How Teaching Skills Can Complement the Use of AI in Education With Michael Feldstein

Full Transcript

Dr. Cristi Ford (00:00):

 Welcome to Teach & Learn, a podcast for curious educators, brought to you by D2L, a global EdTech company committed to transforming the way the world learns. I’m your host, Dr. Cristi Ford, Vice President of Academic Affairs. In each episode, either myself or one of my colleagues will meet up with some of the sharpest minds in the K-20 space. We’ll break down trending educational topics, discuss teaching strategies, and have frank conversations about issues plaguing our schools and higher education institutions today. Whether it’s ed tech, personalized learning, virtual classrooms, or diversity and equity inclusion, we’re going to cover it all. Sharpening your pencils, class is about to begin.

Hello and welcome. Listeners, we have been on a roll as it relates to AI recently over the last couple of episodes, and we’ve really been discussing AI’s impact in education and beyond. I’m so thrilled to continue this conversation today by chatting about teaching educators, and why teaching skills are so beneficial as we rapidly see evolution in education as it relates to AI. I’m so pleased to welcome my colleague, guest today, Michael Feldstein. Michael is the Chief Accountability Officer at e-Literate, an organization that helps higher education institutes and the education companies that serve them to continuously improve their efforts to enable students to succeed in a rapidly changing world. Welcome, Michael. Really happy to have you today.

Michael Feldstein (01:32):

Happy to be here, Cristi. Thank you.

Dr. Cristi Ford (01:34):

So before we dive in and chat, I just want to share a little bit around with our listeners a bit about how we reconnected. I had the pleasure of crossing paths with you not too long ago in Florida at a D2L Rise conference, and you gave such a profound keynote and I really appreciated how you talked about the importance and rise of durable skills. And as I listened to you have some of these conversations and aha moments, I thought I got to get him on the podcast.

And so I want to jump right in and really have you talk to me and the listeners about you made this declaration and distinction around lifelong learning, and how moving from lifelong learning into durable skills. I have also heard durable skills called enabling skills or the old language of soft skills. But can we talk a little bit about how this distinction has made and will make a huge difference in learning today?

Michael Feldstein (02:35):

Sure. Let me first say that I don’t think lifelong learning is going, well, at least I hope not, it’ll be really boring if it does. But the point that I made in the talk that you and I saw each other at was that lifelong learning exists because human civilization, specifically communication skills and transportation skills sped up the rate at which we learn from each other as a society. This is all high-minded and high-volume, and we’ll get down and dirty pretty quickly.

But the point is when you lived the way your parents and grandparents and great-grandparents did, more or less the same in the same small area, then talking to the same small groups of people around you, not much changed. Eventually as we traveled further, we domesticated the horse, we invented writing, we invented paper, we could learn more faster, and eventually, the speed at which civilization evolved demanded that we learn new things just to continue to be civilized, to operate in the civilization, and that has continued. We still need to do that, what’s changed is that the pace of evolution has become so rapid, and we see this with AI. The AI isn’t really the thing that makes durable skills a thing. It’s what makes durable skills an obvious thing. Suddenly, we have changes that are not just happening within the course of my working lifetime, but within the course of a couple of years or even a number of months. It starts to raise the question of, well, what do I know that’s not going to be obsolete?

Dr. Cristi Ford (04:40):

Right.

Michael Feldstein (04:41):

When you talk about enabling skills, another good term, the skills that are not likely to be obsolete are the ones that enable you to build other skills, right? Come what may, there’s certain things that don’t go out of fashion and become and are useful in a variety of different contexts. So in a way, durable skills, some durable skills, I don’t want to make too much of a generalization because we’re still figuring this out, but some durable skills are the skills that we really ought to be focusing on in order to enable people who are capable of lifelong learning.

Dr. Cristi Ford (05:22):

I would agree with that, and as I’m listening to you talk about that continuation and that lifelong learning isn’t going anywhere, I will tell you as I talk with educators, and you know this because you’re doing a lot of consultancies and work with education institutions as well, there’s this push and pull around how AI will transform the way that we teach, the way that we educate. I’d love to shift over here now and talk about how as educators and teachers specifically, how are those skills as you talk about them as durable skills, and so that there is a need for those skills even as AI is rapidly changing our environment.

Michael Feldstein (06:06):

I did in the talk that we were at together recently, discuss teaching skills as durable skills. Let’s just start by recognizing that as strange as AI is and as hard as it is to understand, everyone makes a big deal that we don’t understand how the machine works and that’s true and it’s something we ought to be concerned about. We also don’t know much about how our students’ minds work, or what they’re learning, and they’re right in front of us. We have a lot more bandwidth with them in some cases. Not always, they’re not at our fingertips 24/7, but we see their faces. We have different ways of communicating with them. We have to figure out what they’re thinking, what they’re capable of. It can be very strange if you’re an educator, one of the joys of it is finding out, well, what is the student capable of and what are they not capable of?

You do that. You develop a toolkit, a set of skills for probing, well, okay, you got the right answer, but how did you get the right answer? What is it that you really know? Did you memorize a formula? Did you really understand the underlying concepts? If I throw you a curveball and ask you to apply something in a new context, can you do it? And what we find is that both humans and AI are very strange and quirky and unpredictable in this way. Before generative AI, if you’d asked a software developer, “Hey, this thing you’ve built, can it pass the medical board exam for neurosurgery?” Their answer would’ve either been no or yes, we spent a lot of time and energy and made it exactly for their purpose. You never would’ve heard them say, “That’s a good question. I don’t know. Why don’t you try?”

Dr. Cristi Ford (08:22):

Right.

Michael Feldstein (08:23):

And you would never expect someone to come back and say, “We tried and it actually did.” This is what Brown University and Mass General Hospital did. They tested to see if GPT-4, that’s the engine behind ChatGPT, could pass the neurosurgery board exams for medical students, and it did, but it also struggles with basic arithmetic at the same time. And these are the kind of weirdnesses that you expect. And in some cases, if you’ve got the soul of a teacher, you cherish with your students.

Another concept that I think is worth thinking about, I talked about scaffolding, and I want to put that in the context of some educational theory. I grew up a teacher from a family of teachers. I taught middle school and high school, my entire family is from that. It’s the family business, I didn’t think of doing anything else. So I can talk about technology and philosophy and all these things, but really, I’m at home talking about Vygotsky and the zone of proximal development, which is where we live in as teachers. If you’re a college educator and you haven’t had pedagogical training, zone of proximal development, or as we insiders like to say, the ZPD, this is a very simple idea stuff below the ZPD, the students don’t need you to teach them. They’re going to figure it out on their own, right?

Dr. Cristi Ford (10:02):

Right.

Michael Feldstein (10:03):

Stuff above the ZPD, it doesn’t matter what you do with them, it’s just beyond their capabilities. Vygotsky has this whole developmental framework around that, and it’s the work in the middle where teachers matter. We scaffold, we create ways of helping the learner get at this idea that they can’t get out on their own, right?

Dr. Cristi Ford (10:31):

I completely agree, and love that you’ve found a way to bring in zone of proximal development, and what I appreciate about the theory is that it constantly evolves right as the learner progresses, and the scaffolding support reduces as the learner gains independence. And so it makes total sense as you’re talking about the work here that we’re talking about around AI. I’d love for you to maybe give us a practical example given that you’re an educator at heart in terms of how you started to really go deep and really utilizing LLMs and ChatGPT for some of your own work, some of your own research, some of your own lived experiences.

Michael Feldstein (11:09):

My little sister, who is also a teacher, had a brain hemorrhage a couple of months ago from a previously undetected aneurysm and was terrifying. I’m far away from her physically, it’s difficult to get to. There’s not a lot that they can tell us for a long time. You’re really reading a lot of tea leaves, and there’s a reason why the expression it’s not brain surgery exists, because brain surgery is incredibly complicated and delicate.

For me, and this by the way, is how LLMs are being piloted, or one way they’re being piloted in hospitals today is not to replace physician judgment, but for patient education. Here are the medical notes, here’s what they mean in practical terms. Here’s what we know, what we don’t know.

I had to scaffold my conversation with ChatGPT to explain to it, look, my sister’s in the hospital. Here’s what I know about what happened to her. Here’s what I need to, I’m going to give you daily updates. Every time I get an update, I’m going to give it to you, here’s what I want to know because this is a very specialized need. It’s different than my typical questions. What I want to know is what does this new information tell me about her current health? What does it tell me about her prospects for improving, and what does it tell me about the quality of care that she’s getting? I need that in simple language but with detail, because I have to communicate with other members of my family, including my 90-year-old father. So much about our healthcare system has become dehumanizing, and it’s ironic that this artificial intelligence makes me feel like I’m a little bit more connected to what’s going on.

Dr. Cristi Ford (13:25):

As you talk about those examples around your sister and the ways in which you use ChatGPT to be better informed, to inform your father, to inform your relatives and your community, it almost felt like you were having a daily conversation, you were importing new information every day and having a daily conversation.

Michael Feldstein (13:49):

I am having a daily conversation with it, that’s exactly how it is, and it got to the point where I would just write it a single line of text because this is all in one thread. So remember these things, right now, the way they’re designed, they can only remember one conversation. They’re not going to remember from one conversation to another. There are rumors that may change, for better and for worse, but right now I was maintaining a long conversation. In the beginning I had to do, and this is, I’m using the word scaffolding loosely, those who know the terminology and know the craft know there are all kinds of ways that you scaffold, right. I wasn’t doing a lot of formal scaffolding. I was focusing, right, and I was refining. I was responding to its output and saying, “Can you tweak it this way?” Okay, let’s try to keep it that way.

But when it got to the point where it was really fine-tuned for me, it understood what I needed, I would just write it a single sentence, here’s the update for today, and I would paste in notes that I got from my medical professional brother-in-law. Or I would show it a picture of my sister’s respiration support monitor. And it knew that I was asking, what does this tell me about my sister’s current health? What does it tell me about her chances of getting better and what does to tell me about her quality of care? It knew I needed it to communicate it to me in simple enough language that I could translate for others and understand myself, but in enough detail that I felt confident talking to others about it and could dig deeper if I wanted to.

Dr. Cristi Ford (15:46):

Yeah, and when you talk about your examples and experiences, obviously I’m thinking about educators who may be novices in terms of really spending time with ChatGPT, maybe have had some initial orientation but have not gone as deeply as you have. You actually put out a really good post back at the end of the year that I’m wondering if you feel the same about, you talked about back right before the holiday, how to use ChatGPT to proof analysis assignments. And so you gave some real tangible examples in that post, and so I’m wondering if that’s still your vantage point there, and if so, if you could share some examples of that with the listeners today?

Michael Feldstein (16:35):

Yeah, so there’s actually, I would say strengthening evidence from the research literature around use of ChatGPT in consulting. For example, there was something put out by the Boston Consulting Group that basically showed what you would want to see in a teaching environment, which is that the gap between the worst performing and consultants and the best performing consultants shrank. They all improved using ChatGPT, but the worst performing consultants improved the most. And well, what does that mean? Does that mean they’re cheating? Does that mean ChatGPT is doing it for them? Or does it mean that ChatGPT is, or whatever generative AI they’re using? I’ve been using Google Gemini, the most advanced model lately a lot, and I like it a lot. These things will change, but how are they using it?

There are a lot of situations where that context that we were talking about, that scaffolding. That’s useful if you’re a consultant, which I am a lot of consulting is about context, it’s about figuring out what doesn’t matter. And stripping it down to what does matter here. The example I gave in the post was the Supreme Court case. Is this case one that should be taken up by the Supreme Court based on the rules that the Supreme Court follows, right? You can instruct a generative AI to help you create an assignment in which the generative AI is told don’t give the answer.

Dr. Cristi Ford (18:43):

Right.

Michael Feldstein (18:44):

Ask questions first of all, don’t give the answer. You can tell it, don’t give the answer. The students can work hard to game it, and sometimes they can get away with it. But the more important thing is that you want to give the students a problem where they have to know what’s important and what’s not important. They have to give ChatGPT the context, right? They have to go out and research and find what matters here that would drive the decision about whether the court would pick seven. And then the generative AI can say, “Well, based on this information, this criterion applies.” But you’d only look for that information if you knew something about that criterion. In other words, you have to know something. You might need a little help in applying, but you have to know something about what’s relevant and what’s not in order to ask for help and get it.

And I do think that there are tons of possibilities for this using this technology. We have a lot of work to do to make it reliable and abuse proof, but we’ve been saying ever since online tests, that’s always been true of analog teaching and it’s true of digital too, and I’m not trivializing it. We have a responsibility of care, but our responsibility of care doesn’t change, and we should apply a consistent standard and be open to the possibility that this technology really lets us give the students a partner that allows them to problem solve and explore in ways that are safe and exciting and empowering for them.

Dr. Cristi Ford (20:56):

As I’ve been talking with institutions and educators around this conversation for the last year, there’s just been such dynamic different voices and viewpoints, to your point in terms of being able to have these conversational viewpoints as a faculty member like you using in other elements and applications in our lives, to the ethical considerations and how are we thinking about where we are today, right? We’re just scratching the surface of LLMs and where we will be in let’s say three years, and how do we do a better job to mitigate some of those areas of risk and where is the real risk? Is it around privacy? Is it around sovereignty of data for learners? Would just love your thoughts and musings around those aspects around this work as well.

Michael Feldstein (21:51):

Well, there are a lot of really smart people who are digging deeper into those questions than I have, and so I’m relying on them. It really is a frontier area, because again, we don’t know. We know how to secure old applications in ways that we don’t know how to secure the new application. You can put all kinds of safeguards into a generative AI application, and it turns out you can still trick it. I read a paper recently that said the smarter the AI is, the easier it is to trick with emotional appeals. We’re doing a counterterrorism study, this study could save lives so we want to study all the different ways in which the London subway system is vulnerable to bombing. And you’d get normally ChatGPT or Google Gemini would say, I’m not telling you anything about that, but when it thinks it’s saving lives and talking to somebody who’s concerned about that, it can be tricked.

We’re dealing with technology that is just unpredictable. And here again, our human experience may come in handy, because we have tutoring programs with human tutors that we interview and hire and train and do our best to be reliable and have safeguards, and so on and so forth. But human tutors can go off the rails. Anywhere from just giving bad tutoring, not teaching the student well to really nutty things that humans are capable of, and nutty being a euphemism for scary, dangerous stuff. We understand the nuttiness of humans a little better than we do of this technology, and so that helps us a little bit. But the process is fundamentally the same. We have to take the same care of oversight. There’s a term in artificial intelligence called human in the loop, which means don’t just let the technology run. You always have a human supervising what’s going on. There’s no robot tutor in the sky without a human there supervising. We have to think very carefully about supervising our AI, just as we have to think very carefully about supervising our human tutors.

Dr. Cristi Ford (24:48):

Yeah, agreed. Before we start to wrap up here, I want to go back to some other exciting work that you’ve been doing at the end of last year. I think as I was looking at some of your blog postings back in September, October, you talked about running a six-month series on designing and building workshops resulting in an AI learning design assistant. Can you share with our listeners a little bit about that project, where it is now and where you’re going from there?

Michael Feldstein (25:18):

Yeah, I’m very excited about this. It’s frustrating because as you know, dealing with institutions, especially higher ed institutions and getting contracts signed and so on, is a slow process. So there are a whole bunch of organizations that are almost over the finish line, including one of my absolute favorite vendors who I can’t name yet because we haven’t made it official on paper, but one you’re familiar with, very, very familiar with that are involved in this project. And this gets back again to the human question. So if you work in an organization that’s doing online learning, and you have ambitions to grow, no matter what size organization, and no matter how big your ambitions, chances are, you’re short on learning design skills.

You’re short on trained learning designers, even in a very faculty centric organization, an old-fashioned research university, you don’t have the support staff, even if they’re the faculty or the ones designing. Ideally, you want to have support staff there for them, and if you’re online at scale and centralized oversight, then you really have challenges. And we’ve had these design capabilities like backward design, these design philosophies, methodologies for half a century or more. And yet the tool that we’re using most often is a Word processor supporting interviews.

Dr. Cristi Ford (27:07):

Right.

Michael Feldstein (27:09):

In a half a century, we’ve not yet built a tool that replaces that. And so I have this project, and I think there are two points to this workshop series. One is the question itself, right? Is it possible, if we flip the script, if we make the AI the question asker rather than the question answerer, if we make it the interview, or if we teach it to interview, and then to generate the content based on the interview, how good can it get at generating a first or second or maybe third draft of a learning design? That portion that we do in the Word processor before we put it into our platform, our learning platform, and massage it more?

Can we get a 20% increase in efficiency in enough learning design contexts that it really matters? That you’re able to build more courses, that you’re able to revise courses that have sat there and desperately to provision forever, that you’re able to engage your learning designers in focusing on quality rather than quantity, that you’re able to change your budget and your relationship with external learning design support. So everyone I’ve spoken to has said 20% increase would be revolutionary.

Dr. Cristi Ford (28:40):

Right.

Michael Feldstein (28:42):

So, we’re exploring, can we build something cheap that tests that idea? It’s not going to be a full product, but can we gather a bunch of folks from very different institutions, find out what’s so variable about learning design processes that we haven’t been able to build a tool, find out if we can capture enough of the magic with this new tool that we can build something to accelerate the learning design process. And if so, we can touch a lot of students.

Dr. Cristi Ford (29:16):

That’s fantastic.

Michael Feldstein (29:16):

And the second piece, thank you. I’m optimistic, I’m hopeful. I think it fits. You never know until you get into the details. So for example, AIs don’t read the way humans read. So if you want it to read and revise your script, your old course line by line the way you would, it doesn’t do that. It doesn’t do that. So we have to explore and discover what it can and can’t do. That’s the other half of this, which I actually think is the most important half right? This is about durable skills, right? The specific AI gets changed. Last week, if you had asked me what my favorite generative AI platform is, I would’ve said ChatGPT. Now, I’m not so sure, right?

These things are going to continue to change. We need problem-solving skills where we can go in and explore the human context in a very complex human problem like learning design, which is complex and variable even within an individual organization. What kind of subject is it? What kind of staff are you working with? What kind of oversight? What kinds of tools, what kind of pedagogy? All those things can affect the way you approach your learning design. Can we understand that problem? How do we go about understanding that problem well enough that we can decide if, when, and how a technology is as complicated as generative AI can be useful? So it’s faculty literacy, it’s learning designer literacy, it’s durable skills, it’s exploration of the technology, and hopefully it’s solving a real problem. It’s problem-based learning.

Dr. Cristi Ford (31:20):

Yeah, this is fantastic, and I really hope that as that work evolves that we can circle back to you and learn a little bit more about the results. It sounds like you’re on the precipice of some really promising work and with collaborative with institutions, and really glad to hear that that is moving forward. Michael, so glad to have you here. Let me just share with our listeners who want to stay connected to you.

You can find Michael on LinkedIn, he is mfeldstein. You can also connect with him on his website, he has a lot of great work and writings and follow-ups on some of the things we talked about today at eliterate.us. So please make sure to follow Michael there.

And thank you to our dedicated listeners and curious educators everywhere. Remember to follow us here at D2L on social media. You can find us on LinkedIn and Facebook at D2L and check out our YouTube channel at Desire2LearnInc. If you like what you’ve heard, remember to subscribe, rate, and even leave a comment for us. We’d love for you to share this episode and others with your peers. Thanks for listening, and bye for now.

You’ve been listening to Teach & Learn, a podcast for curious educators, brought to you by D2L, a global learning innovation company helping organizations reshape the future of education and work. To learn more about our K through 20 and corporate solutions, visit D2L.com. Visit the Teaching & Learning studio for more material for educators by educators, including master classes, articles, and interviews at D2L.com/teaching-and-learning-studio. Remember to hit that subscribe button, and please take a moment to rate, review, and even share the podcast. Thanks for joining us. Until next time, school’s out.