How AI Is Fueling an Existential Crisis in Education  ## Metadata - Author: [[Decoder with Nilay Patel]] - Full Title: How AI Is Fueling an Existential Crisis in Education - Category: #podcasts - URL: https://share.snipd.com/episode/5ca2832a-89f5-4f6c-b867-9889fb3000dc ## Document Notes ## Highlights - **Digital Natives Don't Exist** - "Digital natives" is a myth; familiarity with tech doesn't equal ability to use it for learning. - Teachers wrongly assume students know how to use tech educationally, which causes gaps in instruction. Transcript: Nilay Patel Yes. Can you briefly explain what you mean by that? Dr. Adam Dubé The term digital natives was coined back in the early 2000s. And it was this idea that perhaps kids that are born into technology understand it better than, say, what he called digital immigrants. And there's been a lot of problems with this type of language and the framing about how he talked about it. And then there was 20 years of research to see if this is actually true. Are kids that are born growing up with technology better at using it than people that adopt it later on in life? And the research for 20 years has shown this isn't actually the case. It's not that because you're young and you grew up around it. It's just about how much you've used it, how much exposure you've had to it. And this really matters when we talk about AI and education and technology and education, because even though a kid grows up using YouTube or a phone for playing Roblox, it doesn't mean They know how to use technology to actually learn, but we assume that they do. Teachers assume that kids know how to use technology in the classroom. They assume they know how to use it for learning purposes because they use it for YouTube. And so we aren't digital natives. We just have previous experience using technology for specific things. ([Time 0:03:06](https://share.snipd.com/snip/26652265-e14b-439c-98df-595d5f020fa2)) - **Language Interfaces Shift Required Skills** - Natural language interfaces change how we interact with computers but underlying logic still matters. - Students may not interpret or evaluate AI answers correctly without understanding system logic. Transcript: Dr. Adam Dubé Think the way that we interact with computers could be changing by having us engage with it through natural language interfaces. Unfortunately, the logic that underlies that system is still really important for being able to interpret the answers that it's giving us. And so, yes, kids, people growing up using a computer where it's, say, primarily a text-based or a voice-based system, they're not going to think of it the same way than someone who grew Up with a file system and engaging with individual applications instead of everything being launched through ChatGPT, say, for example. That's going to be a problem not for them using and interacting with the system and asking it to do things, but then how they actually interpret the way those systems give them answers And how they evaluate it. Can they actually make sense of these responses and then make critical judgments about it? And so actually, this is an area of research that myself and my PhD candidate Nundini are working on where we're looking at children's theory of artificial minds. We're trying to understand how do children think computers think,, how do they think that AI reasons, if we can say that it reasons? And then what's the impact that that's going to have on how they learn from AI that's put in their schools and in their homes, like smart speakers that are already everywhere? ([Time 0:05:11](https://share.snipd.com/snip/6b9c89d5-b8cf-4969-a814-0b07d9d49d6b)) - **Teacher Sees AI Replace Student Thinking** - Anne Lutz-Fernandez observed students using generative AI to replace reading, thinking, and creative writing. - She warned we treat children like guinea pigs with untested educational products. Transcript: Speaker 3 Lutz-Fernandez. During my last year of teaching, I began to see more students using generative AI to replace their own reading, thinking, and writing, even creative and personal writing. We're treating children like guinea pigs on an untested and unproven and unregulated host of products. It feels to me like we haven't learned some key lessons, a lot of them very recent. One of those during the pandemic was the costs of unhuman teaching and learning. ([Time 0:06:50](https://share.snipd.com/snip/55ae10a2-cd7c-45e5-94ce-8333ffaafac6)) - **Easy Interfaces Can Hide Poor Learning** - Many learning apps oversimplify interactions so students guess rather than learn. - Seemingly easy AI tutors risk creating passive, shallow learning despite appearing effective. Transcript: Dr. Adam Dubé Technologies were introduced. I was a math cognition researcher that looked at how children understand simple things like learning how to add. And then I got into studying technology because Apple came out and was pitching the iPad as the future of education. And then a bunch of math apps were launching that were saying, okay, this is how your child's going to learn math the best. We were actually testing this early on, back in 2011, 2012, we were giving kids a bunch of different iPads with learning apps. And you would think that the kids knew how to use them, but 90% of the interactions they had with that learning app were actually complete mistakes and errors. They were just randomly tapping around the screen. And so there was a lot of guessing, but the apps actually had no negative consequences for getting stuff wrong. So we call this as just being the app was too dumb to cause a mistake. It didn't matter if you interacted with it in a random way. It always just progressed. So this sort of idea that kids get how to work with technology is actually a byproduct of an oversimplified design of a lot of the apps that are used for learning. And then now we've got new systems coming in where it seems like, well, kids can just talk to AI speakers and they can just talk back to them. And an example of this is that video that Sal Khan put out with Contmigo Tutor. It's like, well, look, I can just sit my kid down in front of this. I can say, teach him how to solve this math problem. And it just does the thing. Okay. But is that child actually benefiting from that experience? Are they interpreting ([Time 0:07:50](https://share.snipd.com/snip/3068670b-4ea5-4f71-bde5-f65c2ae5e5e0)) - **AI Use Is Mostly Assistance, Not Full Cheating** - Most students use AI to explain concepts or generate ideas, not to produce whole assignments. - Only about 10% report using AI to generate entire assignments, a stable cheating baseline. Transcript: Dr. Adam Dubé Students and what matters, not just if they're using it, but what they're using it for. And so the biggest use of them is that students in high schools are using it to explain concepts to themselves. So that's 80% of them. And that's gone up from a year previously. They're using it to generate ideas for assignments. And that's about 70% of the use cases. And then they're using it to summarize texts instead of reading it. 40 are using it to edit portions of text. And then at the very bottom, and I think this is something that probably people are going to disagree with, is that only 10 percent of students are actually reporting that they're using It to generate their whole assignment, which is what people are really worried about when they think about AI cheating in schools. But that number is pretty consistent and hasn't really changed over the years. The percentage of students that report actually cheating has stayed around 10%. It's just how they cheat changes over time. So paying someone or copying stuff off the internet and then now. And so the numbers that you say reflect what we see, but then we're always like, how are they using it? They're using it in different ways for different purposes. And then we can debate whether or not these different uses are even good whatsoever. Obviously, cheating is bad. But is it good at summarizing text? It's not. But, you know, that's where we can get more nuanced questions about these different uses. ([Time 0:10:09](https://share.snipd.com/snip/68d98536-ac3e-485f-b616-a10d8923bdca)) - **Don't Force Teachers Into Specific AI Tools** - Avoid top-down mandates forcing specific AI tools on teachers; that demotivates staff. - Let educators choose tools and maintain autonomy to preserve motivation. Transcript: Dr. Adam Dubé And that leaves it up to the factors at the local level influencing whether or not AI is seen as a potential positive or negative, and whether or not it's a positive or negative for teachers, The admin, or for students. There's also differences there. A lot of teachers think students shouldn't be using it, but it's okay for them to use it. Or the admin thinks this is going to help us save time for teachers marking students' assignments so we can save some money there. But we don't want our students using it, but we're going to use it to analyze student data, right? So there's even a mishmash and a disagreement within schools about the role of generative AI. Right now, these systems are being sold to educators to generate lesson plans, to evaluate student work, to do learning analytics. And if you're having this deployed in your school and before you teach a class, you're being told, it was like, okay, well, we're going to cut back on how much teacher preparation time There is, but we bought Magic School for you. And so it's going to generate a lesson plan for you. So it's, don't worry, you're going to have plenty of time to do it. ([Time 0:15:43](https://share.snipd.com/snip/af71b191-e5ee-499e-8e80-f44e8f73b135)) - **Machine Translation Hallucinated History Texts** - Anne Rubenstein tried publisher machine translations and found AI hallucinated added sentences. - Fixing those errors with human translators cost roughly twice hiring a human translator initially. Transcript: Speaker 3 My name is Anne Rubenstein. I'm a historian and a professor of history at York University. Documents from the past on specific topics that are then published as part of a digital history project that goes out to university libraries mostly. Because I am a historian of Mexico, the documents that I'm preparing for them are Mexican and they're in Spanish. The publisher decided that we should, along with providing the original documents, we should provide translations into English, since that's the language that the majority of people Using these teaching tools are going to be comfortable with. So great. I said, great, I've got a friend who's a translator. We'll get them to translate these documents, no problem. And they said, oh, no, no, we've bought new software that will translate for us, and we don't need to go to the expense and trouble of hiring a human translator because this translation Software is going to be great. And I was skeptical, but I said, sure, let's try it. And so we tried the software, and here's what it did. It hallucinated. It made crap up. It inserted entire sentences and in a couple of cases, entire paragraphs into the document that did not exist in the original. That is a very, very big problem in a collection of translated primary source documents for history students, I invite you to come take some history classes and then you'll understand Why that's an enormous problem. Luckily, the publisher also understood this was an enormous problem. So what they decided to do was hire a translator whose job it was to go through these machine translated documents and restore accuracy and clarity to them. And that ended up costing just about twice as much as just hiring a human translator would have done. ([Time 0:18:34](https://share.snipd.com/snip/43de7105-1110-4c31-b850-013f9aebcd0e)) - **Apparent Efficiency Can Be Illusory** - Generative AI can feel time-saving but often requires extensive fixing that erodes efficiency. - Early evidence from coding tools suggests similar slowdown may happen for educators. Transcript: Dr. Adam Dubé Is it actually saving you time? I speak to teachers and they say like, well, I use generative AI and it helps me generate my lessons. It helps me write emails. Actually monitor and try to keep track of, is this actually speeding things up? There's a lot of research that shows that, say, for example, with coders, they actually end up being slower when they use these systems because they have to fix all the mistakes. And I think that we might see a similar pattern that happens with educators. They're using these tools to be more efficient, they think. But if they actually tracked how long it takes them to generate a lesson versus how long it takes to fix the lessons that generative AI produces, it might not actually be faster. I see some educators that are enthusiastic about the time saving it can give them, but I'm not sure it's actually saving anybody's time. ([Time 0:21:01](https://share.snipd.com/snip/1b68a919-ad60-44d3-b0c3-5bcb2d3ab00e)) - **AI Undermines Deep Learning Practice** - Using AI to produce answers reduces effortful practice and memory encoding. - Students may produce polished work but lack internal knowledge to evaluate or apply it later. Transcript: Dr. Adam Dubé Classic example, as you said, is the calculator. And some people say that it didn't matter that we put calculators in classrooms. I actually had a colleague, Joanne Lefebvre, that looked at this throughout the 90s and 2000s in Canada. And we actually saw a decrease in math scores directly correlated with increasing calculator use. Because when you're using a tool to do thinking for you, you're not practicing and actually encoding the information well enough to actually recall it later on. And so it's not surprising that when people use generative AI to do work for them, that they're not able to do it independently. Being able to store information in your memory requires effortful practice. It requires effortful memorization. It requires reflection. It requires thinking about, okay, what am I trying to understand and connecting that to my other understanding? That's what builds a strong knowledge network. And when you use systems that just generate answers on your behalf, you don't engage in those practices. It just gives you the response and you passively consume it, and maybe you don't reflect on it. So it's not surprising that with the use of generative AI, one of the big effects that we see is that people are able to produce maybe work that looks more polished, but they don't remember The work that they actually wrote. That MIT study is the example that a lot of people have heard about where they had people writing essays with or without ChatGPT or Google, and students had very poor memories for the Essays that they wrote using ChatGPT. Well, that's because they actually weren't reflecting on their writing. They weren't engaged in the work that it takes to actually form substantial memories so you can remember it later on. ([Time 0:27:02](https://share.snipd.com/snip/d1060c76-7d5a-42e9-a95f-334b7c7e6105)) - **Explain AI's Limits In Class** - Teach students how AI works and why it fails for discipline-specific tasks. - Use classroom discussion to surface limits and ethical constraints of AI in research. Transcript: Speaker 3 I've worked out to do with the first year undergraduates especially is not so much to tell them that they can or can't use this stuff to help them in their process of becoming history students, But to think with them about the ways in which this particular kind of software is likely to lead to bad results for historians in particular. What I tell them is, as historians, we have a social responsibility to get our facts exactly right. Not only to say exactly when the specific event happened, but if we're quoting to back up our assertions, which we frequently do, we have to say who was speaking and also how we know what Words they said and also to quote their words precisely in the precise order, not to add any, not to leave any out, and to say where and when this quotation was made and to put it in its context. Once we've sort of gone through that with the beginning history students, then I talk to them about ChatGPT and similar software and I say, okay, how does it work? And I pretend to be slightly more naive than I actually am. And I say, okay, explain to me how this works. And usually in any group of, say, 10 or more undergraduates, one of them is going to have a very clear understanding of how this software works. So I get them to explain it to me and incidentally to explain it to themselves and each other. And we talk about how the software can't by its nature actually know a thing. What it can tell you is what order words are likely to be in. And if we're very lucky, what will happen in the classroom as we're figuring this out together and going over it together is that someone will say, oh, so they're bullshitting. And I say, yes, that's bullshit. And then we talk about what bullshit is and why we want to avoid bullshitting and why bullshitting is sometimes useful and important in life, but historians aren't allowed to bullshit. We absolutely cannot bullshit. We are the only people in the world who are never, ever allowed to bullshit. And so then the lesson is you can use ChatGPT and similar software for all kinds of things, but you cannot use it in conducting historical research or writing about history because it Is the exact opposite of what historians are supposed to do. ([Time 0:30:13](https://share.snipd.com/snip/9bf3278f-33d7-409e-828d-369113e873dc)) - **Grades Drive Outsourcing To AI** - Grades and finished products drive student incentives more than learning processes. - Tools that produce polished outputs align with student pressures, reducing motivation to learn skills. Transcript: Speaker 6 Research One university. I have a lot of engineering majors in my classes, and my job is to teach them how to speak about engineering things to non-engineers. The big effect large language models have had on my job and my students is that it's really forced me to recognize how differently I see what's valuable in the classes from at least some Of my students. In my classes, I teach a lot of how to write things. Some of it is format, but more of it is about tailoring a message to an audience, translating concepts from expert to non-expert. But as with most writing classes, the grades are based on the finished product, the user manual or the proposal or the report. I use those to evaluate how well the students have internalized the tools I've been showing them how to use. LLMs promise that they can create those documents without having to learn all those intermediate steps. And when they're being used by a person who already has those skills, they can take some of the grunt work out of it. My students don't have those skills yet. And if they lean on LLMs now, they'll never develop those skills. But for a pretty good-sized portion of my student body, that's not a problem because, one, they have limited time and they have physics exams and so on. And, two, because they get graded on the finished product. For me, the finished product doesn't really matter. I've read enough proposals for free student parking on campus for multiple lifetimes, but for them, it matters because it can affect their academic standing, potentially their financial Aid. And they believe that it can affect their ability to get a job or an internship. The grade matters a lot more to them than to me, in other words. So it makes sense that if there's a tool that promises a product that will help them pass so that they can concentrate on the stuff that they feel is more important to their career, of course, They'll think about using it. That's the real tension at play here. How do we convince students that the value in the class is learning how to do things when the thing we measure is the end product, especially when there's a tool that can take some of the Pain out of producing that product? ([Time 0:33:20](https://share.snipd.com/snip/2f1d1664-c26d-4e41-8044-4ad08cebb039)) - **Prioritize Process Over Product** - Redesign assessments to value process and reduce external stressors that push students to outsource. - Focus support on lowering financial and time pressures so students can engage in real learning. Transcript: Todd Harper Thing I've observed in 15-ish years of teaching at the college level is how much refocus there has been on the product when it comes to course assignments, papers, presentations, whatever. Students are aiming for the grade because the grade is the thing that hooks into important metrics. It's the thing that hooks into whether they graduate or not. Sometimes it influences financial aid, et cetera, et cetera. And students are under tremendous pressure that affects how they approach their college education. My university largely has students who are focused on getting out of here and getting a job. A lot of them work full-time or are full-time caregivers or have some kind of equivalent everyday pressure on them. They're taking multiple courses at a time. They're trying to make it all work out. And if a tool comes along and says, oh, you got a paper due? Just plug the question into me, and I'll give you a plausible sounding result. And then the student can be like, great, that's one thing I can check off the list so I don't lose my mind trying to be alive in 2025. I get why that would have some appeal. But pedagogically, educationally, we don't assign homework, papers, presentations, projects. We don't give those to students because we want something in return. The thing that they give to us is not the point. The point is that when they're looking up sources or drawing the art or creating the thing that they turn into us. They are exercising the skills and the learning that we want them to develop in our classroom, that they have come to our classroom to develop. And yes, we have to evaluate them and we can't be there for the process. So they do have to turn something into us, right? Like the product is how we evaluate the process through the result. But the process is the important bit. And if what's turned into me, you know, air quotes, looks right, you know, looks plausible, which LLMs can be good at. In fact, it's probably the only thing they're air quotes good at, is making things that look plausible. But if the student didn't do it, if there was no process, then what are we doing here? No real learning has happened. All that's happened is that somebody ticked off a box on a to-do list. And I think it hurts students when that happens. ([Time 0:35:40](https://share.snipd.com/snip/27552d54-1113-41dd-a9a4-95aa160fe998))