an interpretation from the literature of collaborative problem solving, and that basically meant that we had sort of 6 stages of problem solving and 3 stages of collaboration, and we used that as a framework within which you could then start to analyze the data. So you put your mapping, learning science, if you like, to data, and then you go deeper than that. So in order to look at particularly those three parts of the collaboration piece, we look deeper into literature around whether there were any observable features when groups work together that were signifiers, as you said, of potentially collaborative problem solving being effective. And we found literature in psychology that evidence that coordination or synchronization of eye movement or hand movements was often associated with effective collaborative problem solving. Now it's important to say we're not saying that's enough in and of its own. These are two little signifiers that you would combine with many more, but they're two signifiers that you can collect data about automatically inexpensively and that you can analyze automatically.

Yeah.

So in order to test whether that worked for us, we did exactly that. We analyzed the hand movements and the eye movements to see how synchronized they were. And then completely independently, we asked an expert professor from another university who hadn't been involved in the project at all to watch videos of the students and identify 3-minute chunks where she was saying, "Yes, I think these students are collaborating effectively. No, they're not. Yes, they are. No, they're not." And then we mapped what the data from analyzing the hand movements and the eye movements said to what the expert said, and absolutely there was an interesting mapping.

Interesting.

Between— yeah, you could see it visually and it was there in the numbers that, yeah, absolutely it did look like these were a couple of useful signifiers. Now, as I say, you want to have multiple signifiers.

Yeah.

But these are the kind of things that you can collect easily, that you can then reflect back to the students, you can reflect back to the teacher and say, 'Look, hang on a minute guys, is everything all right?' or 'Perhaps you need to do more of this.' You can flag to the teacher, 'This group over here is really struggling,' or— and that's where, at that level, I think I'm happy for the AI to make the decision. That, okay, the AI is perhaps scanning 50 different signifiers and because of the way it's been designed, we're happy for it to make decisions at a low level about what those signifiers are signifying, if you see what I mean. But then you need the human to come in and say, okay, let me interpret that for this group. What does this actually mean at that higher level?

And they've got contextual information about those students, or, and then what do they do about that? How do they act on that information is then the next level, I suppose.

Yeah.

Interesting. And just as a final question, I just wanted to think a bit more practically than, because often in education, you know, we let the tech lead, right? And, you know, it's like, we've got this tool that does this thing, you know, let's install it and see, rather than leading with the problem and seeing where is the problem that we need to solve and how can we then have AI or tech to support us with that problem solving. And with Educate Ventures, as I understand it, you are trying to pull together that process a bit more clearly. Is that right?

Absolutely. I mean, I think it's important for for all educational technology, but I actually think it's even more important for AI technologies. We have to build better relationships between the people who use the technology, the people who design the technology, and, and researchers who can help both of those groups understand how you know whether it's working or not, to be blunt. And so we formed Educate as a program based on this notion of the golden triangle, three points of the triangle being those three communities, the developers, the educators, and students. And the researchers and the gold was the evidence, you know, and try and connect them around understanding the evidence. And of course, where AI is concerned, that evidence becomes data. And so it's all about engaging those communities in conversations. And I think where AI is concerned, that's even more important because to be blunt, most AI developers, and there are some good companies who do work very closely with educators, but A lot of AI companies know nothing about teaching and learning. In fact, a lot of edtech companies know nothing about teaching and learning. They don't need to. You don't need to start a business. And understandably, most educators and students don't know anything about AI, and a lot of them don't really understand much about technology. But actually, both have very valuable contributions to make when it comes to designing the most appropriate technology and the most appropriate AI. So we really do need to get collaboration between those communities. And not only will we then end up with more appropriate technologies, we'll also end up with a community of developers who understand more about teaching and learning, a community of teachers and learners who understand more about the tech and the AI, and hopefully a more integrated research community and more knowledgeable educators and developers who will constantly be looking at the data with a sort of researcher's mindset to say, okay, what's this telling me about whether 'Is this actually helping to solve the problem here or not?' And I feel we need to be much more sophisticated in the way that we look at the efficacy of technology's use when it comes to education. And incidentally, we'll be publishing properly next month, but we've just finished analyzing the study that we conducted last year during COVID to look at the way that educators and parents and students within schools were using technology. And it's— the data is fascinating. The biggest problem is the disconnection between the different communities. It really is. But where the connections were good, you could see huge benefits. And one of the connections that really developed on the data that we have, which is UK-focused, is the relationship between the edtech community and the educator community was really strengthened.

Good.

And that was very, very positive when that happened. So those connections are fundamental, I think, to getting much better design and application of technologies.

Yeah, no, that's brilliant. That's definitely necessary. Well, thank you, Rose. That's really such an interesting conversation. And as you say, there's a lot we have learned, but there's a lot more to learn, I think. But, you know, I just think it's such an important kind of literacy to develop amongst the teaching community as well.

Yeah, definitely. I agree. I agree completely.

Yeah, brilliant. Well, thank you very much. I appreciate it.

My pleasure. Thanks.

Brilliant. Cheers.

Bye.

We hope you've enjoyed this episode. Please feel free to continue the dialogues with our guests, with us on our blog or on social media, or within your own networks.