Sesh Murthy, UCSD, on what separates AI that helps students learn from AI that just hands over the answer
In episode 1 of The Learning Curve, Sesh Murthy, Strategic Research Initiatives Specialist at Halıcıoğlu Data Science Institute at UC San Diego, talks about what it takes to design AI support that actually helps learning. The conversation explores productive struggle, scaffolding, instant feedback, and how thoughtful design can help learners build confidence, practice more, and improve outcomes.
Sesh Murthy leads the Aspire project at UCSD, which studies how AI tutoring changes outcomes in gateway STEM courses. He was previously dean of computer science and software engineering at Western Governors University.
About the guest
Strategic Research Initiatives Specialist, Halıcıoğlu Data Science Institute, UC San Diego
Murthy has spent his career on applied AI, and came to education through tutoring. He was a research staff member in IBM Research's AI group at the T.J. Watson Research Center, building multi-agent systems for paper and steel mills, then held vice president roles at IBM Global Services including IBM Cloud Managed Services, and co-founded the security automation company Cloud Raxak. As dean of computer science and software engineering at Western Governors University he began the AI-tutoring work he now leads at UC San Diego as the Aspire project. He holds a PhD from Carnegie Mellon, has more than 20 patents, and has won the AAAI best paper award and the Daniel H. Wagner Prize from INFORMS.
Murthy’s starting obstacle wasn’t the technology. It was that “the same 5% of students who would go out of the way to do anything required to succeed are the ones who use tutoring systems” — leaving the other 95% untouched, which has dogged intelligent tutoring systems for decades. The fix came from a faculty member rather than an engineer: add a second homework, due Monday, where AI help is allowed. Usage went past 60%.
Roughly 250 students were tested on entry to establish their mastery of each concept, then tracked across two midterms and a final. Incoming knowledge was a strong predictor of the final result, as everyone expects. What surprised them is that the number of questions a student attempted was comparable — Murthy’s own words, hedged carefully: “It’s not statistically significant. It is probably at the same order of magnitude. But that itself is remarkable.” If practice weighs about as much as preparation, a student who starts behind still has a route through.
In Math 3B the failure rate dropped on the order of two thirds, and Murthy is careful about the credit: a significant portion came from the course redesign, not the tutor. “But that was possible because the AI tutor was there. So you can’t take those two apart.” His candidate for what mattered most is the least glamorous feature — knowing instantly that an answer is wrong. Programmers have had that since compilers; maths students never did.
Sesh Murthy  · UC San Diego  · 01:40
Edited for clarity and length. Substance unchanged.
Sanjay SrivastavaThanks, Sesh, for joining this podcast. As you know, this is the very first one, so I'm just hoping to learn a few things and have some fun. I just wanted to get started by asking you to give some context — describe your professional journey, and how you got to where you are. And then, of course, we'll have a lot of fun talking about where you are right now.
Sesh MurthyYeah, so, big question. I'll try to keep a short answer and we can explore. I was working with inner city kids, tutoring them at Yonkers Partners for Education. And you learn a lot by working with students. And then the question always was — I've got a long history with AI, so I was wondering, can't we use AI to do some of this?
Sesh MurthyAnd well, ChatGPT came out, and it turned out that I had been recruited to work at Western Governors University. And my boss at that time, Ashutosh Tiwary, and I sat down at dinner one day. And there was a lot of fuss. We said, God, this is going to change so much of what we have to do. From the content creation process to the grading process to diagnosing what the students know and did not — I mean, we didn't know what was possible, but we thought it would do that. So we started that project that day.
Sesh MurthySo I started it at Western Governors and I continued this at UCSD. It's been a great journey trying to see how you can use AI to help improve student outcomes. One of the key things I have been interested in is how to prevent AI from making student learning worse.
Sanjay SrivastavaThat's a good starting point.
Sesh MurthyYes. So what can cause problems? And so yeah, that's what we're doing. I've had the good fortune at UCSD to work with an extremely strong set of faculty across cognitive science, data science, AI, the education department, and most importantly — well, the math department has been extraordinarily generous with their time, right from the head of the department, Michael Holst, all the way down to all of the individual faculty who spent far more time than one can believe. So we've been very lucky that we got them. Other instructors in other courses, like in a course at the GPS school here, have also contributed. So we've got a very wide set of people who work with us. So that's my journey. That's what we've done.
Sanjay SrivastavaThat was interesting. You mentioned you had an experience tutoring inner city kids. Can you connect the dots in some sense — that experience, and looking at exploring AI? Is there a dot to be connected?
Sesh MurthyYeah, there is a dot to be connected. So there is this paper by Bloom which talks about one-on-one tutoring. And I find that every time I read it, I learn some more. There are some papers, like some of Simon and Newell's papers in AI, that are that way for me. So I go back and look at that.
Sesh MurthyAnd it says that the outcome for students who have one-on-one tutoring is two sigma better than students who are in a general classroom. I'm paraphrasing. And ever since then, intelligent tutoring systems and the people in them have been trying to figure out how to provide that two sigma effect without actually having a one-on-one tutor.
Sesh MurthySo I saw the value of the one-on-one tutoring when I was working with students. You instantly figure out what they don't know. You don't have to make them go through, like, hundreds of questions. One question that's reasonably difficult and you have an estimate of where they're sitting. That's what we set out — that's what we wanted to create.
Sesh MurthyA lot of people believe that AI, if used correctly, can improve student learning. And most of them believe that the best way to go about it is to learn the mechanisms that a tutor uses in one-on-one tutoring — that a human uses in one-on-one tutoring. And so I've been going back to my experience tutoring kids. And of course, all the faculty here have lots of experience with that. So I think that's the link.
Sanjay SrivastavaI was going to say — so what do you think? I mean, at this point we've been working together for two years. What has this project been like, and what have we discovered? Do you want to talk through the specifics of Aspire?
Sesh MurthyThere's one important thing that I want to point out, that I have watched other such projects flounder on, which is that they tried to build everything on their own.
Sesh MurthySo the reason that I chose to use Vocareum as the platform on which we've built the AI capabilities was that about 90% of what we required — in terms of the user interface, the connectivity to backend systems like Canvas and being able to ingest materials from there, the FERPA regulations that you need to comply with — all these already existed. And so we were able to get started on day one with the key problem that we wanted to solve, which was: which workflows are likely to succeed?
Sesh MurthyEverybody says that they use evidence-based workflows from learning science. And these are pretty well known — the best papers are out there, we can take a look at them. The difficulty in the past was that you built out these workflows and they were extremely brittle. So the combination of using the Vocareum platform and then having these extremely powerful LLMs at the backend enabled us to build out these workflows — personalized, mastery-based learning — extremely quickly.
Sesh MurthyThe literature on intelligent tutoring systems says that you need to have a domain model and you need to have a student model. And it was far easier to build this out with Vocareum than to do this independently in some other place. And we did that with you. A student model is available, and then you're able to improve your approximation of — so, the student model. The ITS people call it the student model, but everybody else calls it a digital twin. Because it's the AI tutor's estimate of what the student knows at any given time on each concept. There are a few more things that you have in a digital twin and we're improving that. And what they know, the level of mastery, is not just a scalar that goes from 0 to 10. We think that there is more to learn from there, but anyway, going off topic.
Sesh MurthySo having that domain model built quickly, having the student model of what each student knows, and then being able to connect up with the existing courses — for example, in Math 3B — we were able to do this in weeks rather than in years. Then we were able to iterate quickly. We learned a few things, which is that if you just put a tutoring system out there and say, go use it, hardly anybody uses it. We can touch more on that if you want to.
Sanjay SrivastavaYeah, I remember you saying that — the whole 5% problem. That was pretty eye-opening. Why don't you just talk about it? I have mentioned it to other people and sent the reference, and everybody says it definitely hits very close to home. Because when you see it articulated that way it's pretty amazing. But go ahead.
Sesh MurthyYeah, this is a really good paper — I'll send the link later — which talks about the 5% problem, which is that the same 5% of students who would go out of the way to do anything required to succeed are the ones who use tutoring systems. And as a result they improve. But the remaining 95% of them, for whatever reason, don't get around to using it. Maybe they don't know how simple it is. Maybe they don't know what value it provides, et cetera. And this has been a problem that has been plaguing ITSs, and AI in education in general.
Sesh MurthyWe were lucky that we had strong faculty members who understood this. Professor Ko-Woon Ohm, who is now a faculty member at Berkeley, suggested that the way we could try and solve this problem of not enough people using the AI tutor was to give them extra homework that was due on Monday.
Sesh MurthyJust for context: the materials are taught during week zero, then in week one they have to submit a homework on that. And in previous courses, they had a homework that was due on Wednesday. So she suggested that we add another homework that was due on Monday. And this was where they could use AI help to get started. That's where we first saw the 5% problem going away.
Sesh MurthySome students did choose not to use AI. We set it up that way so that they could submit their homework without it. We agree that if you can get started on the problem, work on it, go through the productive struggle, then you will learn it far better than if you depend on a crutch. Some percentage of students did not use the AI at all — they tended to be people who knew the material already, or worked harder, and so on. But a significant percentage of the students, some 60-plus percent of the students, used it.
Sesh MurthyAnd it was instructive to see how they were using it. Many of them started the homework at eight or later on a Sunday for a Monday submission. Why am I mentioning this? This is usually a recipe for disaster. You can get very bad grades on it. But the AI provided the scaffolding so that the students could get started. It takes away one excuse for not getting started. This is our theory anyway. It provides enough of an incentive for the students to get started on that homework.
Sesh MurthyAnd the interesting thing is that they all got 100 on their homeworks. And if you go back and take a look at the transcripts, you will see that the AI has not given the answer away. It has scaffolded them enough so that they moved far enough that they'd solve the problem. So they do go through the productive struggle.
Sesh MurthySo that Monday homework, which the students had to do, provided the first instance where they relearned the material after the class, submitted the homework, and then they were able to gain some confidence that they knew the material. On Wednesday, they had a homework which they could not use AI for — that had to be just written on hand and paper and uploaded. We haven't got the evidence back about whether they improved on that, though anecdotal evidence says that they did do better. And then we had a quiz on Friday, a proctored quiz, where they could not use AI.
Sesh MurthySo they had seen this material three times already — the lecture, then on a Sunday night, probably at midnight, where they were submitting their homework, and then doing stuff on Wednesday. So they tended to do reasonably well. And then they were able to use the AI for doing practice tests before the midterms and the finals. And because they had used it through the course, they saw the value, and a significant percentage of them used it. So that 5% problem went away.
Sesh MurthyAnd they do say in that paper on the 5% problem that you have to redesign the course to do it. But this is a concrete example of how we redesigned the course to get past that 5% problem. And that's probably the reason why we were able to get significant results.
Sanjay SrivastavaYeah, so that's what's exciting. It's a redesign, but it's not disruptive, right? You're just adding another homework and saying you can use AI. It feels like it's incorporated relatively painlessly.
Sesh MurthyYeah. I credit that to two things. One is the LLMs have got damn good, and we're using them at the back end. And the other one is the Vocareum platform — that you guys made it extremely simple for the students to submit the homework. Otherwise it's "the equation editor is not working, I don't know how to do this," and that becomes a problem. This is like when I don't want to go to the gym, I say I can't find my clothes, I'm not going to the gym. That's the kind of excuse that I make. So now I know how to get past that problem.
Sesh MurthySo we're trying to get past that exact same issue with the students. Just get everything that could stop them out of the way, so that they attempt to solve the problem, provide the help necessary, scaffolding, and then help them to go forward.
Sanjay SrivastavaOne thing which you mentioned, I just want to drill down on. You said on Wednesday it gave them the confidence that they had seen the problem and they had gotten 100 with AI help. So that confidence — you believe it was significant? Maybe they're engaging more because they're confident. You think it was a contributor to the improved outcomes?
Sesh MurthyYeah, now we're getting into third-level effects. Let me put the disclaimer upfront. We've got this wealth of data that we have gathered in 2024, the fall of 2025, and now in winter 2026. And there's a lot of stuff to take a look at there. But that data will show me correlations and it will not show me causation. We understand that. But even then, I'm going to speculate.
Sesh MurthyI think — and now I'm going back to my work with one-on-one tutoring with students — I think that practice makes perfect. And the more you do it, it doesn't matter what anybody else tells you. It is the results of what you did that you are able to then use to build the confidence. I think that that actually is a critical part of it.
Sesh MurthyOverall, we have reasonably strong evidence that students who used the AI tutor for attempting and answering more questions — so they didn't have to use it just for the homework, they could say, give me another question like this, and solve that; they could, at homework time and for the midterm and the finals, they had practice questions and they could attempt those and ask for extra questions — so there's a pretty strong correlation between students' attempts at more questions and their outcome on the course. In fact, and we have to check this, but we believe that it is better correlated with the outcome than their incoming knowledge level.
Sanjay SrivastavaYou know, that is amazing. That is just — that's astounding. Because you always believed that the biggest indicator was where do you start, right?
Sesh MurthyYeah, where we start. So this is really interesting, Sanjay. As we progressed quarter by quarter, we recognized in the summer of 2025 that one of the things that we need to do is to test the students coming in, to understand their knowledge level, and then to give them guidance on what their mastery level is of various concepts — remember the student model that we were talking about.
Sesh MurthyAnd so we provided that, but that was then a body of evidence that showed us where, across these 250-odd students, they came in for each concept that was being taught and tested. And then we were able to look at the progress in the first midterm, second midterm, and the final for students who attempted more questions versus students that attempted less questions. And it was very clear that the students who attempted more questions, their progress on the concepts was higher for the midterms and the finals.
Sesh MurthyAnd we found that how much they knew when they came in was a very strong indicator of how they did in the course. But the number of questions they attempted was even stronger. It's not statistically significant. It is probably at the same order of magnitude. But that itself is remarkable, isn't it, Sanjay?
Sanjay SrivastavaThat's amazing. That's insane.
Sanjay SrivastavaSo in terms of the data, anything else before we move on? I realize that some of it is just correlation, as you said, so we need to do more research. But from a data perspective, the Harvard study talks about something like a one sigma difference with a well-designed tutor. So that was pretty exciting — that more and more people are beginning to come to the same conclusion, that a well-designed AI intervention — in my mind, there is no doubt left that it makes an impact. Do you agree?
Sesh MurthyYeah, so let's address that Harvard study. I want to paraphrase it and probably get parts of it wrong. But there were two concepts, and students alternated. So half the class got AI assistance on the concept, and half got personalized tutoring with a person. And what was interesting was that the people who got the AI assistance actually did better than the people who got the personalized tutoring assistance. And they swapped it in the second concept, so they got rid of this issue we were talking about — fairness. And so that was a very designed study. That was the first instance I saw that there was evidence that AI could help to improve the student outcome. We can always say that it helps to make the teacher learn something faster. We can say that the students liked it in the end-of-quarter quiz.
Sesh MurthyBut what we have been able to replicate with your help here is that for this course, 3B, we were able to reduce the failure rate in this gateway math course by a significant number, of the order of two thirds. We're not going to allocate all of that to the math tutor, to the tutor alone. A significant portion of it was the course redesign to have the Monday, Wednesday, Friday. But that was possible because the AI tutor was there. So you can't take those two apart.
Sesh MurthyWe think that a significant portion of it was the instant feedback that the system is able to give. If you provide an answer that's wrong, there's incredible value in knowing immediately that it's wrong. You know, you and I both program. When I had to take a card deck and then feed that into the system, and I only ran it once a day, debugging that program was a little difficult.
Sanjay SrivastavaOh my God. And what if you dropped the card stack and lost the order?
Sesh MurthyNobody could go there, man. Sorry, you shuffled it. Right. So in programming, we all accepted that you get instant feedback on whether your program will run or not. And that was not there in math. And now we have that with the help of Vocareum, to put that there. We think that was a critical part of it. We think that the scaffolding that we provide is really a critical part of it. But net net, I think that that Harvard study that you referenced and this work that we have — I feel confident that a well-designed AI system, one that preserves productive struggle, is going to help improve student outcomes. So it has to be well designed. And the well-designed is part of the research you and I are conducting jointly, with UCSD and now several other universities.
Sanjay SrivastavaRight. And as I was saying before, I get excited about a couple of things here. Number one is that we are targeting it from day one — let's go see how we can reduce failure rate. How can we really solve a real meaningful problem in our education system, where students coming in with weak math proficiency could really derail their career? The second thing you're talking about, well designed — I totally believe that. And what gets me excited also is that it's not something to sit on the side. It can be integrated into the current architecture of how we teach. That is important. Otherwise we'll be writing a research paper and waiting forever for the new system to emerge, the new whole schooling system, and we'll all be dead by then.
Sesh MurthySo I'm really lucky that I met you and that you share the same goal of trying to improve the outcomes for these students. For whatever reason, these students come in either not knowing the prerequisites, or not confident that they know the prerequisites. And they are equally damaging, both of them. They might know it, but they're not confident, and so they perform badly.
Sanjay SrivastavaThat's a good point.
Sesh MurthySo it was always my thesis that with one-on-one tutoring you could fix the problem. But that doesn't scale. That's the problem. If you're talking of thousands of students, you need a lot of time. I actually don't understand why society cannot devote that time, but let's leave that alone for the time being — that's a discussion for another day. I think society should devote that time.
Sesh MurthyBut in the absence of the societal will to provide those resources, I think that what we have jointly come up with, and the entire mass of researchers who are working on using AI in education — there are some significant results that show that you can use AI, designed correctly, to improve the student outcomes. And in a world where people keep talking about all the bad things that can happen, I think we have focused on one good thing that AI can do, and how it can do it.
Sanjay SrivastavaRight. And I do worry that we're just not prepared for it as a society, only because we don't get good news like this. I mean, the last time the education system changed, I think, someone figured out in the industrial revolution that we can put lots of students in a class and teach them the same thing so they can come out. But if you think about it, nothing really has changed since then. So we haven't had the pressure to say, there is this really meaningful outcome, let's go incorporate it in the classroom. So it'll be pretty fascinating to see what happens this time.
Sesh MurthyI'm really excited. All the college and high school teachers are really excited about this. I think it's going to take — I mean, the AI is going to evolve. And the right decision we made is not to try to create an LLM on our own. We could have gone down that path and spent seven years trying to get up to speed. So we're using the off-the-shelf LLMs, which are progressing, and then our research is progressing, and we're able to put this together. So the faculty both in schools and colleges are quite excited about this. And the question is, what can we do next?
Sesh MurthyThat's a significant frontier. And if we're able to overcome disadvantages that occur because of poorly resourced schools, or lack of resources that are there at home, or whatever it is, and get people to the same level — the thing we talked about, which is that the level you came in with is at about the same correlation with the outcome as the number of practice attempts. That is quite powerful.
Sanjay SrivastavaThat's amazing. You have a shot. You have a shot. You don't have to give up just because you're starting a little behind. Very concrete way to level the playing field. That's amazing.
Sesh MurthyYeah. When we saw that, Sanjay, the first thing that we did was we recorded a short video that said, our evidence shows that if you practice more, you do better. Practice before your midterm and your final.
Sanjay SrivastavaYeah, and here's the evidence. If there's something really concrete you can do, that's amazing.
Sesh MurthyI think that we are focused, in the Aspire project at UCSD, on STEM. And I personally don't know whether AI will help or worsen outcomes for students in the humanities, and in other areas. So I don't know.
Sanjay SrivastavaYeah, and it's the well-designed part — at least my personal opinion, and I'm just quoting a very small anecdote, where I thought it was just a marvelous experience. But going back to the whole well-designed part, I don't know what that would mean. Hopefully other people have experimented with it. But you can sort of see the power.
Sesh MurthyI think that, as I mentioned, if I have to do a task — the students were doing stuff at 10, 11 o'clock, but they were getting the 100. That is true across the board. So if I'm trying to learn a Gita shloka — I mean, I could go back and say that I never learned Sanskrit properly when I was in school, I didn't do this, I didn't do that. Or I can say there's a wealth of resources available, I can get started and then go forward. And so, yeah, the AI has been extraordinarily helpful out there. So essentially I'm excited about all the things I wanted to do that I felt that I would not have the time to do.
Sesh MurthySo this is not just for others — I'm using it at work, actually. If I don't know a subject, reading a paper, I have to go through a productive struggle to understand what they're doing. That can be done an order of magnitude faster. I hope I'm learning as much. There's always that worry.
Sanjay SrivastavaNo, that's totally true. But I'm certainly doing a lot more. I do worry about that too, because I'm doing a lot more — am I learning as much? The other flip side is just the whole excuse thing you're saying. I would not be either learning or doing without the tools. So I still think net net it is positive.
Sesh MurthyYeah, this is the equivalent of my packing my gym clothes the night before, so that I don't have an excuse in the morning. That's the exact same thing here. The AI takes that excuse away — that it's going to be too difficult to get started.
Sanjay SrivastavaSo Sesh, it's been a lot of fun. Any final thoughts before we wrap it up?
Sesh MurthyI am really excited about what we can do in STEM. I think that for too many years we have let people who didn't have the requisite knowledge — we said that they could not succeed and go forward. I think that we have a chance now to open STEM careers for everybody.
Sesh MurthyYou know, it doesn't have to be somebody who is doing topology at a PhD level. To do many of these — a carpenter test — you need to be able to add and subtract fractions and do things like that. That might be where somebody is getting blocked in getting into the trades and doing it. I think that we could help there. So somebody, for example, coming out from the military who wants to get into a vocation — we could figure out quickly what they don't know, teach that to them exactly, and then go forward. So I'm really excited about what we can do with this.
Sanjay SrivastavaAbsolutely, and I absolutely share the excitement. Thank you, Sesh.
Sesh MurthyThank you.
Referenced in this episode: Bloom's two-sigma paper on one-on-one tutoring (02:50) · The 5% problem, as Murthy presented it at UC San Diego's 2026 Faculty AI Symposium (07:21) · Ko-Woon Ohm, UC Berkeley, who proposed the Monday homework (08:02) · UC San Diego Math 3B and the Aspire project (22:22) · UC San Diego's own account of the math tutor
That voluntary tutoring tools get used almost entirely by students who were already going to succeed. Murthy cites a paper on it: the same 5% who will do whatever it takes are the ones who log in, while the remaining 95% never get round to it — maybe not knowing how simple it is, maybe not knowing what it offers. The paper’s own conclusion is that solving it requires redesigning the course, which is what UC San Diego then did.
Material is taught in the first week. A new homework falls due Monday, and AI help is allowed on it. A second homework falls due Wednesday with no AI — handwritten and uploaded. A proctored quiz follows on Friday, also without AI. Practice tests with the tutor are available before midterms and finals. By the quiz, a student has met the same material three times. The Monday assignment was the specific change that broke the 5% problem.
No, and they checked. Every student scored 100 on the Monday homework, which sounds like the tutor doing the work — but the conversation transcripts show it withholding answers and scaffolding instead, moving students far enough along to finish themselves. Many started at eight or later on a Sunday night, normally a recipe for a bad grade. Murthy’s reading is that the tutor removes the excuse not to start: “It takes away one excuse for not getting started.”
Murthy paraphrases it and flags that he may get details wrong. Students alternated across two concepts: half received AI assistance on the first, half received personalised human tutoring, then the groups swapped for the second — a design that removes the fairness objection to giving only some students help. On the concept where they had AI, students did better than those with human tutoring. He treats it as the first real evidence that AI could improve outcomes rather than just satisfaction.
Murthy doesn’t claim it does. The Aspire project is focused on STEM, and asked directly about the humanities he says plainly that he doesn’t know whether AI would help or worsen outcomes there. Worth reading alongside the two-thirds figure: the result is specific to a gateway maths course, not a general claim about AI and learning.
Conversations on AI, learning, and workforce readiness with leaders across education, technology, and industry.
