How Learning-Focused AI Looks Different From General AI
The challenge with using general AI for learning
When Rajen Sheth published 10 Lessons on How to Drive Learning with AI, one theme stood out: educational AI should not be evaluated like general-purpose AI.
Large language models excel at retrieving information: ask a question and they generate an answer. But learning requires something different.
Research shows that approximately 75% of students do not know enough about a subject to ask the right question in the first place. If learning depends entirely on student prompts, many learners will struggle before they even begin.
Why AI should ask the questions
Traditional AI systems wait for a user to initiate a conversation.
Learning-focused AI reverses that model.
Sheth argues that effective educational AI should identify where students are struggling, ask the right questions, and guide learners toward conceptual understanding. Rather than simply delivering information, the system should function more like an instructor helping students work through ideas step by step.
The goal is not faster answers. The goal is deeper understanding.
Why guardrails are essential
Another major difference between general AI and educational AI is the role of guardrails.
Most conversational AI systems are optimized to keep interactions going, but educational systems require a different objective. Educational AI must know when to continue a conversation and when to stop.
Students need support that remains aligned to specific learning goals rather than wandering into unrelated topics. They also need protection from inaccurate, distracting, or counterproductive interactions.
In many cases, ending a learning interaction at the right moment is just as important as starting one.
Supporting instructors, not replacing them
Sheth repeatedly emphasizes that AI should function as an extension of faculty rather than a replacement for instructors.
Educational AI must align with classroom content, course materials, and teaching approaches. Faculty should maintain control over learning objectives, instructional methods, and the student experience.
The technology can then provide valuable feedback by identifying concepts students struggle to understand and highlighting where additional instruction may be needed.
The future of AI in higher education
Sheth believes higher education’s greatest AI opportunities will come from systems intentionally designed around learning outcomes.
The distinction matters.
General AI helps users find information. Learning-focused AI helps students develop understanding.
As institutions continue evaluating AI strategies, that difference may determine whether technology becomes another digital tool or a meaningful driver of student success.
Transcript
Wes Smith (01:58.35) So you recently published a piece called 10 Lessons on How to Drive Learning with AI. And what I appreciate about it is that it really cuts through the noise with
practical and buildable principles. So rather than staying at 30,000 feet, I just want to walk through a few of those takeaways and make them concrete for campus leaders and faculty and frankly, the policymakers who we’re trying to get this information to right now.
Rajen Sheth (02:43.778) Perfect, that’s great.
Wes Smith (02:46.498) So OK, let’s start with one of the principles that you’ve articulated. You’ve said that effective learning is instructor-led. What does it look like for AI to ask the question instead of waiting for the student? And I know that you have a stat in there that 75 % of students, don’t yet understand the concept well enough to even ask the right question. So what does it mean for how AI has to behave?
Rajen Sheth (03:13.526) Yeah, I think it’s a great question. It’s interesting because, you know, at Google, I was part of the development of a lot of the underpinnings of what became Gemini. And what was interesting there is when a lot of that was built, a lot of it was built around the concept of information retrieval, which is ask a question, get an answer, ask for something, get content, that kind of a thing. But it wasn’t built for learning.l l
And that stat is actually true. What we’ve seen from studies is 75 % of students actually don’t have a question to answer. So if you ask them to just use a chatbot, they’re not going to know exactly what to ask for. We’ve now taught over 100,000 students with Chiron. And we’ve seen exactly that play out. And so what we chose to do from the very beginning is we asked the question. We figure out what is the right question to
ask to that student at that time, and then use that as a way to stimulate learning and stimulate their understanding and then guide them to the answer with the right teaching rules. And we found that as a result of that, students actually get to a deeper level of understanding. It’s very different than how students are using AI right now, but it leads to better results.
Wes Smith (04:31.938) That makes so much sense to me because usually when you’re starting into a new subject, a teacher can assess where students are. That back and forth with students gives them a little bit of an ability to assess to say, okay, we’re missing a few key concepts. So that’s essentially where you’re starting.
Rajen Sheth (04:53.974) That’s exactly right. And the highest, hardest bar here is conceptual understanding. And if you don’t understand the concept, you can’t keep practicing. can’t get deeper and deeper in the subject. What we find with lot of students is that they’ll have holes in different concepts that they haven’t been able to get over. And then that hurts them down the line. And so we wanted to figure out
how do you use AI to get them to that conceptual understanding and aid the teacher and aid the instructor and faculty in helping their students get there.
Wes Smith (05:28.63) Right, right. Well, OK, so you also write that most AI tools, they’re designed to keep conversations going, not to keep them on track. And I think we all see that in our daily use of AI, right? It always ends, your prompt always is answered, and then another question is posed. Do you want me to do this? Do you need help on this? But you’ve talked about guardrails need to be up here.
Rajen Sheth (05:41.889) Yep.
Rajen Sheth (05:55.138) Mm-hmm.
Wes Smith (05:57.39) What do those look like when it comes to learning? What kind of guardrails do you have there?
Rajen Sheth (06:01.644) Yeah, safety is paramount here because a general AI system can take you in all different directions and can be distracting and in some cases even destructive. And so what we need to do is keep it on topic and keep the learning objective in mind as we talk to the student. And so that’s really what we’ve done is that we’ve enforced really strong guardrails to keep it on topic and guide the student in the right learning direction towards the learning objective.
The other thing is, of course, know, gargling against harmful conversations and making sure that those are captured as well. Another thing you’ve brought up that we’ve had to work really hard to do is not only learn how to do that, but learn when to stop the conversation. And that has been actually one of the trickiest parts about AI, because as you said, the tendency is to keep going and going and going. In some of our early trials, know, the AI would ask like,
20 questions and keep going back and forth with the student and the student would eventually give up. But we now have gotten smart about when to end the conversation to know how to get the student to where we need them to go to and then move
Wes Smith (07:12.546) Yeah, that makes sense. I mean, that’s different than just general AI in my experiences. You’ve had to program it for the purpose of learning. That also kind of leads me into this next question. We have decades, maybe centuries of learning science. We know how people learn. And so if an institution now is evaluating an AI learning tool, what
Rajen Sheth (07:33.548) Yeah.
Wes Smith (07:40.736) are the learning science principles that they should look to or look for that these tools can use. So it’s teaching and not just answering questions.
Rajen Sheth (07:53.292) Yeah, absolutely. And I think the interesting conundrum here is that everyone wants to look for proven outcomes. And AI is so new that we’re just starting to show those proven outcomes. But what is proven, to point you made, is learning science. We know what techniques work and we know what techniques don’t work. And so what we’ve decided to do is build our system around learning science and around those proven techniques to
show that those can actually lead to impact. And a few of the key things that are there. One is this concept of backwards by design. And so when a student has a learning objective, we go backwards by design. So what we do is we take that learning objective, we think about what are the questions we want that student to be able to answer at the end of it. And we work backwards from there to try to get to the right learning modules to get them to that answer.
The second thing is kind of this concept of the zone of practical development. What is the right question to ask to get them into that productive struggle? And then that is shown to be a way that you can actually really guide students towards getting and stretching themselves. A third is analyzing what are the right teaching moves to put in place. And so a generally our system will always go towards kind of giving you the comprehensive answer.
What we’re doing is we actually classify it to the right teaching move. And then we build the next response based on that teaching. And so that makes it such that we’re acting in the way that a strong, pedagogically strong instructor would do. And then the final thing is analyzing and making sure that we understand not only that we help the student, but where does a student have more holes that we can help them with? And that can help the progress on going.
Wes Smith (09:51.951) Yeah, I mean, this is just music to higher education ears, understanding that AI alone can be helpful. There are ways that it could be helpful. But when it’s built on the right pedagogy, when it’s built on learning science that has been refined and proven out over centuries of learning,
It makes it just so much more reliable for instructors to be able to use. And I know that’s a huge issue. You have to have instructors that feel confident in how these tools are used.
Rajen Sheth (10:31.264) Yeah, absolutely. And I think that is the key thing is it cannot be the AI alone. It has to be the AI in concert with the instructor. And how do you kind of go back and forth in the right way? How do we dovetail to be kind of an authentic extension of that instructor? And then how do we feed the right data back to that instructor so the instructor can know what to do next for their class?
Wes Smith (10:55.33) Yeah, yeah. So how do you make sure that AI reinforces what’s happening in the classroom instead of teaching something that’s totally different or it’s a different version of the course at least?
Rajen Sheth (11:02.423) Yeah.
Rajen Sheth (11:08.566) Yeah, I think that is something we’ve had to a lot of time and effort into because AI by itself might teach something in a very different way. So for example, if I’m learning a concept in math, there are probably about 10 different ways to teach every one concept that’s there, but how do we reinforce the way that it’s being taught in the classroom? So part of that is that we dovetail with the material that the teacher has. They can upload in their
material and then we build the lesson with that in mind, with those concepts in mind and the way that they’re teaching in mind. A second thing is to give them full control. And so rather than them just kind of handing this tool to the student, they can control what does it say? How does it say it? What is it leading the student to? They can tweak it so that it is truly an extension of themselves. And then the third is that loop back that
we analyze the conversations and we come back to the instructor with, hey, you know what, 12 of your students are really struggling with this thing. Five of them are struggling with this thing. And that helps them figure out what to do next. And that kind of makes it such that it is a true extension of the classroom.
Wes Smith (12:17.57) Yeah, yeah, OK, so I want to finish with this question. The goal is impact, right? And when we’re talking to members of Congress and their staff, the question wasn’t just, how is AI being deployed in higher ed? It was, how is it impacting students? So it’s not just necessarily about deploying it. It’s about, how are we making a difference? And how do we measure?
what the impact is in higher education. Do you have some insight on that?
Rajen Sheth (12:47.362) Yeah, absolutely. think what there are two parts of this. One is how it’s built and the second is is what are the proven results. And we talked a lot about how it’s built with a lot of learning science and pedagogy in mind. And we’re now seeing that in the results. We’re seeing institutions where their pass rates are going up significantly six to nine percent in classes. They’re getting to the highest pass rates that they’ve ever seen as a result of putting chiron in.
their engagement with students is going way up. In one case, we saw engagement go up by about 7x in comparison to other material that’s there. And students have said that they really love the experience and they’re learning more out of it. And so it really kind of drives towards that goal, which is how do we help the students that need the help the most? And how can we get them through the higher education experience so that they can get to their goal?
And that’s really what we’re seeing in the results.
Wes Smith (13:48.056) Right. OK, so I want our listeners to remember you have a lot of experience in AI. I know your Google experience, pretty significant, and working with AI in a lot of different ways. But now at Chiron Learning, you’ve focused in on the use case for education. How has that focus changed the way
that you see AI.
Rajen Sheth (14:17.954) Yeah, the way that that has changed the way I see AI is that when you’re looking at a particular goal, you can do everything to reach that particular goal. Not only the technology, but how we work with customers and how we work with institutions. All of that goes towards this. You could take a raw LLM technology and get that to students, but it’s not purpose built. And so all of the things we talked about building in
learning science, understanding the student, understanding what they need, driving that engagement. All of that is what it’s taken to them lead to these great results.
Wes Smith (14:54.412) Yeah, okay. Well, Rajan, thank you so much for coming back on the show and for translating what you’re seeing into practice. And these are good practical lessons that our leaders can really act on. We appreciate your time.
Rajen Sheth (15:08.842) Great, thank you, we really appreciate it.