Google DeepMind + Microsoft AI Fellow on the Future of AI in Education

All right, I think it's started. So welcome back to everyone. We see a number of people joining the participants. Thank you for waiting while we were warming up in the practice sessions and making sure that we could initiate this conversation today. So very, very pleased to welcome Kathy Jane to our MicroOne Forum and Sky Zhao. You're calling from two different parts of of the world. But what you have in common is that you've been engaged in the conversation about technology, AI, and education for a large part of your career. So if I do justice to what I discover about you, Katie, if I can start from you, you're currently working for Google DeepMind. You had been working on this LearnLM, if I understand it correctly, where basically you're kind of fine-tuning learning science principle and deploy them with Gemini. And so you've been working basically in this space in which research educators, learners, they're all kind of looking at what would be the best way to translate pedagogical intuition into AI behavior. And I have a lot of interesting questions for— at least for me, I'd love to learn more from you. So very, very nice to have you here, Kathy, and welcome to the forum. I'm going to ask you the first question in a sec. What I also like to welcome is Skye, who is actually a lecturer at the University of Manchester. But Skye, we also found out your work at Microsoft as a senior fellow of the AI Economy Institute. What I think is fascinating about your work is, again, you are deep into the conversation about AI in education. And you have been kind of exploring across multiple different multilateral institutions. I noticed your work for the UNESCO in the AI competency frameworks, the Generative AI Panel at the United Nations. But one thing that I discovered that I found very, very interesting that I'd like to go deeper is the work that you've done on disabled students and of course the how they interact with GenAI tools. So there's a lot of different things that you're bringing to the table, both of you, Cathy and Skye. So thank you very much for being with us today. Cathy, look, I'm going to start from you, and this is our question usually just to break the ice. So tell us a little bit about how you got into where you are, like kind of what is your story? How did you become basically engaged in the conversation that today keeps you busy? And yeah, and welcome, very nice to to have you as part of this forum.

Hi, thank you for having me here. Um, my— how I ended up here was actually a little bit interesting. I wanted to be a theoretical physicist when I was in undergrad, and during the pandemic where I saw the world around us change, we moved to learning through technology, learning from home, is where I started exploring other fields that would interest me. That's where I got into machine learning and interned at Google a few years ago. Soon after that, I worked with Google Classroom, so I was involved a lot in building features for teachers, previously without AI, and then after GenAI became the talk of the town, I was working on building GenAI features. I worked very closely with a lot of education experts and, um, also with product teams, and that helped me understand the problems of deploying, um, AI in schools. So the natural next step for me was to then get exposure into how AI models itself are built and That's how I moved to DeepMind eventually. I moved last year.

That's fantastic. So I know you're connected today from Montreal. That's where I grew up, by the way. And I was in Montreal not too long ago, right? And I'm always reminded about the Montreal winters every time I'm back. So I hope you've been coping with it.

Well, I'm just waiting for the summers for now because it's still subzero.

I know, I know. It's surreal to believe that everywhere in the world is spring in the Northern Hemisphere, but in Montreal, I guess, and probably Toronto. So, Kelly, thank you very much for this. I will go straight into a bit more conversation with you shortly, but I'd like to welcome you, Skye, as well. Same question. You and I share a bit more of a commonality. Like you, I'm a professor, and I've been a professor for a bit longer because I'm older. But I know that you are an academic. You are basically in your assistant professorship right now, University Manchester. But how did you get into AI, right? This is a topic that maybe when we start— when you started your PhD, it was— I don't know whether it was central, was marginal, it became somehow part of what you currently do. Just share with us your journey as well, Skye.

Sure. Actually, my PhD is only on general education theory. So I did my PhD at the University of Bath. And at that point, I feel like there's something missing. For education in the age of, you know, technology. So at that time, AI wasn't a big phenomenon at that point. So I, I felt I need to work in the information and the technology field to be able to develop my understanding of how this will impact on education and how students and teachers could utilize such tools better. So I moved to Information School at Sheffield University. I worked there where I develop my, I wouldn't say expertise, but more kind of my understanding of AI and how that impacts education. And at that point when I realized English is not my first language, I always work as a second language, first as an international student in the UK, then as an international background staff. So English academic writing wasn't my strong point. So at that point I was using Wordtune, Grammarly, a lot of tools to help me to come back this barrier, basically. So then I realized there are a lot of issues with such tools if you are not being careful. So I did write a— I think it was a tech review for Wordtune at that point, and then somehow it become lots of citations because of the time of OpenAI launched ChatGPT 3.5. And also during that time, I was the academic integrity officer. As you can Imagine how many cases I have to review because AI-generated assessments. So that kind of bring me more kind of on the ground in terms of teaching and learning process involved with AI and also research perspective because of my own experience and background. So I've been working with initially Wordtune and then with different technology companies to see what is the impact, how we can combine academia research-like evidence, teaching practice, tech, and learners as a community, as a platform to be able to discuss what we can best benefit and minimize the risk associated with it. So that's my story.

That's your journey. So I love it. You look at one of the things I discovered when I'm interviewing in this forum, the Microsoft Senior Fellows, right, is how much academics are kind of engaging into the industry with real problems that they want to integrate a multi-stakeholder platform similar to your fellowship, but similar to many of the things you do. So I look forward to going a bit deeper into your work And I'm really happy that we have the two of you today here to really help us navigate the topic where at the moment we have more than 100 people currently connected live and more will watch you, right? Kelly, I'm going to go over. Yeah, so it's amazing. And I'm so proud that they decided to spend their time with us. So it's great. Look forward also to the question later on. Kelly, I have a question. I'll start basically on your work in DeepMind, right? So one of the things that I appreciate a lot with basically the core argument about the work that DeepMind Mind has been doing on LearnLM is that we have been having weak evaluation practices, which are one of the reasons among the many where we haven't seen the full potential of AI in education. And, and I think this is an interesting part because it's positioning the work you're doing in kind of looking at connecting this specific gap in many ways. And so one of the questions that people will ask you, right, considering some of the shortcomings that we have seen these days with AI is When you're looking at the work you're doing, how do you decide when a model is fundamentally generally teaching rather than maybe responding? And so how do you define the space as somebody who is basically working actively in this environment where the intent is really to connect rather than necessarily to create transactions between learning and responses, right? So what would be your take on that?

I'd be curious to know how you define responding responding versus learning, though.

Yeah, so mostly I think just the inputs in the sense that the model is simply responding to, let's say, a prompting exercise that is basically reacting to what is the infrastructure of the prompts. But I think there's a space for AI to eventually take a bigger pedagogical role, from what I understand from the intent of LearnLM. So maybe help us to understand basically how the model has been positioning itself in the learning space.

Absolutely. Learning as a field itself is very complicated because of the essential human element of pedagogy being nuanced. And the way our team in general has been focusing on is infusing the pedagogically best practices into Gemini. And position. So when you talk about responses, the way a good AI assistant is supposed to respond versus a good teacher is supposed to respond end up being very different. But, uh, for an obvious example is if I ask an assistant to do 2 3, I just need an answer, but I would probably want me to break it down myself. But even with education, there's— it's not a straightforward evaluation for us to do, um, to just evaluate for the learning science principles because you want to personalize, you want to also have the context of that user, the specific case that the model adapts to. And that's where a lot of this gets tricky because it's, it's not that straightforward an evaluation. When deploying this in classrooms, what often we do is have teachers in the loop and have human evaluation. So pedagogical experts actually see this deployed in classrooms and do some of our evaluations. I think that's one of the research you were referring to with the evaluations that you were talking about is we use Gemini in classrooms itself, had teachers actually moderate each of the responses, and before it gets sent to the student, they actually can choose to make edits if they feel like. And that— an evaluation like that helps us get very useful signals on what a good teacher model would be like, as opposed to more auto evals that we do.

Yeah. So there is definitely, uh, kind of a human-in-the-loop element, right, in what you're sharing with us. You mentioned evaluations, uh, Katie, and teachers. And Skye, if I go over to you, I think, you know, we, we, we both work in universities and your research has been quite interesting because you have been doing research on disabled students. And one of the things that I guess has emerged from what I was kind of doing some preliminary reading about your work is that while there are some educators that view GenAI skeptically, some of the findings, if I interpret them correctly, that you did on disabled students actually are very encouraging, right? That the AI, the use of AI in the classroom could play almost like a playing field, right, where you basically are lifting learners up. But there will still be people, right, that will start asking questions. How does that work? So share a little bit more about that research, Kai, on what have you been able to find out? And if I don't remember correctly, that was in Sheffield, right?

Yeah, when I was in Sheffield. Now I moved to Manchester University, Manchester Institute of Education. But when I was at Sheffield, I was also the disability liaison officer. So as you know, for academic, we have multiple roles. Yes, it's part of our challenging jobs, I would say. At that point, I realized that students are already using AI to help, you know, with overcoming their conditions, but it was not officially recognized. It's still not recognized as assistive technology, and students have been asking us whether we could provide some subscription, like for, for for the software, but because it was not considered as assistive technology, we cannot do that. And also, um, for departments such as disability support department, unlike academic department, they don't have funding, they don't have like researchers bringing in money, they don't have teachers bringing students income. So there's limited funding in the service department itself. So there's a lot of struggle within that as well. As a result, we see students have been, um, you know, paying themselves or just using the free version to help with their— overcome their challenges. So I remember when we surveyed the students registered with disability department, the 3 top conditions that they see most benefit with use of AI is ADHD, autism, and dyslexia. And when I report that back to the institution, like Sheffield at that time, it was people say, oh, we need to think about how do we support students better in this way. So we definitely see, you know, there are areas where AI could support them, especially in assessment environments where traditionally students, marginalized students tend to receive less, you know, support or struggle more. But at the same time, when we ask students their concerns, they also that when they use those tools, right, it helps them in short term with within a set assignment assessment to help them to maybe compete with students without, for example, learning disabilities. And but long term, they are also very worried because it also can automate part of their skills that they should develop. So it's kind of trade-off between students. Do they want to make sure that they get a certificate and get good grades, get a good job afterward, and then also stop using it. And so there's a lot of— even within the survey itself, students are having diverse viewpoints. But what is clear is that most students find them short-term, like it helps them with the assessment within higher education.

Wow, fascinating. So I'll try to build basically what I'm learning from the two of you, Skye and Cathy. And so building from what you say, Skye, and over to you, Cathy, right? One of the things that I was kind of reading about LearnLM, right, and the trial with the with Eddy, the Eddy partnership, where basically students who were receiving short AI tutoring, they were actually more likely to solve novel problems. And so when you kind of have that kind of evidence from running a trial and you're trying to basically go back to your team and try to get them to prioritize what we should be doing now, right? So how do you basically teach your teams about this lesson? How do you basically— integrate them? How do you find them useful for the deployment of the next iteration of the model? And if I can ask you, when you were working with this kind of trial, what are the things that usually surprise you, right, where you were not expecting prior, but then suddenly the evidence is showing something else? If you can share a little bit more about that thinking, I'm trying to create a bridge between and where Sky is also helping us, right, with this conversation and your work as well, Cathy.

Yeah, I think one of the things that Skye mentioned was something I'd just like to double-click on, which is just users wanting to use AI to learn, but at the same time, how do you draw that balance to make sure that you do not lose, like you continue to develop the skills that you had to while getting the grades. And that's something that very often comes up. Comes up in our conversations because in one place, yes, we are developing a pedagogical-first model with LearnLM and having LearnLM be trained to be a good tutor, but often as a user, as a student, there is an inclination to do, to get a quick answer and get your grades that are just as important in progressing to the next level. And how do you keep a student motivated? To actually learn becomes a very tricky problem, which becomes much harder to solve outside of contained experiments like the experiment we did with EDI, where there it's more contained. There it's trusted testers who agreed, so the motivation problem is solved. And there it becomes more defined as a problem to take insights from, as opposed to more open world kind of problems that Skye is talking about. Yeah, so that, that's definitely something that needs more active involvement and real classroom implementations and larger scale deployments. But with our experiments that we're able to do with EDI, and we're also running a bunch of other randomized control trials, um, in our global suite, so we're running one in India, in Sierra Leone, in UK, so running across schools with different access levels, um, and different resource levels. It has been interesting to take— some of the takeaways have been really interesting because when as engineers we're sitting in our offices, um, it's, it's very hard to build empathy for the kind of diverse set of classrooms that exist across the world. Unless experiments like this. That's right.

Yeah, sorry, sorry, go ahead, go ahead.

Sorry. No, so I was just giving an example of interesting insights that you can get, which end up differing a lot. For example, one is the kind of language a tutor uses, and the cultural expectation of how much praise a student gets for the same kind of answer. Is very different in different parts of the world. And if the tutor is not personalized to that local area, it just sounds fake and it's going to be repelled by it. And that interesting cultural insights like that have been some of an interesting takeaway.

That's right. And so, Kelly, look, if I can do a follow-up just to go a little bit deeper in something you said that really intrigued me, So you're getting these insights, right, from the experiments you do. Then eventually you have to translate the pedagogy into prompts. You do have an engineering team that is working on this. What kind of knowledge did you find is the hardest one to operationalize? It's because I would imagine there would be area where you find that maybe the relationship between the planning part in the pre-deployment and the post-deployment, they're kind of linearly distributed so that you feel there is a way to control that and you can improve it. And I am sure there are examples where no matter how much you try to codify knowledge, it becomes difficult, right, to really render it for what it is. Is there anything that comes to your mind about some kind of pedagogical pockets that are so difficult to operationalize?

That's an interesting question you ask. On that, I actually thought it's a good point to note that now we do not have a separate learnLM model. We did have a LearnLM model until last year, but now we've infused the same capabilities to the Gemini Flash and the Gemini Pro model. So as Google, we're working on one big model with instruction-following capabilities and pedagogical capabilities that teachers can get the same LearnLM behavior out of the same model. And that's where I feel some of the other challenges also come from, because when you're training the same model to be able to perform in multiple very different kind of roles, finding the right kind of examples to be able to point out the losses, the places where the model gaps are, end up becoming a lot more tricky, as opposed to training a model purely for education or for one purpose. A general purpose model ends up becoming much more tricky.

Fascinating. Katie, thank you again for walking us into this. Sky, look, I'm gonna take you back to a common place, which are universities where you work, you are currently working, and also some of the roles that you played before, right? Especially we're talking about the integrity officer, right? Now, universities are currently experiencing, I think, quite existential challenges, right? There's a lot of anxiety. I'm sure you feel it in the UK like we feel it here in the United States, right? The question The question is, how do we eventually— are eventually protecting academic integrity? So look, it's not uncommon that here in the university you will see a colleague that has paper exams because they don't trust the technology, right? But we can't necessarily go back to a world in which technology does not exist anymore. So mostly based on your empirical work, the work you're doing as a professor, right? Do you think that somehow this anxiety is justified? Or do you think that, like it has happened the past. We are trying to solve the wrong problem. What's your take on this? And I'm intrigued by kind of the work you've really done in this field.

I think in terms of academic integrity in different contexts, that's— I mean, what I want to say, like when I did a focus group in Saudi Arabia and I will talk to you, I also did some follow-up interview with Indian context. Some of the feedback I got is we have a lot of exams and then we see students cheating with, you know, AI. Maybe it's time for us to switch to essay-based, whereas I'm based in the UK, right? And most of the assessment in higher education is essay-based. And we're talking, oh, maybe we should go back to the exam-based where there's pen and paper and then put students just in the one room, and then until, you know, they finished answer, they get out. We know traditional essay exam like that, locking students in the room, is not best for students with learning disability already. So I think in a way we have to define what do we mean By, you know, breach of academic integrity in this age, we have lots of policies, very definition-driven, like putting there to say, oh, students, please do not commit plagiarism, do not use, you know, AI-generated contents and then pass them on your own. But we don't have training for the students to say what is right, what is good practice, what is bad practice. Even within our own educators' community, we can't have an agreement for that. So I You see the low lack of trust between students and staff as a result of that, because students say, if you are penalising us for, you know, without clarifying what it is academic integrity, then it's not fair. Whereas for staff members, increasingly seeing all the things very similar, you know, when you look at essays, there are dashes, there are bold capital, you know, letters, it's very easy to notify find what is AI-generated, we can pick it up immediately, right? So that also reduced the trust between students and staff. So I think there is a time where we need to at least define what is bad practice, not just giving a guide. And then we need to explain to students better because even in the courtroom we say, well, you assume innocence until proved otherwise. Whereas increasingly, I think in higher education, we are trying to either switch the exam or even having viva for master's students or undergraduate students, not because we want to develop their oral defense skills or how they could be using in academic scenario in the conference in the future, but we just want to say, oh, that is what you write. And that is, for me, that sounds cruel because we don't trust the students. So let me just find out that's yours. So the purpose of having a defense is no longer to help students to develop the spoken skill, you know, for academic spoken skills, but become, okay, prove that it's yours, otherwise I will not give you a degree. So I think we need to challenge what we— have we done our job well as an educator? Do we have an educative stance for academic integrity, or do we just take a punitive style like, I'm going to penalize everyone until nobody offends? And I think that's something we need to fundamentally change, not just to say, well, let's just give everybody a low mark.

Mm-hmm. That's so good. That is so interesting. I'll try to ask you a follow-up as we stay still at the university, right? So let's imagine that we are able to finally move away from the incorrect focus, right? Rather than they talk to me, we are able to understand that Trust is really the currency we're building in universities. Somebody could argue, sure, but this technology is making people so isolated, right? But you actually gave a lecture, if I did my homework correct, where you were talking about building collective agency in time of technology, right? So digitally mediated world, if I understand correctly. So how do you do that? How do you build that collective agency that maybe before was transpiring from the classroom or the class experience But today we're so distracted by this personalized use of AI. But you have a way to think through this, and I'd love for you to share it as we kind of stay inside of the university for a moment before we're going to go back to the industry with Katty. Please, Sky, what do you think?

I'm not sure my approach is the correct approach either, but I do go to the class, you know, being very transparent, saying that I'm also very vulnerable at this stage. This is how I use AI in my teaching. I don't know whether is right or wrong, tell me what you think. And now you tell me what you are using and also explain to them, you know, many of the university will have this template where people could acknowledge whether they've used AI or not, right?

For—

so it's more kind of skills, I would say, or changed into a more dialogue in a way that we can be having a safe space to say this is what I do. I'm not sure whether it's right or please reflect on my teaching perspective as a student and tell me how I can improve. And now it's time for me to do the same for you. But I don't think this approach is replicable for large classroom settings either. Imagine if you have like a class of 50 students or even more than that, how would you going to go about doing this? So I think it's just one of the scenarios that I did. I'm not sure whether that's what you are referring to.

Yeah, yeah, yeah. This is what I was— I found when I was looking at your work, right? And I found it was extremely compelling, the idea of building collective agency in time on digitally mediated life, right? And I think this is an area of tension that we need to keep on curating more and more. So, Sky, look, thank you for having taken us through two dimensions of the university. One is like the educator that are kind of maybe pushing on the wrong rationale, and then the question of building that collective space, right? Technology is not the enemy, right? Just a tool in this case. So if I bring this over to you, Kathy, again, kind of in terms of the work that we can learn so much from what Google DeepMind has been able to do. So look, you were mentioning this before, right? You're reaching hundreds of thousands of students across the world. I know there are tons of partnerships that you have made in different parts of the world. So the question is, reaching out is important. Important because it's really about almost like closing a gap of access. But how do you eventually measure that the reach out has converted into shifting the outcome? In other words, how do you make sure that it's not just about reaching that sets the KPI, but creating durable learning? And what is basically your thinking behind this? This is an area where we don't have an easy answer. Many companies, they think sometimes the reaching out is solving most of the issue, but somehow it's more about creating in a culture that didn't initiate from that shift, right? So share your thinking with us, Cady. I find this conversation with you ladies fascinating, to be honest. So I'm enjoying the ability of learning from you so much, and I look forward to the question that some of our attendees— again, we haven't gone away from about 100, right? They stay with us for the entire time. So Cady, over to you.

Well, that's not at all intimidating to answer, But I, I, I don't know what the official Google view is here, so I, I just share mine. That may or may not reflect what the company thinks, but I think there are like two parts to it. One, of course, is the business part where, uh, it's good business to be in schools, to have that kind of brand value that is recognizable, which incentivizes all the edtech companies, Google included, all the other AI competitors we have to actively reach out and have that kind of breadth where they're able to analyze what the students want, what the teachers want, and improve their models. But at the same time, there's a shared collective responsibility and a shared collective long-term approach of companies working together with teachers and with educators to understand how do we inculcate AI in classrooms that is more sustainable. Similar to some of the things Skye mentioned earlier, that a lot of, there's one part of a lot of times that companies are able to reach out and have technology used in classrooms, but they'd probably be used in more unethical practices or they derail from the conversation. For example, there is a lot of research that is talking about how using of technology is adding to frustration in classroom because of the technical issues or because of teachers feeling pressurized to cover more content in the remaining amount of time because covering the content with AI or with the fanciest new tech tool in classroom just took longer time duration. So while the outreach has helped industries or just technology infiltrate classrooms, it still remains an open question about how do we actually use this for sustainably improving learning outcomes without having more unintended side effects. And a lot of those interesting insights I've read, only when I've like spoken to teachers going to classrooms, for example. Um, one of the common things, one of the things that actually I didn't think, uh, I'd have expected was, uh, this part where I was telling you about tech— use of technology actually adding more pressure to teachers to cover more content because the amount of content they have to cover is still the same.

Curriculum.

I think the other concern that is coming in is that we are talking about using AI to teach the existing curriculum, but there's not, uh, it's still something that we have to work together. I, I don't have answers to that yet, but how, what is the curriculum that needs to be updated itself? Like, what are the skills that need to be taught, uh, given that there is using AI. I've been working— so I come from an engineering background, so this is a very engineering-focused answer, but I work a lot with undergrad students, and the most common question now that I've been asked, and I'm on a spot, is, okay, why do we even need to learn coding given the AI models can do it directly? So sure, uh, 'your pedagogically sound model is helping us learn, but why do I need to do it in the first place?' And, uh, those questions are something that are much more trickier to answer. Like, for this one, I have an answer. Like, I don't think AI can improve your coding skills unless you have those essential skills in place. Like, uh, AI can only speed up. But there are a lot of skills also going obsolete.

Yeah.

So I'm not sure if I fully answered your question.

You did, you did, you did, Kelly. I mean, you did because of somehow, first of all, I think it shows that we're making lots of steps forward, but they are equally at the same time, there's a lot of considerations to be made, right, as we're making this progress because it doesn't come as a mainstream. There's a lot of nuances that I think are emerging from the conversations. The other thing is There is a lot of progress being made that is encouraging, but the job is not done fully, right? It continues to be basically an evolving field. So I think that the area of experimentation that needs to continue to basically pivot is where more and more questions similar to the one that you're answering will give us a sense about how do we keep on basically engaged. The answer might not be necessarily getting to a fully integrated system the way we wish to, but to a system that is functional enough to allow us to keep on progressing. And I found this part to be so interesting about the work that, well, first of all, you do at Google, but equally what DeepMind has been able to do. So absolutely, you did. I'm conscious of time, ladies. So I'm going to try to put it to a point where maybe we can kind of wrap up and then see if questions come in. There's one question that I have for you, Sky, before we kind of move into the Q&A, right? So one of the things that I, if I understand correctly, your fellowship with Microsoft is examining is the global pathway of AI diffusion. So similar to what Kelly was mentioning about before about reach, right, or outreach, you are studying diffusion across countries, different skills, different governance structure, different policy strategies. So if you kind of look at that experience, in part of your research, what are some of the most interesting divergences that you have come across, something that is generally different from mainstream that maybe we're missing out? Because I guess every single system generates its own sense of meaningfulness or significance. But there must be something that you were looking at in terms of diffusion that kind of caught your attention across the work you had done.

So the fellowship is recent, so we're still collecting data as you've seen. But the fellowship was built on a previous finding from a UNESCO global survey that I did, which we look at how AI was adopted by students and teachers. Again, this is in higher education. So I can say from there and how we're going to build the interviews based on that. Yeah, it's all right. So, well, we have around 3,000 responses for that global survey. And what we did find out is there's, again, it's not surprising there's a huge digital divide, right, for the Global South, Global North in terms of digital infrastructure. But in terms of training, actually a lot of countries in Africa have leading hours to provide the training. And what we see is teacher and students talking about, okay, we have a lot of theory-driven, this is what this tool, this is what AI, this is what this GenAI, but there was no tool for them to play with. So I think the digital first-level digital divide is still there. So we need to have more hands-on and more kind of tool access globally. Another thing is a— what I can remember at the moment is the linguistic dominance of English and Mandarin Chinese. So a lot of tools in other regions are only this kind of two languages. So for local language, for them to be able to use those tools, they have to have English level or Mandarin level to be able to do it. Then the local language translation is not that great. So that creates another barrier. And then when we look at AI frameworks, we try to give a, okay, here's a top-down, here's a framework, let us know what you think of it. And then we realize that lots of cultures, for example, I've talked to people from Bangladesh context and the feedback is, well, we are very community-based knowledge co-construction. So giving us one guide, ask us to adhere to it, it's not working. We need something where it's grassroots community generated. So I think there's lots of localized culture, geopolitical reasons. For example, in Macau and Hong Kong, where I talk to people, they don't have German, they don't have ChatGPT, they can only use Microsoft, right? So they are also talking to Belarus people who, because of sanctions, they don't have those tools either. So there are lots of factors that affect people's use of access to AI, hence their skill development. And for that, I think once acknowledged the digital divide, things we can do, but in terms of funding, but also we need to to see what is the current emerging practice even within those digital divide situations. What are the people on the ground are combating? What are their, you know, solutions? We need to collect those solutions so that we can better bridge this digital divide. So this is what the interview is about, not to say, oh, how your country is adopting this AI tool. What these people, what are people doing on the ground? What do you think is the best practices? Let us know, and then we cross-compare, and then we put similar situations in a group, in a theme, so that we can generate more knowledge in a way that is— one is kind of policy, the one is funding, but also the practice, the voices from the educators, from the learners, from the local schools, or in higher education, the leaders, what their perspectives are. So these are something that we want to find out, but those are the striking things that That's immediately from that.

Yeah, fascinating. Well, Sky, I look forward to reading your work once your fellowship gets to that point. So I hope that we can continue to be in touch and, and, and, and follow each other's work. Kathy, look, there's a question, and I don't know who is asking because it's anonymous, but I think it's an interesting question in our Q&A. And the question is basically, where is the limit where we basically are defining this, this boundary between technology and humans, right? In many ways. So it looks like we are humanizing a lot of the AI in many ways, right? And I think the trajectory feels much more that not for everybody, the difference will be clear, right? There will be moments where we might be confused about, am I dealing with a human who is being elevated by AI or am I using with a humanized AI, right? For the work you do, like from engineering, that is basically part of your work. Where do you see the trajectory going? Do you find that there is more, more for more. So we definitely are humanizing technology more, and therefore there is a risk similar to how this question kind of is being positioned. Or do you find that no, actually this is maybe a perception problem, but there is a clear space between humans and machine that we're not going to eventually cross, right? Where do you see this trajectory going from, from your perspective?

So I wouldn't say this is limited to its impact in education itself, but in general use case of humans with AI, anthropomorphization of AI has been shown to be like a serious concern with interaction because it leads to emotional dependency. And that is something that Gemini actively works on, that the models do not pretend to be a human, but as the AI capabilities get very good. Good. LLMs can talk as well as humans, or AI-generated images can look almost as good as a real image. So it definitely is a real concern that, um, for somebody who's not well-versed with technology, how do they distinguish? Or even for somebody who knows that this is technology, it is very hard for them to draw those boundaries in that interaction between between a human being and AI. For example, it's a running joke where a person replies thank you to the Gemini chat and they're wasting tokens, but they're being nice to an AI model because they've somehow associated the same kind of human politeness to it. But where this goes or what lies ahead is something I don't have an answer to, but I also think as a society we don't have an answer to. Um, that's right. It's— this is unprecedented technology that's moving at a pace where what's imperative is we continue to have a lot of these discussions, not just in forums like these, but also in pol— with policymakers, with politicians, and bring a wider consensus into what is right for society.

Yeah, 100%. I think— yeah, yeah, sorry, sorry.

But as a concern, I definitely agree with that question, right? Yeah, about—

yeah, that's right. Sky, look, there's a question on integrity that seems to be written specifically for you, right? I'll try to summarize it somehow. And this is no, no, no, no easy, right? Yeah, asking this question, right? So if you think about the fact that we want to create a space for people to learn what is ethical AI integration. So somebody will tell you, yes, we want to integrate AI, we want to integrate it ethically, right? So where do we start? Is, is it the value of the institution that are currently bringing this up? Do you find this is more about frameworks that we can eventually inspire and adopt similar to what has happened in some sustainability certification where we were basically teaching companies what was the right way to basically increase their triple bottom line? So do you see this top-down, bottom-up, a combination of the two? So how, where do you find AI integrity currently happening from an ethical perspective? Method?

I think first, ethical AI, what's considered ethics across different contexts will be different. Give you one example. When I was doing the focus group in Saudi Arabia, the students are saying, we do not have this high percentage threshold for turn it in. So we are allowed to, we don't care about the guide that much. Like we have this trial and error and we want to take control of, you know, we want to learn through our mistakes, whereas in the UK we have a high percentage, you know, no, actually low percentage threshold on turning 18. So if we see there's any issues, lots of meeting with academic integrity officer. First, I think ethics is also kind of contextual. So when we're talking about AI competency, we need to say, well, if you are coming to this context, you look at their local guide, look at the institutional guide, make sure you're aware of, you know, what is defined because it's a socially constructed concept. But also in terms of top-down or, you know, bottom-up, I think there is a middle ground that you need to meet. So my own experience, a lot of AI leaderships talking about ethics AI, they do not have much of teaching experience or they they had teaching experience, but because of years of leadership, their teaching got reduced. So they don't have much of current teaching experience, what is on the ground anymore. Whereas teachers, because of the barrier, like it's sometimes difficult to get our voice heard. So I think that, you know, for both parties, it's very difficult. It's very important to find the middle ground where the voices could be heard. And another thing is when we're talking about ethical AI, we need to train the students. We train, also need to train the staff. So it seems like, oh, it's all about the students issue, but there's lots of bridges, believe me, that staff are using it, creating exam papers, marking students' essays. We just haven't talked about that openly yet.

Yeah.

So I think the training that for should be for both students and staff, but for staff, especially if you want staff to take on training, engage with community as well as teaching students about this, we need to build in the incentive incentives. So at the moment, work allocation model does not recognize— I mean, at least the people that I've talked to, educators— if you want to engage with AI dialogue, go with your own time. I don't know with our audience today, how many of you come here, your time is recognized there. If you're working in education, you know, institution, how many of your time has been recognized by saying, this could be used for your promotion because you engage with, you know, dialogue outside and you are developing your knowledge and you're having dialogue. But many of the educators don't have this luxury. They are giving more teaching and then more AI skill capacity building. And then on top of that, as Mark said, there's a lot of anxiety because the more training you have, somehow we see a pattern is the more anxiety you have. Because I'm quite worried that people who are at the forefront fronts of AI, you know, what we call AI world, right? You see daily update, you see news, you see people doing innovative things, you see people with unethical use. Every day you are surrounded by this information cocoon in a way that reduce kind of your mental health, digital well-being, I would say. So there's also, I would say, education department make sure that staff are incentivized for workload building, but also make sure there are supporting mechanisms for people who are digitally fatigued to protect their digital well-being. I don't know whether that answers—

It does, it does, it does. I also realize a lot of the answers that you both are providing are material for even more conversation that we can constantly have and engage. And, you know, similar to, I think, the spirit of the forum is to create a space for conversation to start. We start them, but eventually they continue also outside of our perimeter. So I will— I'm respectful of your time, Kadi and Kai. I want to ask you the very final question, the same question for both of you. It's coming from Ivan, and I think it's a question that I can use for both of you. So if you're thinking about about kind of the integration of AI in education. And it's both for you, Cady, working with DeepMind, and for you, Skyler, working as an educator, right? So do you find then somehow inevitably we will make some teachers obsolete? Do you find it is more about a reforming of the teaching role in a way we had never seen before? Cady, you said it rightly, right? This is a generation. Like no one— sorry, this is technology like none of the technology before, right? So where do you see the trajectory going? Do you see this as in sync with technology so that we're simply expanding our capabilities? Do you find it that inevitably we're going to make some teacher obsolete? And I'm going to ask both in terms of your work as you see this from a technology company and your work, Skye, as you see this as a professor. So same question, both of you. Kelly.

So, I mean, that's a very good question, but I think teaching as a job would probably be the last one to get obsolete in society. Um, we have a severe shortage of teachers in the world, and teachers are extremely burnt out and overworked. And in most parts of the world, the teacher-to-student ratio is very, very poor. Um, so I don't think the concern here is teachers' jobs being taken way, uh, a bigger issue that we see in all our school interactions, or we see in, uh, policymaker interactions, is the shortage of teachers, or is the shortage of teachers leading to lack of personalization in classrooms. So AI brings that kind of opportunity that— this is where I am personally very optimistic about, um, being able to have an opportunity to personalize learning for students where a teacher can now have more of their time freed up to orchestrate these things in classroom and focus on higher, higher cognition skills, or focus on motivation for students, or help them in their personal growth in a way that they're currently not able to. And that's the kind of future I am optimistic about, and I don't think educators is a job that AI will ever take away.

That's— well, as an educator, I thank you for that, Cathy. You know, it feels good to hear it. Skye, over to you. Your same question about the obsolescence maybe of our job or our profession with more technology. So yeah, your take on this, and it will be my last question to you.

Sure. Funny enough, we did a survey and one of the qualitative open questions for students to say, what do you think AI could replace, you know, the jobs and making job loss, etc.? And we got a very funny quotation from a student saying, professors, I'm sorry, your job is the first to go from the student's perspective. But working in this area for many years, I'm less concerned about, you know, people losing jobs as a result of AI, but more kind of people have to change the job and their roles as a teacher and how people react to that. So I realized that, you know, the research suggests that the first level digital divide is, you know, sometimes the digital infrastructure, but also not interested or the fear of technology is another level. So when we have this, you know, against the enemy type of AI, it will make it either us or them. But sometimes we are already co-design teaching together, and then we need to think about how we could— what is the core value of the teacher? As you said, knowledge is not transactional, it's more kind of relational. So in a way, what is our role changing when we're adopting AI? I think 6 days ago it was news from Japan, they are using AI teaching assistant in the classroom for the future of AI in Japan. I think that a teaching assistant is called seiya, if I'm not wrong. So what the teacher does is like teacher ask still the same classroom with students, teacher ask questions, what their view about this painting or this picture, students explained. And then this seiya AI-generated tutor, right? The teacher will ask this teaching assistant, AI teaching assistant, what do you think? So it comes with different diverse points of view. And then the teacher asks students, what do you think Seiya's point? So it becomes rather, okay, it's Seiya replacing me, but I'm orchestrating everything, but also looking at authentic safeguarding. What if Seiya says something ridiculous? I'm there to safeguard. And also my pedagogy is facilitating, make sure that, oh, I may not think about this perspective, so yeah, does. So teachers still there, the role is still there, but extra workload in terms of orchestrating the whole thing. And again, come back to my point of workload allocation. If schools would recognize the changes that required people's time and effort, offers and then pay, not like pay, but in a way to incentivize whether working as a promotion or giving them more time, adjust the, you know, work type will be very useful. But I definitely think just immediately going to see AI is the enemy taking our jobs is not a good point at this stage going forward.

That's right. Well, Skye and Kelly, look, I realize that the only thing that can stop this is the fact that time is up, but we could continue on this conversation for longer because it's simply so interesting and it requires engagement and ownership of the dialogue. And hopefully, as you said, Kelly, this is not only in a forum like this where it's easy to feel comforted by the fact that we're like-minded and we're working somehow, you know, independently but interdependently in the same direction. But it's true, we like to see this a conversation that happens also outside of our perimeter. So this was extremely educational for me to learn from the work you do. It's a good way for us to continue to engage with people from the field who are pioneering this conversation. So Cathy, thank you so much for being part of this conversation. I hope this is only the beginning of engaging with our community, and I wish you all the best as you continue your work with DeepMind. Sky, very nice to know more about your work. You are one of the first of the second cohort of Microsoft Senior Fellows that we are basically bringing on, on the forum. So it's a partnership that we created for quite some time now where we really want to learn about this incredible community of scholars who are coming and working for the AI Economy Institute. So ladies, thank you so much for making time. I know that it's never easy to find time during the week for this, but Cathy, thank you so much. Enjoy Montreal. Skye, thank you again. Enjoy Manchester. And I'll see you hopefully in another event. And for anyone else who's been watching us in the next episode of Micro One Forum in the near future. Thank you so much, guys. All the best. Cheers. Bye-bye.