Welcome to Tech Talk by

spent a summer doing some education consulting work with the foundation. And as a part of that, I was lucky to get to meet a lot of different, school leaders, so folks running charter schools and, just kind of involved in reforming education in in a variety of ways. And it was really interesting to learn from them. And it was also interesting to kinda hear some of the questions they had about how can we use data and technology and, you know, AI, etcetera, in education in order to to to try to improve outcomes for students. And, it was really interesting just to hear, it sounded like, you know, AI was something this is 2019, by the way, so way way before kinda Chad GPT. But even at the time, they were talking about it, and it's interesting to kinda hear their, speculations about AI and really just, you know, mentioning of AI as as something they felt like could really change things for students, you know, personalized learning, etcetera. But but it didn't seem like there was first of all, just a lot of it was really kind of this buzzword, I think, that was in the air at the time and obviously still is. And so there's, I don't think there were a lot of concrete ideas necessarily about what to do with it and and even what it was. And so I think, it was interesting to see this excitement about it, but, you know, maybe not, some of the the deeper understanding of of what's going on and and and really also the limitations. Right? I think that that's something as as education experts, you know, they were actually much better suited than the AI people or the engineers to really understand what could and couldn't, be useful in their context, to to serve students. And so it just felt like there was maybe an opportunity to try to try to have a little paper that tries to bridge the conversation between people in education and and people in engineering to clarify some of these mystical terms like AI and and, see if we can sorta start to get more to the notes and nuts and bolts of how this could be useful for for learning. So speaking of this mystical term AI,

Yeah. So, it's interesting. I think right now, everything is about generative AI, of course, right, with everything going on with Chatt GPT. And, you know, what what generative AI can do is is pretty amazing, I think, as we've seen. Right? I think it very quickly, has really kinda changed our our thinking about what's possible with AI. But but it's not the only type of AI. Right? I mean, AI has been around for a while. It's really we try in the paper to present it as a loose umbrella term, you know, with a lot of different kind of schools of thought or or different kind of subcategories. One of those subcategories, of course, is machine learning, which generative AI is is a subcategory of in turn. The idea behind machine learning, of course, is you have a bunch of data and you're basically trying to, teach machines to kind of identify patterns in those datasets in order to do some downstream task, like, you know, identify if a picture has cats in it. Right? Or identify if if, you know, certain attributes are present in videos, etcetera, etcetera. So, so that's kind of the dominant, I think, right, especially right now, dominant understanding or school of AI. But there's a lot of other approaches to to AI. We we try to talk in the paper about, rule based AI. So maybe AI where, the goal isn't to learn the rules from the data, right, which is kind of what machine learning is about, but where you kind of, explicitly state certain rules and then, have a machine go and try to find answers. And so an example of that in our own work, actually, that's relevant in education is, you know, school redistricting. Right? So here, you know, the, a school district might wanna redraw something like school attendance boundaries or or do some kind of assignment task like that, and they're trying to basically figure out how do I assign neighborhoods to schools in order to achieve some goal. You know, maybe it's to reduce segregation. Maybe it's to balance out capacities at schools, etcetera. There, you state the rules. You know? You say this is my goal. This is these are the constraints you should, respect. Like, hey. You know, family shouldn't have to travel too far when they're being reassigned to schools, etcetera. So you're kinda stating the rules there, and then you're just having because there's so many possible solutions, there's so many ways to assign neighborhoods to schools. You have a you have a machine basically help you kinda parse through that solution space and identify different assignments that could really maximize your goal while still respecting the constraints. So there, you're not learning some rules from data. You're you're really just trying to state the rules and then see, what's the best solution you can find. So it's a little bit of a different approach. And, of course, these two can be kind of machine learning and rule based AI can be brought together. But we try to just distinguish this partly because, it felt like if we have if we're only thinking about AI one way, then we're probably only gonna think about solutions or improvements in education one way. And so if we kinda broaden what we mean by AI in this context, maybe it'll help us be a little bit more creative about some of the different ways we can actually try to positively impact education, positively, impact learners and in this process. And, you know, one of the points, that we try to make in the paper too is that the eye in AI is not actually that impressive. Right? And and I think there's been a lot more, in the past year or so around people really probing what's going on, you know, on under the hood of a large language model. How is it actually learning, quote, unquote, learning things? What is it doing? And I think what's interesting is that, machine learning systems are really kind of glorified pattern recognizers. Right? I mean, they're really trained on, a simple task. Right? Like a chat g p t, roughly speaking is let me see conditional on some set of words and phrases if I can how how well I can predict, you know, the next word or the next sentence. And and, you know, from that, you're able to generate really kind of impressive pieces of text and other things. But at the end of the day, the the task there that the model was trained on is a very simple task. It's not actually, engaging in abstract reasoning necessarily. It's not doing a lot of the things that people think human do humans do when they think about, human intelligence. So, I think part of the reason we wrote this too was to just try to bring a little bit more clarity to that. Not that it that changes sort of what the outputs of these models can be, but at least makes us feel a little bit less enamored by them and maybe able to focus with more clarity on on, on what's actually going on and how we can use them.

Yeah. It's a great it's a great question. I mean, I think there's a reason to be excited right now. Right? I I think there's a lot of great opportunity ahead. There's a lot of problems in education, and, certainly, AI is not gonna solve all of them. But but, you know, it has the capacity, especially some of the the emerging capabilities with generative AI, to to help, you know, reposition or reframe some of these things so that we can make a dent. I think one of the things, and I think every improvement kinda comes with its its risk as well. It's kinda to your point. But, you know, one of the, opportunities are, for improvement, I think, are one, you know, expanding personalized support. So this is kind of the, I think, probably the main thing that that people think about when they think about AI edge AI in education. How can AI kinda meet students where they're at, understand sort of what gaps they have in their knowledge, and then really work with them to to fill those gaps, you know, by recommending certain problems to work on, by by even coaching them through certain problems, which, you know, Khan Academy has a new tool called Conmigo, which is really kind of this advanced, AI tutor that's that's trying to to really meet students where they're at and and offer them some of these pathways to to higher grounds of learning. And so so there's there's an exciting kind of possibility there. I think another exciting, improvement area is really in helping, you know, very broadly helping people in education kinda cheaply explore counterfactual realities. And and what I mean by that is really using AI as a tool for simulation. So for example, you know, teachers need to, get better at teaching by just working with students. Right? You have to practice. And it's kind of dangerous and high stakes to make teachers practice, in a live setting, right, on on real students. And and, you know, at some point, you just gotta do that. But wouldn't it be useful if we could somehow get them authentic kind of practice, experiences that don't, you know, require them to try things out or use students as the guinea pig. Right? They're able to kinda build that teaching muscle before they go in front of a classroom. So so a lot of interesting work right now, I think, on, really assessment and feedback for teachers. How do we kinda simulate teaching and learning experiences for teachers using some of these these generative AI and other tools, in order to make the real thing as good as possible before, teachers can turn to the real thing. And and I think another thing again, just just sort of drawing a little bit back on on our work, you know, exploring hypothetical policies. Right? So, again, we maybe we don't think about, education policy as kind of where, AI might might, impact education. But if I go back to the example from earlier, you know, redrawing, school attendance boundaries, which are responsible for assigning the vast majority of students to to schools across this country, can we try out different policies, right, with different goals? If we wanna maximize in, racial and socioeconomic integration, if we wanna balance out school utilization, all of these things have implications for the equity of student opportunities and and outcomes. And so can we can we use AI to simulate possibilities before we actually pick one and and, you know, go to the community for their feedback and eventually, implement them. So I think those are some exciting, possibilities. I think, of course, there there's a lot of risks, I think, kinda going to the personalized support piece. You know, one thing I worry about is we might be able to build an amazing chatbot tutor. And and, you know, in some ways, it's probably worthwhile and really kind of exploring how we do that. But a lot of learning is about the emotional connection that students feel with educators. You know, the mentorship that they receive, can they see their own story and somebody else's story? You know, chatbot doesn't have a story. Right? And then so I think it's it's, in some ways, it's scary to imagine not not to be a a a doomsday, you know, sayer here, but I I think, we kinda play this forward. You we might start having really amazing robot tutors, but then will that will that keep us from, connecting students to to other human beings, mentors that that might be able to provide similar offerings maybe at less scale. But that offers something that the chatbot can't, which is which is, you know, inspiration. Right? Helping helping a student see what's possible for themselves. So I think, those are that that's sort of one concern. Doesn't mean we should invest in those technologies, but I think it's about kind of, you know, mitigating them and and trying to trying to, anticipate those upfront. And then I think the other concern I see is that, we are really, almost obsessed, I think, right now with personalization and and with tutoring and and with a lot of these kind of within classroom applications of AI. And, again, those things can kinda go a long way in improving student outcomes, But learning is a complex system. It's not just about what's happening in the classroom. It's about what's happening out of the classroom. It's about what's happening at the policy level. And I've seen less discussion right now, kinda thinking about how how AI can be used to to support policy making, maybe in some of the ways that I've I've talked about a little bit earlier. And so, I worry that if we focus too much on the narrow problem of of teaching and learning, which itself is a broad problem, but if we just focus on the within classroom, we might actually lose sight of some of the the higher leverage opportunities to really shape and and improve students' outcomes. Right? If we change things at the policy level, if we change things at the social network level, kind of the home network level, can we even obviate the need for additional tutoring or, you know, the achievement gaps that we see kinda down stream, right, that a lot of these AI tools are are focused on right now. So those are a couple of the I mean, some of the the broader risk. I could keep going on and on, but I'll I'll, I'll stop myself here.

Yeah. So so it's kinda it's interesting. Deep learning has a we talk about it in kind of the the the machine learning AI sense. It it, of course, has a dual meaning in the education sense. But, when when we talk about it in in the the technical sense, you know, deep learning is really this advance in machine learning. It's kind of been around actually for decades, but people couldn't figure out how to apply it practically until about a decade and a half ago. And and that's really when this kind of new AI revolution really took off. And and, you know, transformers, which is what the t in GPT stands for, is is a type of, a neural network which which leverages deep learning. Like, these things have now enabled things like JET GPT and some of these other advances that we're seeing. So it's this advance in in how machine learning is done. You know, typically, a machine learning model, like, let's say you wanted to you know, one application in education sometimes is trying to predict, whether a student is gonna is at risk of dropping out right of high school. So this is something that that, schools and districts will will, use models for. And so a typical model would kind of take in a bunch of student data, and then, use that those inputs to try to predict some output like a binary guess of whether they're gonna drop out or not. Deep learning is is interesting. And, you know, that that prediction, I should say, in a classical machine learning sense is is really sensitive to how those inputs are included in the model. Right? So, for example, maybe one of the inputs is the student's GPA. You know, should we, there's a lot of questions when you're kinda designing those inputs. You know, should we just include the raw GPA? Should we include, the GPA as a representation of how much it differs from, you know, the average GPA in the class? Right? These all kinda seem like nuance to kinda details of data representation or what's called feature engineering. But they actually can have a big impact on on how, a machine learning model, makes its predictions and how accurately it's actually able to learn those patterns from the data. And so deep learning, one of the things that it it's it's brought is really what it does is it takes a lot of these kinda smaller, machine learning models, so to speak. And and often, they're neural networks, which are basically a type of machine learning architecture that are trying to learn some of these patterns in the data to do some tasks like a like a prediction we're talking about. They take these neural networks, and they basically kinda stack them on top of each other so that the, you know, one neural network gets some data. It it does some transformations to them, right, to kind of, massage those features or those data points into something that will help the model make a better prediction and then passes that massage data to another neural network that does the same thing, etcetera, etcetera. And the idea is through these kind of successive layers or or this successive, like, input output sequences, you're able to the models are kind of able to or they're trying to identify, patterns and and really transform the data into something that, they're gonna be able to make accurate predictions with. So it tries to, like, reduce the one of the many things it does is tries to reduce, the burden of feature engineering, from the the model developer and and sort of automate that to the extent it can. And and a lot of these models are able to do that really well, which is why, you know, we've seen these kind of crazy advances, especially for something like text. Right? Text data is very complicated. The the you know, it's very nuanced. Right? You could put a word at the end of a sentence that completely changes the meaning of, you know, the whole sentence itself. And and so, because of neural network's ability to to do this kind of pattern extraction in this way, they they've they've gotten, really good at at even handling data like text, which other models just aren't able to do well.

Yeah. It's a great question. I think I think, you know, we talked a little bit about personalized learning and tutoring. You know, the goal, I think, of many of these systems is actually to, like, mine a bunch of data on how students have answered certain problems, like math problems, and then, identify problems that, that are kind of at their learning edge. So what are the problems that we think would be a little bit of a stretch for a student, but that they're likely to get right? And that's kind of one way to create, like, a problem recommendation engine. Right? So this is how, I think a lot of, kind of personalized adaptive tutoring, adaptive learning systems to date have worked. I think the limitations we talked about there as well. Right? Like, what is the kind of role of the human? Of course, the goal isn't to replace the human or the teacher, which is what most folks say and I think earnestly believe. But but I think the the risk is that, you know, it's not if we focus so much in designing these really accurate algorithms, like, there's no guarantee that we will champion what the human has has to offer uniquely, right, beyond the machine. So I think that's kind of a interesting what is the true human AI collaboration around a system like this? I think that's an open question. You know, I mentioned early warning systems. Right? Like, hey. We wanna can we try to predict if a student's gonna drop out or not? Right? Can we predict if, how likely they are to, you know, maybe maybe, engage in kind of a disciplinary infraction or or whatever it is? You know, one, of course, there's kind of the issue the same issues of bias there that that we face with any other system where what's the data we're actually feeding to the model to make these predictions, and and how much should we trust trust that data, right, and trust those predictions. I think that's one thing. We might find that, we might identify features about the student like, you know, their grades or, you know, maybe even their their free and reduced lunch status or whatever it is that, correlates with their likelihood to drop out or not. Right? So we can identify great correlations. But those systems don't actually tell us what's causing those dropouts or those potential dropouts. Right? So it's classic correlation does not imply causation. And this is like a general problem with machine learning. I mean, machine learning, is really exploiting a bunch of correlations or relationships between data points, but they don't really tell us, what's what's underneath the trends that are being, observed or produced. And I think that's an issue because then we don't actually know how to intervene. Right? Like, we might know that these students are at the highest risk of dropping out. But if we don't understand why, then how how good are interventions? Right? There's still a lot of work to be done to design good interventions. And I think the other thing with early warning systems that's a risk is that when should we intervene? Right? If we get a indicator that a student is at eighty percent risk of dropping out or even, like, let's say and you say they're at at at forty percent risk of dropping out, where do we draw the line? Right? Like, are we okay? Is or do we say forty percent? Oh, that's low enough. We don't need to intervene. Well, what if something happens? Right? Or do we say, oh, eighty percent. We definitely inter need to intervene. But then what if that leads us to blitz a lot of attention and resources at that student that might somehow harm them. Right? Maybe maybe there there's kind of an, type of surveillance almost that happens to certain kinds of students if if if we if, depending on where we draw that line. So so it's really it's really unclear, I think. There there's a lot of human decision making that still needs to happen to to put these things to use. So yes. I mean, these are a couple I think I think there's an interesting application in, you know, coaching and counseling as well. So this is, again, you know, you can imagine there's students that have a lot of questions, like, often transactional questions, like, hey. How do I fill out my FAFSA or, you know, these kinds of things. Right? That guidance counselors might help with so there's been kinda efforts over the years to use chatbots. You know, chatbots also to send reminders to students to make sure they submit their financial aid documents so that they can start at college. So some of these these simple interventions, these simple tools have had a a big impact on on, student outcomes. But I think there's a question there of, well, what's the right task for a chatbot? Right? Maybe if it's closed ended, there's kind of a very clear answer. That's something a chatbot can can handle. But then if you start to get into open ended questions, you know, issues of mental health, etcetera, we probably really want that student kind of engaging with the human counselor in some way. So I think every one of these, applications has sort of its its benefits, but but possibly also the risks.

Sure. So just some of the questions we we suggested. This is really, you know, if you're a teacher or a district official or a kind of a practitioner in education, and people are you're reading about this tool in AI in education or people are coming to you trying to sell you something or get you to use something, what are what are ways you can kind of critically engage maybe in in a dialogue around that? So that was kind of the purpose of offering these questions. And and so I'll just very quickly, you know, mention some of the questions we suggested were things like asking, what kind of AI is it? Right? What exactly is going on here? Reason for that was, you know, there's a lot of tools that are saying they do AI, you know, in education. But what does it mean? Like, is it is it if thens, if else statement in code? Does that count as AI? Right? So I think I think just really pushing, like, what do you mean? What what is this tool doing? How is it actually, using AI? Can at least start to kind of maybe sift through some of the snake oil, I think, that that's kind of out there right now, especially in in some of the products that are trying to sell to schools. You know, another kind of related question was, does the AI enable something that would be difficult or impossible to achieve without it? I think it's really interesting. You know, we can kinda think pre generative AI. This seems like a really, applicable question, I think, because there's a lot of things that, an AI system like problem recommendation engine could do, but, you know, a much simpler system could do that too. Right? You don't need some fancy, you know, machine learning system that's trying to predict the problem that a student's gonna get right and and and suggest that. Like, you could just have a simple engine that this recommends, you know, the next thing it makes sense for someone to work on. And there's been studies, I think, that show that those kinds of systems are also very, very effective. So, so what is the AI really getting you that you couldn't have before? You know, I think, obviously, with generative AI, like, I don't think we've we've lived in a moment, obviously, where we can generate the things that are being generated right now with chat GPT. So I think I think there's something about, now kinda looking out a couple years after we wrote this paper, that certainly there's a lot that can be, produced by AI that would be impossible without it. I think there's still an interesting question of what is the outcome for students that those things are driving that couldn't be achieved without it. And the reason I say that is because there's clearly a lot of outputs that are being generated. Right? Again, just what a what a chat g b t can generate, how it's able to summarize a document for a student, how it's able to rephrase, you know, parts of an essay for a student. There's all these things that these tools are able to produce as outputs that we just couldn't produce automatically until very recently. But I think there's still a lot of open questions about what is that actually enabling for learning? What is that enabling for the child? What is that enabling for the family, right, in terms of educational outcomes, in terms of quality of life outcomes? I think the evidence there is still extremely thin, and I think that's also why there's really a need for more, you know, more efficacy studies. Right? Really understand the impact that that these new tools can have, not just the outputs they can produce. You know, another question was what are the potential risks or drawbacks of deploying these technologies? I think, you know, for every action, there's an equal and opposite reaction. Right? So I think we can understand the benefits possibly of AI, but also worth understanding what what are what are the costs, what are the, you know, some of the biases that we talked about that might be perpetuated. You know, what are what is the you know, will the human be taken out of the loop? What will that mean for the student? Right? So there's all these sorts of things, I think, that we can, we and and particularly education practitioners might think about. The the the fourth question was, how equitable how equitably are the anticipated benefits and risks distributed across different groups of students and families. So this, of course, is really important because, again, a system a personalized tutoring system could work really well for a student who, is already very, you know, fluent with technology, who kind of already has a a strong starting point. Right? Doesn't maybe doesn't the issues of inspiration and motivation, those are not issues possibly for this particular student. They just need they just need something to kinda stimulate them. Right? So, sure, maybe those tools can be useful for them, but maybe there are other students that are facing a really tough home environment. You know, there's a there's a lot of support that they might need to kind of beyond the pedagogical or the cognitive support that a lot of these tools seem to be focused on right now. So how can we think holistically about the subgroups of learners and families instead of just picking tools based on how well they serve the average or even worse, how well they serve maybe, like, a specific, you know, small affluent well off subset of the population? And then the last thing is is, if you could wave a magic wand and change anything about this technology, what would it be? You know, this is like a typical kind of user research type of question. Right? If you're doing a user study and you're trying to understand, the value, or lack of value people might see in in your product, you often ask the magic wand question. And I I like the question because I think it really, helps people, see that they can change things. Right? And and I think that's really important when you think about educators. There's so much that's being handed to them right now because maybe many of them are not the ones developing these technologies. And so people are developing things and kind of no. Ideally, in tandem with educators, but sometimes they're developing things and then they're kind of handing it off to the educators to to use. And and I think it's really important that educators, again, see that, whether they're building these tools or not, they have kind of a unique vantage point in education on what learners and families need and can benefit from. And so, you know, they have every right to, I think, kind of, question. Right? Like, what what how could this be different? How could we make it better? Right? Because in some ways, they are the voices often of the students and families. And so, yeah, I I think I think that was that was the hope and kinda the the overarching goal of adding this question is is to try to help educationally see that they have a lot of opportunity and power and, frankly, responsibility to,

No. I mean, I appreciate the the chance to to share and all the great questions. I think, you know, maybe it's just reemphasizing something from earlier, which is that, GenAI is extremely powerful. I think there's a great, frontier ahead in in thinking about, what it could unlock for, you know, students, teachers, and families. I think we can also keep in mind AI is, again, this broader umbrella. It's not just generative AI. There's a lot of other things that can it can do for us, and it's not just about the classroom. Right? I think the classroom is is powerful. It's a it's a critical inflection point. But, you know, policy, families, right, parent networks, how do we how do we empower parents? How do we increase family, engagement in schools? How do we foster new connections for students? Right? I mean, these are not things that feel like AI problems right now. But but I think I think if we think about it, we're gonna think of really exciting and interesting ways tools like AI could even serve some of those areas. And I think if we do that, we might even have, a bigger impact, hopefully, in in education and on people's lives than, than if we focus on pedagogy alone. So I guess that would be my plea for anybody who's listening and is is interested in this topic is is to explore some of those other areas as well. Awesome. Gordon, Bill, it's been a pleasure having you with us. Thank you so much for joining today. Thank you so much for the opportunity. Of course. And to all our listeners, to keep up with all the work CDT's policy teams are doing, please visit us at cdt.org