← Harvard Data Science Review Podcast
Harvard Data Science Review Podcast
ChatGPT in the Classroom: Breeding More Cheaters or Better Learners?
00:34:49
Loading summary
Transcript36 lines
- [00:03]
Liberty Vittert
Welcome to the Harvard Data Science Review podcast. I'm Liberty Vittert, the feature editor of the Harvard Data Science Review, and joining me is my co host and editor in chief, Shelley Meng. Teaching the next generation of scholars is no easy task. And recently the boom of generative AI has changed the game in a lot of ways. We're joined today by our esteemed guest, James Zhao, professor at Stanford and member of Stanford AI Lab, and Liz Rose Shulman, professor at Northwestern University and teacher at Township High School, to discuss the challenges and possibilities generative AI poses for educators. How can we, as the data science community, build models to help early learners? What goes into learning how to think and what is the job of a school? Are we raising a generation of cheaters or a generation of kids who simply can't learn to think critically? How can students and educators of all levels use ChatGPT as an effective tool? Or can they at all? Stay tuned for all of this and more on the Harvard Data Science Review podcast.
- [01:15]
Shelley Meng
Well, thank you so much, Liz and James for joining us. So to start our conversation, I want to just get to the scope on today's topic from each of your perspective. First, let me go with James. Thank you, James, for writing multiple articles for Harvard Data Science Review. But can you just give us a bit more about general AI, why that has become such an element of controversial in the education space from your perspective?
- [01:40]
James Zhao
Yeah. Well, first, thanks for organizing this podcast, Shelley. Really excited for this discussion and very nice to meet you, Liz. Generative AI. It's an extremely exciting time and very fast moving space. So when people think about generative AI, they're mostly talking about large language models like ChatGPT. But now they're also like every week there are new language models that are coming into the workspace, into the market. So just last week Beta and Facebook released their latest Llama 3, which looks very promising. And there are new ones coming online every week. So I think these large language models and other generative AI systems I think can really be transformative for education. I'll just give you one concrete example of this because as a professor, I'm sure Shelley has similar experience. We spend a lot of time reading papers and also reading reviews of our papers. And then we've recently actually been noticing that a lot of the reviews that we get look a bit like AI generated. They look a bit strange. And we actually did a study where we tried to estimate statistically what fraction of all the reviews that people post online of the recent AI conference papers have been AI generated. And our estimates are around 17% of the reviews for AI conferences have been substantially generated or written by large language models. And that's really going up very quickly. I think that just one aspect of this, but really highlights that these language models and generative AI is becoming very rapidly, very prevalent in everything that we see and interact with. Right. And it's affecting how our students are conducting their research. It's also affecting how information is being exchanged and interpreted.
- [03:26]
Shelley Meng
Les, you have this important dual roles because you teach both in college and high schools, which is very unique. So I particularly want to ask you, how do your syllabus policies differ from the college version, the high school, and also how the students react to them differently and anything different between the way, you know, high school students react to ChatGPT and the College students react to it?
- [03:52]
Liz Rose Shulman
Yeah. First of all, thank you for having me. So there is a statement in both syllabi. So my syllabus at Northwestern University, I teach future teachers as well as the syllabus that I use in high school at Evanston Township High School, that we don't use it in the classroom because I'm teaching English and the students are really there to learn how to generate their own original thoughts. In that sense, it's considered academic dishonesty. It's also considered plagiarism because if they're using ChatGPT and they are high school students are all using it, I would. I mean, a large percentage are using it to do their homework when they're being told not to use it. So they're missing out on the opportunity to use their imagination and to create their own original thought. And that's the biggest problem that I see in high school, in college, because I'm teaching future teachers, we're also not using it in the classroom because they're in the class with me to learn how to create their and design their own lesson plans. And so similar to my students in high school, they're learning how to use their own original thought and their imagination to design these things that they can use. So it's actually similar for me in both syllabi in both college and in high school.
- [05:18]
Shelley Meng
Thank you. So what basically do you tell them, like, you should not use it or you can use it to report, but you need to disclose what are the specific policies you have for them?
- [05:30]
Liz Rose Shulman
Well, different teachers have different policies. I mean, there are some teachers who use it to, for example, generate a draft for a paper. I don't do that because the drafting stage of writing is so important. If you consider that writing is thinking and thinking. Thinking is really everything. Specifically in K through 12 education. They're in school to learn how to think for themselves. So we have these conversations and we talk about what it means to use it, why they would use it and why we're not using it.
- [06:05]
Shelley Meng
Right, I see. And so in general, how do you enforce your policy on a code or you have ways to check and I'm just curious.
- [06:15]
Liz Rose Shulman
Yeah, there are some ways to check. There are some websites, like some detection websites, zero gbt and there's also GPT zero.
- [06:27]
Shelley Meng
I see.
- [06:28]
Liz Rose Shulman
But you can, you can put the text in and it will give you a percentage that was generated online. And we talk about it as plagiarism because what they're copying from online, it's actually coming from other people who have written things, you know, and there have been lawsuits by writers not wanting their stuff to be copied in that way. And so yeah, I do try to enforce it. The problem is that when you're a high school teacher and you have 125 students, you're plugging in this text into these detection sites and it's a tremendous amount of work. And so I'm trying to nip it in the butt at the beginning of the year and talk to my students about that they deserve to be coming up with their own original thought and that they're capable of it and that if they don't do that, they're not going to have the skills to be able to problem solve when they leave high school.
- [07:24]
Shelley Meng
Thank you. Liz. James, I want to follow up with you with the same question about what's your policy, what you're doing in your own classroom, especially if you're teaching both the undergraduate, undergraduate. But I also have a more technical question for you because Liz talked about using these online tools themselves as GPT liker tools to do the detection. And you are the computer scientists actually working on kind of secret of these large language models. So I want to ask you your opinion on how trustworthy these tools themselves, because that itself is an issue. Right. You can accuse somebody plagiarize, but actually they say, well, they didn't. We know these tools are not 100% accurate, may not even 90% accurate.
- [08:08]
James Zhao
Yeah, these are really interesting and important questions. So in my undergraduate and also graduate classes that I teach here at Stanford, we allow students to use ChatGPT and other language models. We ask them to be accountable so that they should still be really accountable for everything that they submit homework assignments or projects. And they should also try to be explicit and Be transparent and state in their assignments and their homework where are they using these language models. But we do not tell them that they cannot use these tools. And there's a couple of reasons for this. First is that, and this relates to your second question, which is that I think it's actually going to be increasingly difficult to detect for a given assignment whether they're actually using language model to generate that assignment. And we've actually done some evaluations of tools like Detect GPT or other vendors that people are using. Those tools are I think are often relatively easy to fool in the sense that if the students modified or prompt a bit, or maybe the students can say, okay, rewrite it in this style. And often that would actually bypass those detectors. So there's a lot of false negatives. On the other hand, there's also a lot of false positives in that if you are actually, let's say a non native English speaker and you actually did the assignment yourself and you wrote it yourself often. We also found that these detectors can often mistakenly flag a lot of the texts that are actually written by real humans to be GPT generated because there are certain biases in how non negative speakers write that tends to trigger these detectors. So both on the false positive and false negative side, in our experience, we found that it's just not reliable enough to apply it to individual students.
- [10:00]
Shelley Meng
I see. Well, that's a fascinating topic and clearly there are a lot more research that needs to be done. And HDSR is publishing a special issue on generative AI and there will be coming articles on detecting. There's one of articles we'll be talking about how much are generated by AI and how much are edited by humans. So I can see there's a variety of those things that will be coming.
- [10:25]
Liz Rose Shulman
Would it be okay if I added something to that? I appreciate what James is saying. I think that in addition to the detection sites, which clearly are not going to be 100% reliable, of course it's been pretty easy to tell when a student has used it because it just doesn't sound like who they are. Sometimes it's too well written or it doesn't really match their personality. And the other thing that I wanted to add is that ChatGPT was not designed for high school and we're operating as though it was. Teachers are being told that this is another tool in the teacher toolbox and we're being told all these ways that we can use it. It wasn't designed for school and teachers certainly weren't part of the process of it infiltrating schools, especially for English teachers who are here to help our students learn how to think for themselves, really, for the first time in their life, to understand what that means to think critically. And In November of 2023, the National Institute of Health said that young students using it not as a tool but to do their homework that is actually hindering critical thinking. And I'm concerned about what's going to happen when these students graduate high school. And they haven't developed that basic ability to think through problems because they're missing critical thinking skills.
- [11:51]
Shelley Meng
I think that concern is shared across the board, not just for high schools and obviously for any of us are educators. Liz, you're absolutely correct. These tools were generated. I generated without, you know, a lot of these considerations.
- [12:07]
Liberty Vittert
Isn't in some ways chatgpt allowing for some of this critical thinking at least skills or knowledge? You know, I remember when it first came out, I had my students do something where they said, write a poem about Donald Trump. And chatgpt says, we refuse to write a poem admiring Donald Trump because it goes against our rules or something. And then they said, we'll write a poem admiring Joe Biden. And it wrote this long, pretty terrible, frankly, but really long ode to Joe Biden. Is there in a sense the way that if we as educators choose to utilize ChatGPT, we can actually use it as a learning experience for their critical thinking skills because they're gonna be seeing AI generated material online all the time. They're gonna see it when they're scrolling through social media. You know, they're gonna be exposed to it over and over again and isn't by them using it and being able to that arise from using it that in a way that itself can be a teaching moment.
- [13:04]
James Zhao
I agree. Yeah. I think that's actually a really important aspect that you brought up is that I think people will be using these tools increasingly and in some sense it's going to be very hard to fight against that trend and try to not have people not use these tools. So I think it would be very important for us as educators to see how can we teach students to use these tools reliably, but also engage with the output of these AI tools in ways that still are educational and also enhances their own critical thinking abilities. I think ChatGPT is a very sophisticated form of a calculator. It can do all sorts of calculation for us. The advent of a calculator did not reduce the need for students to learn how to think logically about math problems. It reduces the need for people to memorize multiplication tables. Similarly with ChatGPT, maybe there's less of a need for people to memorize some of the more mechanics of things. Now, we don't ask students to memorize which Python were R packages to use. And maybe they don't need to memorize all the specific technical words and jargons, because those are easy to fill in and look up with ChatGPT's help. But on the other hand, it's increasingly even more important for them now to be able to, for example, evaluate an output of either their human colleague or of a language model. They have to look at a program that's written by ChatGPT, or a calculation that's done by the language model and evaluate and say, okay, so where did it go wrong? What are the potential biases? What are the potential limitations of this program? How do I debug it? And I think that kinds of evaluation debugging becomes increasingly valuable part of the education experience that we want to highlight.
- [14:56]
Shelley Meng
Listen, now we know that there's such technology out there, and suppose you are invited enough to say, hey, we have this technology there. What is your wish list? How do we take advantage of these technologies to design these tools that will be useful for you as an educator and make them tailored to helping students to learn? Is there any good use of that? What would you tell them?
- [15:22]
Liz Rose Shulman
I think that there could be. I think the problem is that students need to first know what it means to think critically. It's kind of like you can't fly the plane upside down until you learn how to fly it right side up. And so if they're not even learning how to fly it right side up yet, and then we're kind of expecting them to know what's critical thinking and what's not. And so I think that we need to rethink what is the purpose of school. Is the purpose of school to use AI? Is the purpose of school to learn how to think critically? There's a whole movement of teachers who are going back to blue books and pencils because the technology has, you know, as we know, we have, the research has. Has harmed students. The problem is when English teachers like me say that we're considered Luddites, but, you know, thank God we have the technology during the pandemic when we were all teaching from home. So, you know, we're happy to use the technology, but I think we need to ask what. What is the purpose of school? And, you know, do we want students to learn how to think critically? I'd Be happy to use ChatGPT to help students generate drafts, but first they need to know how to generate a draft because otherwise they're just going to use it in place of their own original thoughts.
- [16:41]
Shelley Meng
At the same time, do you see an opportunity there that using that kind of tool can actually help students to learn the communication, learn the importance of language? Because after all these technologies actually develop on large language models, there's a rich language component to it. Do you see some ways that these tools can be used to enhance the critical thinking communication skills? How do you talk to people? How do you understand what others saying? Do you see possibility there?
- [17:14]
Liz Rose Shulman
Absolutely, I could. Again, I think that teachers need to be included in the process because we have not been included at all in bringing AI into schools. Although now there are conferences for teachers on how to use it. But we haven't really talked about the fact that it wasn't designed for school. But yeah, I do think that it could be helpful. But I think we need to be really careful when we're dealing with the teenage brain that has already been changing because of social media addiction. You know, we have research to show how much they're struggling. And so I think it's just one more thing that is making teaching really difficult when it's kind of being branded as helping us. But yeah, I think it could be. If we can use it in a way that it, you know, isn't going to do any harm, I'd be happy to use it.
- [18:03]
Shelley Meng
Similar question, James, for you that what do you see particularly for you because you're also working on the methodological side, what do you see that we can tailor these tools for the education purpose, for example, in graduate studies and even go beyond the study. You talk about the review problems, but on the other hand, we do know there are these conferences, I guess you probably know better than I do have way too many papers submitted. They don't have enough time, enough manpowers to do these reviews. So could there be actually a real usage of these tools? Because otherwise you send to people not qualified to review them may or may not be better than AI, which in some sense synergize other opinions out there. What do you see? Where are the kind of a safe space you can experiment and what are the things that you need to watch out for trying to be innovative in that space?
- [18:58]
James Zhao
Yeah, that's a great question. I think Liz raised a really important point which is that chatgpt as we know it, it's a very general purpose tool. It's not really specifically designed for Education or for research, it's really more of a general purpose, almost like Google, the semantic conversational search engine. But there are, I think, interesting ways of tailoring these tools potentially to make them much more suitable for education. So for example, right now if a student wants to use ChatGPT to write an essay about Hamlet, the student can just say ok, summarize Hamlet for me in a paragraph and the model would just do it. The current ChatGPT is designed to be as helpful as possible for humans, so it's really want to be very subservient and then just do what humans ask to do. But a good tutor in this case should not just summarize Hamlet for the student and they should say at oh, what do you think about Hamlet? What do you think about this aspect of the play? How do you think this relates to Scarlet Letter? Relates to some other literature. So that process, I think of tailoring a language model from this general purpose tool into something that's more specific is an active area of research that we call alignment, which is basically trying to customize this language model to align it into specific behaviors that we think would be useful in the case of education. Ideally we want to align these language models so that they actually would know how a good useful tutor would behave. Maybe we need to train these models on a lot of actual data from good human teachers and tutors to see how would the teacher answer these questions or how would teacher actually post questions to the students. And a good teacher will not just summarize hamlets and just write to the homework. A good teacher will push back and say well, what do you think about X, Y and Z? Right. And I think that would require some additional training to tailor the language models. But it's definitely within the realm of the technologies to really design it so that it can become useful tools. On the research side, I think because we work with a lot of graduate students, I think these language models can be very useful as a co pilot potentially for students. For researchers like you mentioned, I think the current peer review system is very much overburdened. There are lots of papers and it's very difficult now, especially for junior researchers and especially if you're from like an under resourced country or institution to get high quality feedback from external experts. Right. And that's a place where you could actually have a language model who's read all of the recent papers and provide some summary and some feedback to those junior researchers in a way that's almost instantaneous. Right. And we have seen some evidence that the feedback provided by the language model can be even more helpful in than maybe some of the feedback that they might get from the human peer reviewers, especially if those reviewers are not as dedicated in providing constructive suggestions. So I think there's definitely an opportunity to see how can we again align and customize these models to be useful co pilots or useful tutors and even useful maybe AI research co advisors for the students.
- [22:13]
Liz Rose Shulman
From the perspective of a teacher in the K through 12 education world, teachers can do the very thing that James was mentioning, and that is what we do in the classroom, asking students ideas about Hamlet, if they're generating ideas for a paper, for example. And I think this is the concern that a lot of K12 teachers have is that we go to school and we are experts in our field and we know our students personalities and we want to work with them and we design lessons. And it's that human connection particularly that has really been missing, I think since COVID You know, students are really, really struggling in schools and need even more of the human connection.
- [22:57]
Shelley Meng
Thank you, James. I know you have published research on improving the fairness of AI in general. So I want to talk to you a little bit about that, particularly with respect to generative AI and in the education space. Do you see any particular issues about the fairness of AI in this space? For example, maybe some people have more access to these tools than others, so they have better educated way of using it. And any concerns you have?
- [23:23]
James Zhao
Yeah, I think there's a couple of aspects that's related to both the access and also the quality of these language models. In terms of quality, currently these models are relatively good for English and also for specific kinds of questions. But the model often still reflect biases and limitations in its training data. And currently the training data I think is quite good and quite large for the more general kind of European English and other languages like Chinese, maybe Spanish certainly is much worse for other settings, other countries and other languages. So that will lead to a disparity in both the performance of these models across different settings. So maybe it is much worse, for example, when it's in certain dialect compared to in English. And that can lead to more mistakes, more hallucinations by these algorithms once used in these other settings. Even within English, if we talk to it using specific dialects, maybe outside of the very standard English, if you talk about it with certain slangs or use different abbreviations, that can also lead to more mistakes that are made by the models. So I think it is important for us to continue to evaluate how well these models work and the Potential biases there. And another aspect of the problem that we often see with these models is that they have learned certain potential biases about different users. If you tell it, oh, I am a user, Jung, I am from this background and I am a professional, then the model tends to have one set of responses. But if you change the name into something maybe a bit more ethnic, and you say that, oh, I'm from this other city, and then the model actually based on its own priors, would adapt and change its response. So we see that the model does give different responses, even if you ask the same question, but just change how you represent yourself as a user. And that's something that could potentially lead to biases and disparities in the models.
- [25:27]
Liberty Vittert
You know, I have to ask, and Liz, I really like your thoughts on this about sort of how these generative AI platforms will really impact the future of students. You know, during COVID and you know, still today, we see a lot of individuals in the news saying that students are becoming too reliant on technology. And we're basically raising a generation of cheaters. You know, in order to combat that for my class, I said, everyone can use ChatGPT. Use it as much as you want. And I put the onus sort of on myself to come up with assignments that they needed to use their actual critical thinking skills, either around ChatGPT or over it or in some capacity so that they would have to try to use that. But at the same time, I was learning right along with them. I didn't have some magic wand to show me ChatGPT four years ago, and I could spend three years learning how to use it. I was learning at the same time of how to use it. But do you think that this fundamental change is going to shape our students? Either they're going to become this sort of generation of cheaters when they're not allowed to use ChatGPT, but they do. Or are we going to have a generation of people that can't think for themselves if we allow them to use ChatGPT? Or is there an opening there for really letting ChatGPT and other things like that enhance our learning?
- [26:43]
Liz Rose Shulman
Yeah, I think those are really important questions. I mean, I. I think that there are definitely creative ways to use AI, but I think when we're talking about the teenage brain, I think it gets very, very tricky because I am already seeing the consequences of students using it. Even students who are told by teachers that they can use it because they're just at the beginning of their life of learning how to think through, through problems and to problem solve on their own. And so I think that we're already seeing some of the effects of them using it, even when they are using it as a tool. I know I said this before, but it was not designed for schools. Teachers were not included in the process of bringing this massive thing into schools. And we're already dealing with social media addiction and the changes in the brain that, you know, we all know about, you know, the cell phones and all of that. And I think in, you know, if you're not in the, in the high school classroom every day in the trenches with these teenagers, I think it's difficult to understand how much they are being affected. And it doesn't mean technology is bad. And like I said before, it saved us during remote learning. We're happy to use it, many of us teachers, but I think if we're talking about being at the beginning of learning how to think critically and learning how to problem solve, I think we need to not assume that they already know how to do that when they're just at the beginning of their academic career, even learning what that means.
- [28:22]
Shelley Meng
Well, clearly there will be a lot for us to do. And in that space, James, I want to ask you, because you have co authored an article in Harvard Data Science Review to particularly talk about the impact of generative AI ChatGPT on data science education itself. And I think that is another incredibly important topic, particularly for the audience of this podcast. So could you help to summarize briefly about that article and what are the findings and what are your recommendations?
- [28:57]
James Zhao
Yeah, thank you, Shelley. I think there's a lot of discussions in data science community, especially among students. Is generative AI going to, to replace certain kinds of data scientists? I don't think that will happen, but I do think data scientists who knows how to use generative AI will be much more competitive and will start to replace data scientists who do not know how to effectively use tools like ChatGPT. And the reason is that I think tools like ChatGPT, they have a lot of knowledge and if we use them properly and data scientists can really free up a lot of their bandwidth to do more higher level critical thinkings. Because a lot of the data science as we know, it's often involved around the more higher level thinkings and designs, but also a lot of time spent on, for example, wrangling data, putting things in different formatting, looking at different writing different scripts to put different data together, pre processing data. And those are the kinds of tasks, especially if the more common type of Data modalities. The language model can be very good at doing and can relatively quickly generate some of these initial first pass analysis that could then open up a lot of time and bandwidth for the data scientist to really look at the output of the language model and then see, okay, so what are the higher level, the more critical questions? How do we do the evaluation? What are the hypothesis that we should test? What are the potential biases in the model, in the data? These are the more higher level questions that the data scientists can focus their time on. I think that would be one way to sort of up level the expertise that's more uniquely human to the human data scientists.
- [30:37]
Liberty Vittert
Well, I know that we could talk forever on this. I find this whole concept so fascinating and so hard to sort of navigate. And I think it's something that we're all, I guess, especially in higher education, trying to navigate together. But it does bring us to the end of our conversation. And we always ask one question at the end of all of our guests, and that is, is if you could wave your magic wand, so this is for both of you. If you had the power to establish a policy so you could wave your magic wand and establish a policy on generative AI for the entire education system, what would it be? Liz, I'll start with you.
- [31:15]
Liz Rose Shulman
I think that part of waving the magic wand would to really rethink what is the purpose of school? Is the purpose of school to learn how to use AI? Is the purpose of school to learn the skills so that they can use AI because they're going to need to know how to use it when they graduate from high school? Is the purpose of school to learn how to think critically? I think we need to kind of go back and to ask what is the purpose? But if I had a magic wand, I think students would be learning how to think for themselves and to think through problems and to think critically in the best possible way, whether that's using AI or not. But what I am seeing now in K through 12 education and at Northwestern University and talking to the future teachers that I teach, students are not learning the critical thinking skills that they're going to need to think critically with AI when they graduate from high school. So that would be my wish.
- [32:12]
James Zhao
My magic wish would be, I think we do need to incorporate generative AI into all the aspects of our education curriculum and a part of that also be incorporating it into how we interact with teachers, with educators. We need to give the teachers more tools so that they could first have more experience working with these algorithms and also more guardrails to customize these models to their classrooms. Going back to before, I think there are ways to make these tools potentially into align them so that they can be more useful for teachers and more responsive to students, rather than just writing Hamlet essays for students. Yeah, I think incorporating it into our curriculum and teaching all the stakeholders, both the students and the educators, how to responsibly and effectively use these very powerful technologies would be my wish.
- [33:10]
Shelley Meng
Well, thank you Liz and James for this very timely conversation. I can assure all the listeners this conversation, all the questions, all the answers are created by humans. There were no chatgpt used as far as I can tell. But I do wish someday maybe we can conduct a podcast which we will be able to have a conversation with a ChatGPT or AI generated response. I'll be very curious to see how how that goes. That technology may get us there, but clearly we all agree that when these technologies takes place, we obviously should take advantage of them, but we also need to use them in a very critical way. I think the emphasis on critical thinking clearly has been working for all of us because we all think very critically and we hope that our future generations, when they conduct these conversations, it will at least keep the words being critical. Thank you very much.
- [34:10]
Liberty Vittert
Thank you for listening to this week's episode of the Harvard Data Science Review podcast. To stay updated with all things HDSR, you can visit our website at HDSR, MITPress, MIT.edu or follow us on Twitter and InstagramHDSR. A special thanks to our executive producer Rebecca McLeod and producers Tina, Toby, Mac and Ariane with Frank. If you liked this episode, please leave us a review on Spotify, Apple, or wherever you get your podcasts. This has been the Harvard Data Science Review. Everything Data Science and Data Science for everyone.