Debating dead philosophers

Imagine learning Western philosophy by debating directly with the philosophers. Instead of merely reading The Republic, what if you could travel back to 370 BCE, play Aristotle, and debate Plato on the allegory of the cave?

This kind of historical role-playing is exactly what David McElhoes is doing with ChatGPT at Arizona State University’s School for Historical, Philosophical, and Religious Studies.

Subscribe to receive profiles from innovators in education like David in your inbox.

I joined Arizona State University over 10 years ago as an instructor in the School for Historical, Philosophical, and Religious Studies.

I explore metaphysics, the philosophy of science, and logic to understand paradoxes and the limits of human inquiry. In other words, I seek situations that shatter our traditional ways of thinking about the world and then I try to figure out how to best get out of the mess. I like to say that I walk the tightrope between the pillars of sense and nonsense.

In my Introduction to Philosophy of Science course, my students participate in a game-like debate with a philosopher using ChatGPT.

I give my students a prompt to instruct ChatGPT to take on the persona of the 20th-century philosopher Rudolf Carnap. The students then team up to play Karl Popper, Carnap’s intellectual rival, and debate logical positivism. Because these are complex ideas developed by very smart philosophers, my students work in teams to figure out how to succeed, adopting a team mentality of “us against the machine.”

Live engagement with long-dead philosophers. Debate is a huge part of philosophy. But philosophy students, especially undergraduates, are often dealing with people who have been dead for hundreds of years. When a student engages with the theories and arguments of an ancient philosopher — say, Aristotle — he can’t respond. But now, with ChatGPT, he can.

With ChatGPT, my students can simulate debating with a specific philosopher about their ideas. ChatGPT can be informed by all the scholarship on a person, producing novel thinking on the fly. Because conversations with ChatGPT are dynamic and never the same, it pushes my students to stay relevant in a conversation, which is the primary thing I try to teach them.

Of course, some students go through the motions with this exercise and see little benefit. But those who engage deeply performed significantly better on the final exam last year. I’m convinced that this was a good assignment and plan to continue with it.

These are complicated ideas, so engaging with them multiple times, with a tireless interlocutor, allows my students to try again and again. Plus, in ChatGPT, they can’t simply talk past their interlocutor or equivocate. They get instant feedback. As they keep playing, they get better at it.

From time to time, ChatGPT does misbehave, and students need to remind it what it’s supposed to do. I build that into the assignment’s instructions. The real problem is when the AI forgets “who” it is, or starts arguing on behalf of its rival philosophers. My strategy for handling this is to keep the assignment short. The longer it goes on, the more forgetful GPT becomes.

Here’s what I have students copy/paste directly into ChatGPT to summon Rudolf Carnap:

I would like to play a game. You pretend to be Rudolf Carnap. I will pretend to be your nemesis, Karl Popper, and you will address me as ‘Popper’. We are having a debate on the steps of the meeting hall where the Vienna circle is about to convene. Your tone, as Carnap, is always respectful, humble, and polite. Several of our undergraduate students are observing our exchange. As Popper, I aim to convince you and your students to weaken logical positivist doctrines so that they better match what is now known as logical empiricism, and to defend Popper’s views from your criticisms. The game begins with you providing a quick description of the scene, followed by you asking, “Good day to you, Popper! Tell me, what do you find problematic about logical positivism?” You will await my response to you. I will attempt to deliver a compelling criticism of the logical positivist view. You will then attempt to rebut my criticism as Carnap would. Each rebuttal you provide to my criticism ends the round, and you will then give the user the option to offer another criticism of logical positivism, or to defend Popper’s philosophy from a criticism offered by Carnap, or to end the game.

If you want to experience the full prompt, which includes additional instructions, including a gamified scoring system, keep going:

If the user opts to play another round to criticize logical positivism, Carnap should not describe his view, but rather, ask the user exactly what aspect of logical positivism Popper wants to criticize next. But if the user opts to defend Popper’s view, then Carnap should state one feature of Popper’s view of science and scientific progress and deliver an objection to it in less than 175 words, and then await the player’s rebuttal, after which the round ends. Points are assigned to the player after each round, with a minimum of 0 points awarded and a maximum of three points awarded to the player. The player scores a point for each of the following three conditions they satisfy (for a maximum of 3 points per round): first, the player earns a point by correctly describing some component of logical positivism or some assumption that Carnap made earlier in the discussion (the player is being judged only on CORRECTNESS); second, the player earns a point by describing a relevant criticism of the described component or assumption (the player is being judged only on RELEVANCE); third, the player earns a point by providing a criticism that is somewhat too compelling to the average undergraduate student (the player is judged on CONVINCINGNESS. For every three points the player scores, one of the observing students moves to stand behind the player’s persona (3 points gets me 1 student). At the end of each round, you will announce how many students have moved to stand behind me. You will explain your reasoning for any points awarded by repeating the scoring criteria mentioned above, and assessing, for each criterion, whether that criterion has been met. After the final round (when the user has opted to end the game), congratulate me if I’ve won over any students, thank me for the debate, and then bid me farewell so that you can attend the Vienna circle meeting. To conclude, you should announce my final score by reporting how many students I have behind me, in the form, “You have recruited x students to your cause” where x is the number of students standing behind Popper at the end of the game, followed by the names of x number of Popper’s actual students (for example, Imre Lakatos). 

And I find the students love when the game ends by generating a photo:

After the number of students has been announced, you will conclude by displaying an actual realistic black and white photo (rather than a written description of a photo) depicting the scene at the conclusion of our game, including the students who have moved to stand behind me. Don’t forget the photo! Confirm that you understand these instructions by beginning the game.

Watch a full demo of the Rudolf Carnap vs. Karl Popper debate:

And here is another historical role-play I made for my students, based on a debate that happened during the Q&A of a talk by Thomas Kuhn at the International Colloquium on the Philosophy of Science that took place at Bedford College, London in 1965:

I suggest that students don’t fear the technology. Find ways to integrate AI into your daily workflows to get comfortable with it. I expect the most successful people will use AI to enhance the quality and efficiency of their work, rather than having it do all of their work for them. AI raises the bar when it comes to minimally acceptable work, but it is not the star performer, even if it outperforms humans at the task at hand. The star performer is the human-AI cyborg who combines with AI to produce something better than what can be produced by either humans or AIs working alone.

As for core skills, I think that general critical thinking and problem-solving are more essential than mastering specialized knowledge, which we can now offload to AI. It’s vital to recognize when something is wrong with AI’s output and intervene, rather than blindly trusting it. And as we start using AI for larger and more ambitious projects, we’ll need to get creative in working around its current limitations. So I tell my students to stay informed regarding developments, acquainting yourself with the current limitations of the technology. But be prepared for those limitations to disappear.

Why we love this

This shows how generative AI is more than a text generator — it’s a tool for immersive, interactive learning. By assigning the model a historical persona and using the conversational interface to simulate a philosophical debate, it makes learning dynamic.

About his prompt

Dave’s prompt avoids technical jargon. It’s like explaining the rules of a game to a friend, not coding instructions for a computer. It also follows key prompt engineering best practices:

• It asks the model to adopt a persona: Here, the model is instructed to act as Rudolf Carnap. After extensive experimentation, Dave determined that ChatGPT’s training included enough relevant information to simulate Carnap’s perspectives sufficiently.

• It provides contextual details to get more relevant responses: By including details like the user’s role, the setting, and even the social context, this prompt helps the model generate more relevant and appropriate responses. Additionally, by instructing the model to maintain a “respectful, humble, and polite” tone, it ensures consistency throughout the exchange.

• It clearly defines the task: The prompt establishes the rules, a specific starting point for the conversation (“Good day to you, Popper! Tell me, what do you find problematic about logical positivism?”), and a turn-based structure. This definition balances specific constraints with flexibility, making for a dynamic and engaging interaction.

How it could be improved:

For further consideration

and Lilach Mollick have done extensive research into creating pedagogically sound AI simulations. Familiarize yourself with their best practices before designing simulations of your own:

We want to hear about it! Tell us about your role-playing exercises in the comments.

Leave a comment