Does AI work in schools?

What is to be understood from “LLMs in education” studies

What is to be understood from “LLMs in education” studies

I see a lot of thought pieces and takes these days on the impacts of large language models in education. Typically, they are presented as some version of “LLMs do not work for school,” which is almost definitely the case when you look at the studies that people are writing about when making those kinds of claims. There is a lot of evidence to suggest that using LLMs in education to pursue improved/increased/better student learning is frequently a mistake. This study demonstrates this clearly: when students were given open access to GPT-4 as a practice tool, their performance on subsequent unaided exams dropped significantly. Recent large-scale research on secondary students in China echoes these findings; unconstrained AI artificially inflated homework scores, and ultimately led to a decline in exam performance. It’s also shows up in less formal reports, like this one where a professor saw take-home midterm scores artificially skyrocket, only for student performance to collapse when forced to take the in-person final.

I’ve seen enough here. It is absolutely clear to me that having students use chatbot-style LLMs in their vanilla/out-of-the-box capacity and structure is folly, at least if your goal is to increase or improve student learning. I’m never clear on when exactly a particular research finding becomes a product of a corpus of literature, but I definitely think we are at the post-corpus state as far as this claim goes. No more surprise on this one. Your finding is that students use vanilla chatbots poorly? Completely in line with my expectations at this point. Giving a student a vanilla chatbot for learning is a bit like giving a small child a car and being surprised that they don’t know how to use that either.

Is there any place for LLMs in education?

If out-of-the-box chatbots are a dead end for learning, where do we go from here? For all of the above, I am not convinced that there is no place for using LLMs in the educational process. And that’s really what I’m interested in these days as far as the research goes. I want to see where (if anywhere) there are uses for LLMs to help students learn. And as non-useful (or even anti-useful) as a vanilla-style chatbot is, I increasingly find myself wanting to learn more about how tailored LLM applications might be used to help students learn.

There are a variety of ways that you can tailor an LLM. The main ones that I am interested in here are things like wrappers. A wrapper is a tool that surrounds (or wraps) the LLM in a series of instructions and structures and interfaces that constrain its use, and the way in which it can interact with the user. If a vanilla chatbot is like handing that child the keys to a car, a wrapper is more like putting them in a vehicle with dual controls, a speed governor, and a set track.

An example of this approach is CK-12’s Flexi, which is explicitly designed around structured math and science tutoring rather than open-ended chatting. Or this example from Harvard’s computer science department, showing how an AI tailored to guide rather than answer directly supported student learning in software engineering. Rather than sitting down with a chatbot that invites you to do pretty much anything you want (and thereby encourages many students to do the laziest thing they can), a wrapped tool like these is only able to engage in particular modes of use with a student. If those modes of use are well conceived by the people who make the tool, then the likelihood of any particular student using the LLM poorly is greatly reduced, if not altogether eliminated.

Another way in which you might constrain an LLM is to pre-prompt it. This is basically what wrappers do; however, there’s nothing that really stops you from sitting down with your particular vanilla chatbot of choice and entering a lengthy pre-prompt that explains how you want your session to go. This might be Research on tools like This might be as simple as embedding reflective strategies into prompts, or training LLMs on things like student-tutor dialogue. And that’s not even considering approaches where students might be trained in how to use a vanilla chatbot well before they sit down with it.

Admittedly, I have a little bit of personal interest in this given that I am a teacher, a parent of school-age children, and the developer of a few different educational wrapper tools. But I’m not trying to stack the deck here. I don’t think I have a real stake if the research around these kinds of approaches to using LLMs in education shows a net positive, neutral, or negative impact. I just want to know with something like the level of confidence I have around the use of vanilla chatbots. And to this point, I don’t see the same level of research and discourse given to these kinds of approaches that I see being given to some version of “these kids used an unstructured chatbot to study, and it hurt their learning.” The latter kind of finding is no longer surprising; it’s settled. The former is where I hope more research is going soon.

Also, while I’m here, I have to wonder why schools are still providing students with institutional access to vanilla-style chatbots, at least if those schools are going to simultaneously spend a lot of sound and fury regarding the fact that students “just won’t learn these days.” Instead of framing this as a debate about whether to allow AI in schools, educators should really start asking what kind of AI access they are providing and under what constraints. More of the “LLM with constraint” version of this research will help everyone better understand what (if any) kinds of approaches are actually useful for these tools in school.

Thoughts on this one? Am I missing something? What is your experience with LLMs in education? Let me know! Drop a line, find me in the various places, or leave a comment / hit reply.  Thanks for reading! If you find value here, please consider sharing with a friend or supporting my work to help keep the digital lights on.