Critical Reasoning with AI: Using public discourse to teach political science

The teaching superpower of LLMs is assisting and modeling reasoning / analysis, not serving content

I want to show the sort of thing I see too few people talking about when it comes to AI and teaching. And although this is perhaps a lengthy way to demonstrate, I think the best way is to walk you through the process I followed, complete with verbose output from Claude, the LLM I've been using lately.

As you follow the examples, feel free to skim the Claude output and get the gist. I don't want to abbreviate it though because I think the depth provided is important.

Anyway, today's journey starts with this small tweet exchange that I saw on X (formerly Twitter).

One of the things I've been doing is playing around with Toulminian analysis via LLMs. That string of words might mean nothing to you but all you need to know for this example is that “warrants” in Toulminian analysis are the underlying assumptions and values that connect evidence to claims. So for example, if I say look at this, a person got 10 years prison for stealing a car and another got 6 months for bilking a million dollars out of people's retirement funds, one necessary warrant there is that punishment should be proportional to harm caused. I don't have to say it, but without that my claim that this shows injustice doesn't make sense.

Anyway, I decided to ask Claude the following:

take the three participants different warrants. Can the different warrants be categorized according to underlying philosophy, and can those philosophies be matched to specific philosophers, economists, or political theorists?

So take a second and look at those three tweets, and then see what that produced — and imagine how you could use this in a class activity:

Let me analyze the three different warrant systems and connect them to broader philosophical traditions:

1. Qasim Rashid's Warrant:

- Core belief: Government policy should prioritize the many over the few, especially regarding economic relief

- Philosophical alignment: Utilitarian and Social Democratic traditions

- John Stuart Mill's greatest good for greatest number

- John Rawls' theory of justice, particularly the difference principle

- Elizabeth Anderson's democratic equality

- Economic theorists:

- Thomas Piketty on wealth concentration

- Joseph Stiglitz on inequality

- John Maynard Keynes on wealth redistribution

2. @MACRoofing's Warrant:

- Core belief: Economic rewards should be based on merit, effort, and value creation

- Philosophical alignment: Classical Liberal and Meritocratic traditions

- John Locke's labor theory of property

- Friedrich Hayek's defense of market outcomes

- Ayn Rand's objectivism and emphasis on productive achievement

- Economic theorists:

- Milton Friedman on free market capitalism

- Ludwig von Mises on entrepreneurship

- Joseph Schumpeter on creative destruction

3. @ThinkSarcasm's Warrant:

- Core belief: Voluntary agreements are more morally binding than involuntary obligations

- Philosophical alignment: Libertarian and Voluntarist traditions

- Robert Nozick's minimal state theory

- Murray Rothbard's anarcho-capitalism

- Herbert Spencer's social Darwinism

- Economic theorists:

- James Buchanan on public choice theory

- Friedrich Hayek (again) on spontaneous order

- Hans-Hermann Hoppe on private property rights

Interesting patterns emerge:

1. Historical Evolution:

- Rashid's position descends from post-WW2 social democratic thought

- MACRoofing's from 18th-19th century classical liberalism

- ThinkSarcasm's from 20th century libertarianism

2. View of State's Role:

- Rashid: Active force for equality and justice

- MACRoofing: Referee ensuring fair competition

- ThinkSarcasm: Potential violator of individual rights

3. Different Conceptions of Justice:

- Rashid: Distributive justice (outcomes)

- MACRoofing: Commutative justice (exchange)

- ThinkSarcasm: Procedural justice (process)

This analysis reveals how contemporary political debates often replay centuries-old philosophical disagreements about the nature of justice, property, and the role of the state. The student debt/tax policy debate becomes a proxy for these deeper ideological divisions.

Would you like me to expand on any of these philosophical traditions or their modern political manifestations?

Ok — so there's a number of ways you could use this. First, this might just be for you, as a teacher, as starting notes about facilitating a conversation about these tweets. Do you know all this stuff if you are teaching political science? Of course. Frankly, you know a lot more. Maybe you look at this and there's a couple pieces that you feel aren't quite right, or are outright misunderstandings. But would all of this have sprung to mind at this depth without using Claude? I think probably not.

Again, LLMs are very very good at showing you the things that you didn't remember to think about. Its a really useful capability, far more useful than showing you specific facts you ask for, which Google already does quite well.

I also could see using Claude in a front of class facilitated discussion. You plug in this prompt and get the results on the screen, or have students do it in groups, and then ask the students — why do you think Claude links this to John Rawls, and this to John Locke? What do you think of the point that these debates are old, old debates?

This is just one example of how you might use LLMs for teaching, but it is the beginning of a point I will be making over many posts and examples. Everyone is thinking that LLMs will either be tutors or content delivery systems. But the thing the newer versions released over the last year or so do really well is model reasoning. As here, you can take an observed phenomenon and have the LLM model how an expert might talk about it.

LLMs do that incredibly well, both because they mimic discourse as a core competency, but also because after the data gold rush ended, the hundreds of millions of dollars flowing into these systems was focused on building better reasoning and inference capabilities, and that's what you see on display here. And it's this capability — modeling reasoning — into which education needs to lean I think the hardest, not really content production, and not even tutoring as normally envisioned. The ability to have students see, for any given prompt, what disciplinary thinking looks like applied to it is an incredible superpower, and we’d be fools to leave that power untapped.