Offloading to AI Means Not Doing Philosophy: on the importance of exploring the space of ideas

When I write philosophy, my task is not merely to generate the most persuasive argument for some target claim. Rather, my task is to figure out what I ought to say. I often have the experience of setting out to write a paper with a conclusion in view, only to find that spelling out the argument in detail leaves me not believing the conclusion anymore. The arguments instead lead elsewhere.

I expect this experience is familiar. It reveals that writing philosophy is (or at least can be) more a matter of figuring out what to think than it is of figuring out how to express pre-existing ideas. I start a paper with a conclusion in mind, but I end up making discoveries through the process of writing. Careful writing is a way of exploring the space of ideas and settling my own mind about them.

This is not just true for me doing original research intended for publication. It holds also for students asked to think through philosophical questions, even in introductory courses. When a student says that things are perfectly clear in their head but they can’t get it out in writing, they often don’t really understand (at least, not yet). Clarity is illusory if it can’t be made explicit. And the best ideas often start out unsettled and indeterminate. Writing is a way of doing philosophy because it centrally involves articulating and refining ideas.

If someone writes an initial prompt and lets an LLM fill in the details, then they avoid doing the work of figuring out what they really think. One might object that this is no problem, provided the user ends up believing the truth. But that does not answer the concern. As John Stuart Mill argues, merely remembering true claims without understanding them or knowing what justifies them risks leaving the empty shell of mere verbalisms. This is especially true in philosophy, where superficially persuasive, logically valid arguments can be given for almost any position—and for its negation.

Still, one might object that not all philosophical writing aims to discover previously unknown answers. Indeed, there are times when I am not making original arguments, but instead am summarizing an issue or explaining what some other philosopher has argued. Students, especially in introductory classes, are often asked to do that sort of thing. So why aren’t LLMs a useful tool for that?

Before answering that question, note that if it were fine to rely on an LLM to summarize philosophical texts, then it would also be fine to turn to secondary literature written by humans. There is no reason to think that human-written summaries would be any worse. So, this proposed use of LLMs raises the broader question: Why should students read primary texts?

There are many possible answers to that broader question. An important one for me is that, when I assign primary texts in my courses, I do not know exactly what the value of them is. If I could provide students with the value of reading Descartes’ Meditations, paraphrased as a set of discrete insights or conclusions, then I could assign something shorter and easier to read which concisely summarizes it. But it’s more complicated—and more interesting—than that.

When I assign a primary text, I read it ahead of the class meeting where we discuss them. I may have read it a dozen times, but I typically notice things when reading it this time that I either had not noticed or had not thought to emphasize. Of course, my understanding of the Meditations last semester was not totally different than my understanding the time before, but neither was it strictly an incremental improvement. I was presenting how it seemed to me this time.

I assign texts which I think are worth engaging with, but I do not expect students to come away with the same interpretation that I do. I often have some specific point in mind when I include a text on the syllabus, of course, but there are indefinitely many take-aways which would make the reading worth students’ time. My task is to help students find some meaning in the assigned text and learn how to find meanings in texts like it. Just as writing towards a conclusion can lead through the space of ideas to something unexpected and different, reading and rereading a primary text uncovers unexpected details and yields novel insights. Reading a summary would rob students of this experience.

When I made this argument to a colleague of mine, they protested that they were just teaching recent articles—nothing with the enduring greatness of the Meditations. However, my point is not just about a canon of books written long ago by a short list of greats. I feel the same when I have students read Thomas Kuhn’s The Structure of Scientific Revolutions. There is a lot in that book, and I expect them to walk away with their own attitude about the details of its insights. Yet one might still object that Structure is more than half a century old, so it has had time to join the canon.

Perhaps. But a professor teaching similar topics in the 1960s could have assigned Kuhn’s book, and students would have had much to gain from it even then. The issue is not the purported staying power of a text, but the value of putting in the effort to struggle with it on its own terms. Even if most of the recent articles my colleague assigns in a contemporary topics course are forgotten or superseded within the next few decades, we do not know yet which those will be. And the only way to find out will be to engage with them directly, not to let an algorithm distill them down to a summary.

I do not know if LLMs (or AI more generally) will be a valuable tool in philosophy teaching of the future. Perhaps they will, although I suspect that engaging responsibly with LLMs will require skills we do not have yet. Important philosophical questions do not admit of clearcut, universally compelling answers. So, there is an important element of philosophical writing and reading which aims not just to arrive at some defensible position, but to find one’s own position. Deferring to the output of an LLM, the algorithmic average of philosophical discourse, would be a kind of bad faith—a refusal to genuinely do philosophy.