Beyond ChatGPT: Critical thinking in the age of AI

Hari Subramonyam (00:00):

AI doesn't have to be the problem, and it can be part of the solution if we design it right and if we teach kids the right way to use it.

MUSIC (00:08):

(music)

Denise Pope (00:12):

Welcome to School's In, your go-to podcast for cutting-edge insights in learning. From early education to lifelong development, we dive into trends, innovations, and challenges facing learners of all ages. I'm Denise Pope, senior lecturer at Stanford's Graduate School of Education and co-founder of Challenge Success.

Dan Schwartz (00:35):

And I'm Dan Schwartz. I'm the Dean of the Graduate School of Education and the faculty director of the Stanford Accelerator for Learning.

Denise Pope (00:45):

Together, we bring you expert perspectives and conversations to help you stay curious, inspired, and informed.

(00:54):

Hi, Dan.

Dan Schwartz (00:55):

Denise, good to see you as always.

Denise Pope (00:58):

Same. What's going on?

Dan Schwartz (01:01):

Well, it's all AI all the time, so we're gonna have another podcast about it.

Denise Pope (01:06):

And this one's kind of AI with a little bit of a twist. We're gonna be focusing a little bit more on the thinking part. We want the students to be thinking while they're using AI. Not that we didn't want that in the past.

Dan Schwartz (01:20):

Yeah. What twist is that? (Laughs) It's like,

Denise Pope (01:22):

Well, the twist is I wanna start us off with thinking about critical thinking. So instead of talking about AI and technology right off the bat, I'm very interested in what you think critical thinking means. Like, what's the definition to you?

Dan Schwartz (01:36):

Right, 'cause people think the AI is gonna wipe out human critical thinking.

Denise Pope (01:40):

Exactly.

Dan Schwartz (01:41):

You know, some people think it's just pure logic is critical thinking. Some, some people sort of think it's more composition. My decision on critical thinking, what most people mean is don't be stupid is critical thinking.

Denise Pope (01:55):

(Laughs) Use your head. Just use your head.

Dan Schwartz (01:57):

Think about it. Uh, I actually have a definition of it.

Denise Pope (02:01):

Oh. I, I wanna hear it.

Dan Schwartz (02:03):

I'm not sure anybody else subscribes to this definition, but it, it's, it's the processes by which you decide what to believe.

Denise Pope (02:09):

The, wait, let me, let's break that down. The processes by which you decide what to believe?

Dan Schwartz (02:14):

Yes. So it could include what kind of evidence do you look for? Do you look for consistency? Do you even choose to seek evidence? Things like that. So, uh, do, do you check sources? So it's a little more complicated.

Denise Pope (02:30):

Yeah. Well, I, when I teach, uh, qualitative research, one of my questions is, how do you know when a claim is true? This seems kind of similar, right?

Dan Schwartz (02:39):

Uh, yeah.

Denise Pope (02:40):

Like what would you do to check that? How would you even know how to think about it?

Dan Schwartz (02:45):

Yeah, yeah, yeah. No, I think this would be a very specific brand of critical thinking.

Denise Pope (02:48):

Okay. All right.

Dan Schwartz (02:50):

It's a-

Denise Pope (02:50):

Well, yeah.

Dan Schwartz (02:50):

But, but anyway, there, there's a lot of people who say that, uh, ChatGPT, which you know, is very capable of writing prose and, uh, arguments, is going to displace people's opportunity to learn how to do critical thinking.

Denise Pope (03:06):

Because it's gonna just do it for them and they're not even gonna have to think. Is that the argument?

Dan Schwartz (03:11):

I think that's right. So let's take a slightly different approach to this. So, um, let's assume students aren't just offloading everything to the AI while playing some video game with their free time. Instead, the question is, how can AI improve learning important topics, skills, attitudes, things like critical thinking?

Denise Pope (03:30):

Okay, love it.

Dan Schwartz (03:32):

So the question isn't how's it gonna destroy it. Question is, uh, how can we make it so that it helps it?

Denise Pope (03:38):

Okay. We're taking a positive spin, and our expert here is gonna take us on that positive spin.

Dan Schwartz (03:43):

Well, we'll see. Today we're very lucky to have Professor Hari Subramonyam. His expertise is in human computer interaction and bringing it to education. So, Hari, thank you so much for joining us.

Hari Subramonyam (03:56):

Thank you for having me. I'm really excited to be here.

Dan Schwartz (03:58):

Yes. So I wanna, I wanna dive straight in 'cause you, you actually did a study that's sort of, uh, is, if I recall, like one condition does ChatGPT, one condition does nothing, and the other condition does your special secret sauce with AI.

Hari Subramonyam (04:13):

Yes, yes, yes. So I wanna start off by building on the argument that you were making about critical thinking, right? And like, you know, writing is a way for people to engage in critical thinking. It's only when you're writing that you're trying to identify what is the context and like what kind of ideas you have. What kind of connections do you need to make between these ideas? Are you having like a logical chain of reasoning from some like idea to evidence to insight? Are there gaps in your ideas? What are the evidence?

Dan Schwartz (04:43):

And you, you do that when you're writing?

Hari Subramonyam (04:44):

Yes (laughs).

Dan Schwartz (04:45):

You don't, you don't think it all through beforehand? You sort of, on paragraph 14, and I'm thinking, uh, what evidence do I need for paragraph 15?

Denise Pope (04:53):

Some of us are not as genius as you, Dan, and we can't just do it all in our head and then spit out the beautiful, all completed essay, prose, whatever.

Hari Subramonyam (05:03):

Some of it is mental where you're forming the argument, but also like having an external representation, like, you know, the externally written text also like helps you sort of iterate on your critical thinking specifically. So anyways, to the point of the study, and like what we were trying to do is with tools like ChatGPT and like, you know, this is maybe what you were alluding to earlier is students can give very little input and get lots and lots of text as output. So there is this imbalance between like how much input you give and what kind of output you get. So if I'm trying to write an essay, I can just go to ChatGPT and say like, okay, you know, help me write an essay about say, um, you know, healthy eating and nutrition. And it's gonna gimme like, you know, a three or four-paragraph long essay.

(05:49):

And I think that is, when the student or the learner is not actively participating in shaping the essay, a lot of the creative and critical decisions are made by the AI. And this is sort of like what is problematic. This is partly why we sort of see lots and lots of essays that sort of look and sound the same, because people aren't bringing in their authentic ideas and insights into shaping the essay. Now, in theory, you could argue that like, you know, students can break down their arguments and like instead of asking a ChatGPT to give the essay, they can like do this step by step by asking it to list out ideas, brainstorm with it, and all of that. But this kind of engagement is hard to do in practice because these tools aren't designed to support that.

(06:35):

And as a human computer interaction researcher, what we find is that the human AI interface is missing the kind of affordances necessary for like exploratory, dialogic critical thinking and writing. So in our research, we developed this tool called Script & Shift that is explicitly designed to sort of support these low level cognitive thinking process. So instead of treating AI as this magic box that you throw prompts into, uh, Script & Shift gives students like specific text and buttons and interfaces that allow them to engage in like idea formation and structuring ideas and so on.

Dan Schwartz (07:15):

So wait, wait. So, so this is interesting. I think we need to make it a little more concrete. Let's pretend Denise is in front of the computer and she wants to write an essay on the beauty of relationships between an owner and their dog. (Laughs) Do you accept, do you accept your assignment, Denise?

Denise Pope (07:35):

Okay. I accept the assignment, yes.

Dan Schwartz (07:38):

Okay.

Denise Pope (07:39):

And so I could just say, "ChatGPT, write me an essay on the beauty of a relationship between a human and the dog." And out would pop this essay, and I would have done very little thinking about it whatsoever. But now Hari has some kind of tool that I can use. Walk us through how I would use that tool.

Dan Schwartz (07:57):

What's the first thing Denise sees?

Hari Subramonyam (07:59):

The first thing Denise sees is a layer where she indicates what is her goal? Like in this case, writing an essay about her relationship between her and her dog. And then you need to think about like, potential audience. This could be like a general audience. So if you're trying to write it for a specific audience, so you can fill in some of these like, the metadata, like your goals and intentions behind your writing.

Denise Pope (08:22):

Like prompt, you're giving me prompts, like who's your audience? Like, are you kind of acting like an English teacher?

Hari Subramonyam (08:28):

Um, at this point, like you are sort of expressing like these initial goals. At this point there is no AI yet in the tool. Like you, you just, when you open the tool, there is an interface which tells you like, you know, what is your writing goal? What is your topic, what is your audience? Sometimes if you're using other reference materials for your writing, you can upload all those documents as well, which the AI will use at a later stage to sort of facilitate this like low level like thinking and the writing process.

Dan Schwartz (08:54):

But I, I agree you, Denise. Uh, thinking about who your audience is, it's something you gotta teach people 'cause they... Yeah.

Denise Pope (09:00):

Yeah. I mean, that's one of the things I teach people all the time before you write. What's your purpose? Who's your audience? Yeah. Love that.

Hari Subramonyam (09:05):

Yeah.

Denise Pope (09:06):

Okay. So then what happens? So I, we've established all that and then what happens?

Hari Subramonyam (09:10):

Yes. Okay. So at this point, I'm gonna like briefly tell you what the interface looks like. Imagine you were writing this thing on pen and paper, like you were writing this essay on pen and paper. You may have like a bunch of note cards or cue cards where you put in like, you know, different anecdotes or like things about your dog, specific events and experiences, all of these on like different cue cards. And then you're trying to then compose them into a coherent story. You can think about the order and the stuff within like each cue card. You can kind of think about like how you're gonna translate what's in the cue card to how you wanna say it to a specific audience. So the interface itself supports this kind of like, uh, a spatial layout with a bunch of layers that you can lay out flat in the table.

(09:53):

You can put in like different content on different layers. And then you can transform the things in like each layer to tailor to specific audience. So say, suppose that you have a really interesting event about taking your dog, um, on, on a first car ride or like a road trip. And in this case, you wanna figure out how you want to tell the story to different audiences. You know, if you're trying to tell this to like children, like you may wanna phrase it differently as opposed to if you are writing a blog article for like, people traveling with pets. So this is where like Script & Shift allows you to try out these different variations augmented by AI. So you can tell the AI in this, in Script & Shift to tell like, you know, like change this tone to be a little bit more like kid-friendly or like change this tone to be a little bit more like informative. And so this, this is sort of like the, the high level layout of Script & Shift, or like what, how, what the tool looks like.

Dan Schwartz (10:52):

I've sort of said the meta parameters of, of my writing task. It, it, it should be five pages, things like that.

Hari Subramonyam (10:58):

Yeah.

Dan Schwartz (10:59):

And then I fill out a bunch of index cards with different things I'd like to include, right? And then I sort of organize those cards. Do I sort 'em around the screen? Do I push 'em around and stuff like that?

Hari Subramonyam (11:12):

Yes. So yes, you can, you can move things around in the screen, and the, and the way the the tool works is once you stack them in a specific order, it'll then compile the final article for you. So you can look at the final article and the flow. And if you don't like it, you can change the order of the cards. You can remove a card, insert a new card in its place, and generate the article again. So this is what allows writers to be more creative and explore different variations and combinations of ideas and different order, which makes writing more creative and engaging.

Dan Schwartz (11:43):

It is taking what I say in my cards, like a set of bullet points, and turning it, turning it into prose that's connected.

Hari Subramonyam (11:50):

Well, kind of, but you, you don't to have bullet points. You, you, you're doing the entire scripting within like each card. Like I use the card metaphor to sort of simplify things, but it's, it's essentially a page. Each card is a page. You can write text in it. We have what we call the Writer's Friends that live in these pages that you can call up anytime you need help. So we have like an Idea Ivy. So if you're stuck in like the ideation phase, you can call Idea Ivy and like chat with Idea Ivy about what you're trying to ideate on. If you have written an initial draft, you can call like another friend called Structure Sam-

Dan Schwartz (12:26):

(Laughs).

Hari Subramonyam (12:26):

... who's gonna help you like, explore different structures for this idea. Or if you have a vague sort of like argument you've made and you want more details about it, you can invoke Detail Danny, who's gonna then like help you flesh it out. So we have these Writers Friends who live on like, individual pages. Some friends help you with like, writing, what we call scripting, some friends help thinking through like variations of what you have written, which we call shifting, like exploring different rhetorical strategies.

Dan Schwartz (12:54):

Ah.

Denise Pope (12:54):

This is an English teacher's dream. Instead of just having the kids like get better at prompting ChatGPT, you're actually building in some of the things that I as a high school English teacher would've wanted my students to really think about in a kind of fun way by naming 'em and, and, and not just like, here AI, just write my whole thing. I, I love that.

Dan Schwartz (13:16):

So I can't resist. I, I can't. Is there a Critical Thinking Karen?

Hari Subramonyam (13:18):

(Laughs) There can be a Critical Thinking Karen. Um, so these friends are configurable by the teacher. So just to, a little bit of background. Before we built this tool, we interviewed over 20 English and history teachers.

Denise Pope (13:31):

Uh-huh.

Hari Subramonyam (13:32):

And we co-designed a set of prompts, looked at like English grading rubric. And we sort of work with teachers to understand like where would they want students to get help from AI, where are they struggling.

Denise Pope (13:42):

Lovely.

Hari Subramonyam (13:43):

And what are the things they don't want the AI to do for students? So all of this is sort of like built into the tool, which is why as you said, Denise, it's the teacher's dream.

Denise Pope (13:51):

No, it's a dream. Okay, so now this is a really cool part, tell us about the study, 'cause I saw the findings.

Hari Subramonyam (13:57):

Yes.

Denise Pope (13:57):

And I was surprised.

Hari Subramonyam (13:59):

(Laughs) Yes. So the study, um, we, we recruited 90 undergraduate students, uh, and we assigned them to one of three conditions. One is no AI support, one is using a traditional interface like ChatGPT, and one is, uh, that the third condition is using Script & Shift. And what we found in sort of like, and, and they were asked to do this sort of document-based question answering task where they were given like a set of articles around like advertisement. And they were right to, uh, asked to write a critical article about advertisement, role of advertisement in society. And what we found is that the participants, or like students who use Script & Shift engaged a lot more in what we call knowledge transformation. So there are two concepts. There's knowledge telling and knowledge transformation. Knowledge telling is just, you know, retrieving what you have in your mind and saying it as is. So for instance, like going back to the nutrition essay, you can say like, you know, kids should eat vegetables because they have vitamins and minerals.

Denise Pope (15:01):

Right. Just spinning it back. Just not, you're not really thinking, you're just spinning back.

Hari Subramonyam (15:05):

Exactly Yeah. Yeah. Okay. But now imagine you wanna say this to a parent or a care- uh, caregiver. And the way you would say this is like, you know, um, introducing like colorful vegetables in your kid's diet is like nutritious and helpful and uh, it promotes a, a healthy eating habit. So that is sort of knowledge transformation where you're taking this fact that you wanna say and finding out a way to say this well within the context of this argument that is aligned with the kind of audience you're trying to communicate to. And that is knowledge transformation. And there are many different critical thinking steps. Everything from analyzing to synthesis to evaluating evidence. And the biggest finding was that kids who were using Script & Shift, because they were able to engage in these low level thinking process of writing, they carried out a lot more knowledge transformation sort of sub-tasks compared to like the other condition where they just used ChatGPT.

Denise Pope (16:02):

So did you, when, to find out that they were carrying out the knowledge transformation subtext, was that just because you watched them doing those pieces, doing the writing? Or did you have another kind of assessment afterwards to see what they learned?

Hari Subramonyam (16:15):

So we'd given them a set of source content based on which they were trying to produce this essay. So comparing the source content and the student-written like essay, we could observe and code for the specific knowledge transformation that went on from like the initial source that was given to them and the final content.

Dan Schwartz (16:33):

So, so if it, if it's uh, just replication, you're measuring the degree to which they're basically taking sentences from the source materials and maybe changing the grammar.

Hari Subramonyam (16:44):

Yeah.

Dan Schwartz (16:45):

And then knowledge transformation is they're introducing some inferences that aren't in the source material.

Hari Subramonyam (16:50):

The second sort of point is around, uh, agency. And when we talk about agency, it's often like control over the tool. But over here, when we say agency, we mean more control over their thinking process, right? Because the tool is aligned in a way that gives students or writers the space to engage in these sort of low level writing process. And they felt that they had more control in the way they were producing the text as opposed to the text being written for them.

Dan Schwartz (17:18):

Yeah. But isn't, isn't the condition I care about where there's no technology support at all? I thought there was a third condition here in this study.

Denise Pope (17:26):

Yeah. The ones who, who had to go back to like a blue book, right? Without any tech help.

Dan Schwartz (17:31):

How was their knowledge transformation?

Hari Subramonyam (17:33):

Yeah, so the no tech help and like chat LLM conditions did not have significant knowledge transformations along like the, the established set of dimensions, synthesizing, assessing, and all of these things compared to Script & Shift condition. And similarly, with agency, we didn't assess for agency in the baseline condition, but compared to like Script & Shift and a ChatGPT-like condition, like students who used a Script & Shift felt like they, uh, self-reported that they had a lot more control over their writing.

Dan Schwartz (18:06):

Alright, so there you have it. AI can be better than no AI, and good AI is better than ChatGPT.

Denise Pope (18:12):

Yeah. Scaffolded AI could be better than no AI at all. Why? Because yeah, you're teaching something (laughs). Like, to me, this is a, a no-brainer. Very, very exciting, Hari. But also like, yeah, with some help, like a teacher or a Script & Shift, you're gonna learn to do low level thinking instead of just sit down, write this whole thing. Oh my gosh, I don't know how to do that.

Dan Schwartz (18:35):

Okay, there we go.

MUSIC (18:36):

(music)

Dan Schwartz (18:45):

So we, we've heard about a study where they use technology and it, uh, led, led to better sort of knowledge transformation. They're, they're engaging the ideas, elaborating them, connecting them. So here, here's the big question for both Denise and you, Hari. Is the measure of whether this is effective that later on, I give them a writing task without the computer and they do better? Or do I just assume in the future, there will always be the computer? So if I want to test their learning, I need to give them these tools, right? So if I wanna find out if people can write, do I say, "Sorry, they can only use charcoal on parchment paper." Or do I say, "No, I'm gonna let 'em use the tools that they're gonna have in the future."

Denise Pope (19:29):

You kind of just answered your own question, Dan. I want you to, I, I, I'm gonna say, it's not going away, so the test should not be how do you do it without it, right? It's how do you do it with it? That would be my answer. I don't know. Hari, you're the expert here, but that would be my answer.

Hari Subramonyam (19:47):

Yeah (laughs). So I think the short answer is yes, I do agree with you, Denise. Like, I think we wanna test them on like how they would do things with the, with the tool. But you know, going a little bit more deeper into why we get kids to learn how to write and why we give them all these like essay writing assignment so that they can build sort of good internal procedures and ways to think about writing. And that is what we want essays on. Like, you know, learning is representing where like you're building good mental models and how should you be writing an argumentative essay? How should you be writing when you wanna communicate something to like a five-year-old? So the idea is that when you learn to write, you sort of internalize all of these, uh, scripts and concepts and constructs. Then what we wanna be assessing on is can they take these representations, how do you get them to apply these things in like new contexts and situations, and are they able to do that? And I think that is how I would think of assessing. Yeah.

Dan Schwartz (20:44):

Are they allowed to use the AI when, during the test?

Hari Subramonyam (20:47):

They're allowed to use the AI, like, just, just not ChatGPT.

Dan Schwartz (20:52):

They're allowed to use your tool during the test?

Hari Subramonyam (20:54):

They're allowed to use our tool during the test. Yes.

Dan Schwartz (20:57):

So, so we're scaffolding their performance, and we're not saying how could they do without the scaffold?

Hari Subramonyam (21:03):

Yes. Yes. Because I think that that the thinking and the, the, the testing happens in the way the students are leveraging these different like AI features, the way they're sort of prompting and interacting with these features. And that is what makes evident the like, understanding of like how to write.

Denise Pope (21:21):

So even, so I'm just gonna throw this out there. Tell me if I'm right or wrong. Yes, you're gonna test it-

Dan Schwartz (21:26):

Wrong.

Denise Pope (21:26):

... with the scaffold. Can you (laughs)-

Dan Schwartz (21:27):

Sorry. Couldn't resist.

Denise Pope (21:29):

Can they use it, right, now that they've done it once, can they do it again, right? But then we wanna see, when we do take the scaffold away, that is the meaning of a scaffold. Are they gonna look at AI in a whole different way and go, you know what? I'm not just gonna ask ChatGPT to write the paper. I'm gonna kind of think of those scaffolds and I'm gonna be like, "Hey, try it for like a fifth grader. Try it for like, how I would say it this way. Hey, can you change the order around?" So you eventually do wanna take the scaffold away, but you don't wanna take the AI away 'cause that's not going away. Am I getting that right, Hari?

Hari Subramonyam (22:00):

That is correct. That is correct.

Denise Pope (22:02):

Okay.

Hari Subramonyam (22:03):

That's exactly what I meant. Yeah.

Denise Pope (22:03):

Oh, I'm so happy (laughs).

Dan Schwartz (22:06):

Very good. Very good.

Denise Pope (22:07):

Everything you've done so far is so exciting, Hari, because instead of just push a button and have the computer do it for you, which is certainly out there and available right now, you have found ways to really scaffold different things that we know are important for kids to understand in the writing process or in creating a simulation or an understanding, scientific concepts, et cetera. So I think it's super cool. So I just wanna throw it back to you for a second. If you, we got a lot of parents and teachers who are super worried about AI out there. What's your advice, given, given what you know and what you've been able to do, what's your, what's your advice? What's the secret here to, uh, to handle AI in education?

Hari Subramonyam (22:49):

Yeah, yeah. First, I wanna acknowledge this concern. Many AI tools that we use today are designed to give you like polished, finished output rather than help people learn. But the good news, at least like, you know, given the range of work going on in my lab is like AI doesn't have to be the problem and it can be part of the solution if we design it right and if we teach kids the right way to use it. So my advice to parents and teachers would be to don't ban AI tools. Instead, help kids use them more intentionally, like asking for help with like brainstorming or clarifying an idea or getting feedback. I would also suggest that we look for tools that support these processes and not just the outputs. Like, you know, Script & Shift is a good example where the tool is designed to support the human cognitive process, human thinking process.

(23:39):

So those kinds of tools are really good and powerful in facilitating learning and promoting like kids' curiosity. And I would also suggest that parents and teachers stay curious along with their kids. When they're using AI, like ask them like what choices they made, why did they prompt it in, in a specific way, and then like explore different alternative ways of prompting and like using the AI with the kid. Like, you know, one of the teachers when we talked to earlier mentioned they don't care about kids using AI. What they care about is this productive effort in the interchange between the learner or the child and the AI.

Denise Pope (24:14):

Super helpful. Dan, you wanna bring us home?

Dan Schwartz (24:19):

Yeah, so I think one, one summary point is we need to have a show that's about critical thinking.

Denise Pope (24:24):

(Laughs) Okay. Sure.

Dan Schwartz (24:25):

I, I think there's just, there's some work to be done here. And, and may, maybe I'm the one who needs to be cleaned up on this. You know, I like Hari's vision. The vision here isn't, it's really about education, right? And he's making purpose-built tools to support education. And what I particularly like about the tools is they're not sort of doing what a lot of school is, which is just replicating what you've been told, right? You're told what to do and then you do it. And the AI can really push that 'cause it can keep saying, "You didn't say it right. You didn't get the right answer." Here, he, he's using the AI to enable people to surpass themselves. They're going beyond what they can do by themselves. And at the same time, that support is helping them learn how to reach further and do more. So I, I like this vision of AI where it's helping, uh, humans surpass what they could do on their own and then learn in the process.

Denise Pope (25:18):

That has been your line all along, Dan, and I think you're absolutely right, right? How AI can help us surpass ourselves. Hari, thank you so much for being on the show. What a great show. And thank all of you listeners out there for joining this episode of School's In. Be sure to subscribe to the show on Spotify, Apple Podcasts, or wherever you tune in. I'm Denise Pope.

Dan Schwartz (25:39):

Uh, AI is still composing, uh, who I am. So I'll, I'll get back to you later.

MUSIC (25:49):

(music)