Philosophy teachers—Michael Rota, a professor of philosophy at the University of St. Thomas (Minnesota), is about to make your teaching a bit better and your life a bit easier.

Professor Rota recently began learning about how to use artificial intelligence tools to teach philosophy. In the following guest post, he not only shares some suggestions, but also let’s you try out two demos of his GPT-4-based interactive course tutor.

The course tutor is part of a program he is helping develop, and which should be available for other professors to use and customize sometime this summer.

–

Using Generative AI to Teach Philosophy
by Michael Rota

I have a friend who leads AI product strategy at a medium-sized tech company, and for about a year he’s been telling me about various impressive tasks one can accomplish with Large Language Models, like OpenAI’s GPT-4. In December I finally started listening, and began investigating how one might use AI tools as a teacher. (I’m a philosophy professor at the University of St. Thomas.) I’ve been amazed by the promise this new technology holds for instructors—in part because of the potential to increase productivity (of the teacher), but even more because of the potential to improve student learning.

In this post I’ll focus on the practical and discuss three free or low-cost tools that can be employed by a philosophy professor without any special technical expertise: (1) an interactive course tutor for your students, which you can load with your own questions and answers from your course, (2) a tool for quickly drafting a new exam, quiz, or HW assignment, and (3) a chatbot created from your own syllabus and lecture notes, so your students can query the content of your course.

The interactive course tutor

GPT-4 mimics human reasoning remarkably well (it scored in the 88th percentile on the LSAT). But it sometimes just makes stuff up. What if you could provide GPT-4 with good answers to questions you wanted your students to work through? It turns out you can, and thus it is possible to create an alarmingly capable course tutor by supplying GPT-4 with a series of question/answer pairs. This allows each student to have a one-on-one tutoring experience, and get immediate feedback as they work through an assignment.

You can play with a demo of this here.

Take the first assignment in the first module of this demo: “Think up a false conjunctive proposition.” This task has an infinite number of possible correct responses, yet GPT-4 can competently assess student answers,because the instructor-provided answer passed to GPT-4 by the program is general—it’s a recipe for correct answers, as it were. In this demo, the instructor-provided answer GPT-4 has been given is this:

A conjunctive proposition is any proposition of the form A and B, where A is a complete proposition and B is a complete proposition. A and B are called the ‘conjuncts’ of the conjunctive proposition. A conjunctive proposition is false if and only if A is false or B is false or both A and B are false. It counts as true otherwise.

That’s it. That’s enough for the AI tutor to respond accurately to almost any possible student response. A student can get the question wrong in a number of ways: for example, by entering a conjunctive proposition that’s true, or a proposition that’s not a conjunction, or something that’s not a proposition at all. GPT-4 handles all of these possibilities.

Using generative AI in this way offers several advantages over traditional homework assignments:

(a) students get immediate, specific feedback on each question
(b) students who need more practice can get it without having to make other students do busy work
(c) there’s less grading for teachers
(d) there is a decreased need for the teacher to explain the same thing multiple times.

How will grading work? In my view it’s too soon to hand grading over to AIs, so in my classes I plan to split the grading and the learning. The grading will be based on class participation and in-class, pen and paper exams. The learning will be facilitated in the standard ways but also with the help of an interactive course tutor based on questions and answers from my course.

Here is a second demo, allowing an instructor to test functionality by inputting a single question/answer pair and then checking how well the AI tutor handles mock student answers.

The demos linked above use an early version of the product I’m helping to design. It should be available by the summer, at which points professors will be able to create an account, input their own modules of question/answer pairs, and hit ‘submit’ to create a tutor based on their material, accessible for their students as a Web App.

For a larger discussion of the promise of interactive tutors in education, see this TED talk by Sal Khan of Khan Academy.

Assignment generation

The creation of new questions for homeworks, quizzes, and exams can be time-consuming, whether one is designing a new course or just creating a new version of an existing assignment or test. Large language models are great for speeding up this process.

If you go to chat.openai.com, you can sign up for a free account with OpenAI and use GPT 3.5 at no cost. That allows you to type into a textbox, entering a prompt like “Can you give me ten sample questions on ____, suitable for a college level” or “Here’s a question on this topic {insert a question from an old assignment}. Can you can give me a very similar question, but with different details?” Used in this way, GPT 3.5 can provide some value.

But GPT 4 is much better, both because it is better at mimicking human reasoning and because it allows you to attach files. So you can attach an old assignment and ask it for a very similar assignment in the same format. The downside here is that to use GPT-4 you need a GPT Plus account, which costs $20 a month. An upside is that additional functionality comes along with a GPT Plus account: you can access the GPT store. There you will find customized versions of GPT-4 like the “Practice Exam/Quiz/Test Creator for School” GPT, which allows you to upload course content (e.g. your lesson plans on a topic), and then ask for sample questions based on that material. With only a little more work, you can create your own GPT with access to knowledge about your own course, uploading up to 20 files, and use it to generate drafts of assignments tailored to your material.

As with any AI generated content, think of the process like this: with the right prompting from you, the AI produces initial drafts almost instantaneously, but you’ll need to evaluate and refine before you have a final product.

Course chatbot

Another thing you can do with GPT Plus is to create a course chatbot. If you create a GPT and upload files with information about the content in your course (your syllabus, lessons plans, handouts, etc.), then anyone with access to the chatbot can ask questions like “When is the final exam in this class?”, “How did we define a ‘level of confidence’?”, or “what are the steps in Aristotle’s function argument?”. And you can give others access to the chatbot by making it available to anyone with a link. However, your students would need a GPT-Plus account to use it, and that may not be feasible. But there is a free workaround: If you put your course content in a pdf that is no more than 120 pages (or break it up into several), you can give the pdf(s) to your students and direct them to the landing page of ChatPDF, where they can upload the pdf and then query it for free.

If you have further questions about any of this, raise them in the comments or email them to me.

The post Using Generative AI to Teach Philosophy (w/ an interactive demo you can try) (guest post) first appeared on Daily Nous.