4.2: AI for tutoring-style assistance

Before you ask AI for assistance, try asking yourself “Could a human tutor help me in this way without doing my work for me?” Most of us recognize that human tutors can do all kinds of things to stimulate our learning, and they also have to draw a line between helping and overhelping. If we use that line as a standard, we can be pretty sure AI will not interfere with our learning.

I do want to make clear that I’m not saying AI assistance replaces human tutoring or is better than human tutoring. Human tutoring offers actual human connection, empathy, encouragement, witnessing and other forms of support that have been shown to have a huge impact on learning. Your human tutor might actually be interested in what you have to say and might care about you as a person and a learner; AI can say nice things, but it isn’t capable of those feelings or relationships. 

That said we don't always have access to human tutors. And not every human tutor is a good fit for every student. They might not be as flexible or familiar with particular approaches to learning assistance that you want to try. These are reasons why I see AI tutoring-style assistance as a valuable supplement. In my writing classes, I assign students to visit a human tutor first and later reflect on AI feedback*.

Get the chatbot to explain differently

Sometimes we turn to tutors when we just don’t get it, when the course materials don’t make sense. The tutor explains in a different way and keeps trying different explanations until we get it. We can ask a chatbot for that. And with a chatbot, we can get ultra specific about what flavor of explanation we want. 

For example, we might ask for an explanation that fits our intellectual strengths or how smart we’re feeling that day:

• “Explain it for a five-year-old/a twelve-year-old/a graduate student.”
• “Explain it for a physics major who doesn’t get poetry,”
• “Explain it for an English major who doesn’t get math.”

We can also ask for engaging explanations when we’re having trouble related to the material:

• Why it matters: Ask it to help you understand the relevance or application of the concept to help you stay motivated to learn. or example, if you were studying the theory of mind in preschool-aged children, in addition to asking for explanations of what that is, you could ask “Why does it matter if a kid has a theory of mind or not?” 
• Style or genre: Ask it to explain it in a creative way that’s more fun or that reaches you differently, like a podcast (try NotebookLM), a spoken word poem, a rap, a sonnet, an image, a short story, or a game (This would be tough for a human tutor to produce on the spot.) 
• Examples: Ask it to tailor examples to things you care about or are interested in. Maybe you are learning about opportunity cost in an economics class and you ask for an example related to sustainable fashion. 
• Comparisons or metaphors: For a concept that doesn’t have easy-to-relate-to examples, you can ask for a comparison or metaphor instead. Let’s say you are learning about covalent bonding in chemistry and you ask the chatbot to make up a metaphorical explanation involving dogs while still conveying the essential concepts. (Here’s a ChatGPT answer involving dogs sharing bones instead of atoms sharing electrons.)

Get the chatbot to ask you questions and help you practice 

You can also use a chatbot to test and expand your understanding of the course material, such as before a test, before class discussion, or before a meeting with an instructor. Tell it to make up sample questions based on your topic and course materials. If you’re preparing for a test, give the chatbot any information the instructor has given you about the test.

For example, here’s a sample chat where I asked it to ask me questions about logical fallacies based on a textbook chapter (the chapter was published online under an open license, so I wasn’t violating any rights by uploading it to the chatbot).

If you have to do an oral exam or you need to interview someone for an assignment, you can ask a chatbot to simulate the exam or interview to help you prepare. The AI Pedagogy Project at Harvard’s metaLAB gives an example of how students preparing to interview heritage speakers of Spanish could practice their interviewing skills with a chatbot before the real interview. (Consider using voice mode if you prefer to do this out loud)

Learn by teaching the chatbot

They say if you want to learn something deeply, teach it. What if you pretend to be the teacher while the chatbot pretends to be the student? It explains something to you (maybe imperfectly), and then you tell it how to improve the explanation. In “AI as Learner: Challenging Students to Teach,” Ethan and Lillach Mollick provide a prompt for this. They share a chat session where the student critiques the chatbot’s explanation of “first-mover advantage” in business. 

How do you check AI “tutoring”?

Wait a minute, though. Given all the cautions we’ve heard about bias and inaccuracy in chatbots, we’re always being told to check what comes out of them. But how the person who needs a tutor supposed to know whether the tutor is right? As Leslie Allison says in “AI Can Do Your Homework, Now What?” “The less you know about something, the more likely you are to be convinced by ChatGPT's answer.”

There’s no perfect solution to this challenge. It might help to think of a crazy cousin who knows a lot but gets mixed up. They’ll often give accurate explanations and examples, but sometimes they will sound great and be wrong in weird ways. Would you still turn to them for tutoring assistance? How would you know what to trust?

The good news is that there are some ways you can find out if it’s wrong. It takes a little time, but there are faster ways to do it, and the time is not wasted; it’s probably helping you learn the same concepts you were trying to get help with. 

Ways to investigate accuracy

Here are some strategies to help you assess AI tutoring-style assistance.

• Does it make sense to you? Trust your own thinking and investigate when something doesn’t seem quite right.
• Does it match what’s in your course materials? Circle back to readings, notes, and textbook explanations. You can ask the chatbot to double check against your course materials if those are publicly available and quote from them to show how they support its explanation or example.
• Let the chatbot critique itself. You can ask AI to play the role of an expert and double-check itself. For example, when I asked ChatGPT to explain covalent bonding, it compared atoms sharing electrons to dogs sharing bones. I asked, “How might an expert chemist critique this explanation? In what way is it misleading?” 
• Check against credible web sources (with chatbot help). Search, or ask the chatbot to suggest a link to a credible website that supports its claim or explanation. 
• What does a teacher, human tutor, or classmate say? Keep a human in the loop. AI assistance and human discussion can complement each other.

A personal example

Here’s an example of a way I used it in my own learning. “Embedding” is a term for something key in the structure of large language models. I’ve heard “embedding” for years, seen it explained, but not really understood. Recently, I read it somewhere and got frustrated. I knew it has something to do with the math that language models use to represent and analyze text. But what was embedded in what? I felt intimidated and didn’t want to invest the time to absorb a technical article. Besides, I was skeptical that such an article would explain it in the simplest way possible.

So I asked ChatGPT to explain embeddings in five ways that a twelve-year-old could understand. The results appealed to me and reminded me that I had read that embeddings were vectors, ordered sets of numbers that represented words. Here are two

•  “Imagine every word took a huge personality quiz with hundreds of questions like “Are you positive or negative?”, “Do you go with animals?”, or “Are you something people can touch?” The answers to all those questions are stored as a list of numbers—that’s the word’s embedding.” 

• “Embeddings are like a secret number code that the computer uses to remember what a word means instead of how it's spelled. So instead of ‘dog,’ the computer sees something like [0.2, -0.1, 0.7, …]. That code helps the computer figure out how “dog” relates to ‘puppy’ or ‘cat.’”

Then, when I was feeling a little more confident,  I asked it for an explanation aimed at an educated layperson with no background in computer science. The explanation made some sense. It ended with the evocative lines, “Embeddings are how a language model turns words into math so it can work with meaning. Think of them as coordinates on a map of ideas that help the model "navigate" language.” ’

Next, I asked it how a computer scientist would critique its explanation. While I didn’t understand everything in the critique, I did appreciate some clarifications that made sense given credible pieces I had read on LLMs. For example, it explained that the same word could have a different embedding, or sequence of numbers, depending on context, which made sense because a word can mean different things depending on context.

At the end of this chat session, I didn’t feel like I had mastered embeddings, but I did feel better equipped to read and respond to human discussions about them. I was intrigued to learn more.