Should philosophers use AI? A blunt objection and the replies it drew

An argument that has been drifting through the philosophical profession for some time was pulled sharply to the centre yesterday.

Daniel Kaufman, a philosopher at Missouri State University, has written a wide-ranging objection to his colleagues' use of large language models in their research. The piece appeared at Leiter Reports, the most widely read blog in the field, and drew more than ten replies on the day it went up.

What makes the exchange worth attention is not either side but the question itself: in philosophy, can method be separated from result?

Kaufman's four objections#

1. There is no such thing as a pile of truths#

The implicit assumption of the advocates of artificial intelligence, on Kaufman's account, runs like this. The aim of philosophical inquiry is to accumulate important truths; and if that is so, what matters is where one arrives, not how one got there.

Hence the familiar analogy: does it matter who or what found the cure for cancer, so long as the cure works?

Kaufman rejects the comparison. There is not, he says, a single significant question in philosophy on which serious disagreement has not persisted for thousands of years:

Moral realism against anti-realism; deontology against consequentialism; free will against determinism; internalist against externalist epistemologies; foundationalist against coherentist epistemologies; correspondence, coherence and deflationary theories of truth…

What is more, theories pronounced “decisively refuted” keep coming back. Kaufman's example comes from philosophy of mind: functionalism was for years taken to be out of favour, and the excitement over artificial intelligence has brought it back to life.

2. Philosophy's subject matter is human#

Beauty, value, obligation, virtue. Every one of these notions, on Kaufman's view, presupposes human feelings, perceptions and responses.

His question is direct: how is a piece of statistical aggregation software going to help us understand something it cannot feel, cannot see and cannot care about?

3. Philosophy is closer to literature#

This is Kaufman's most contentious thesis. Philosophy, he argues, is not the “hard science of the a priori” that many would like it to be; it is nearer to literature.

A work of philosophy is a disinterested examination of its subject, certainly — but it is also the point at which the philosopher's own voice, history and way of looking meet that subject. The consequence is that where an idea comes from, and how it is put, are decisive in philosophy. Against the promise of a kind of “immortality” through artificial intelligence, the only form of permanence Kaufman offers is this: that distinctive voices go on being heard.

4. Philosophy is a social activity#

Kaufman argues that the enthusiasts overlook philosophy's social dimension. One writer had suggested there was no difference between using artificial intelligence and working with graduate assistants; a reader wrote in to say he was glad to see the machines take over the “stupid drudgery”.

Kaufman's answer has two layers. The first is practical: that drudgery is one of the few sources of income graduate students have. The second is a matter of principle: talking to colleagues is not something outside philosophy, it is philosophy itself. To replace the Socratic exchange with sitting alone in a room with Claude or ChatGPT is to forget what the discipline is.

The replies#

The discussion beneath the piece is as interesting as the objection above it.

David Wallace (Pittsburgh) defended the claim of progress. Kaufman, he suggests, is falling into a selection effect: among the questions counted as philosophical over the centuries were whether matter is composed of atoms, what the nature of motion is, and whether the universe had a beginning. On those questions irreversible progress was made — so thoroughly that they now belong to their own sciences rather than to philosophy. Wallace sees a gain on the human side as well: could the rejection of the claim that slavery is good not be entered on the pile of truths?

Rob Precht offered a correction of the history: functionalism was never a dead issue. The work of Shoemaker, Levin, Polger, Wilson and Piccinini was going on long before large language models, and Chalmers's The Conscious Mind dates from 1996. The story of a doctrine revived by AI enthusiasts, he argues, is chronologically wrong.

A. B. Jimenez-Cordero accepted the objection and reversed its conclusion. Philosophy may not accumulate truths; but it does accumulate results. We know which argument fails and why, and which theory runs aground on which objection: the trouble behaviourism met in Putnam's super-Spartans, or the trouble type-identity theory met in the possibility of a thinking computer. On this reading philosophy produces not settled truths but a growing grasp of the things we do not know. And since the work is largely one of “combinatorial mapping”, one should expect artificial intelligence to be good at it.

The question Jimenez-Cordero leaves behind is this: artificial intelligence may well automate a large part of how analytic philosophy currently operates — but was that ever the point of philosophy?

Michel Xhignesse entered two notes. Beauty appears in Kaufman's list of human values beyond the reach of machines; yet aesthetics is a field in retreat from graduate teaching in the United States. And second: the labour of research assistants already goes largely unacknowledged, though it is both real work and a training.

Galen Strawson agreed with Kaufman but with a reservation: being able to bring dead horses back to life is not a virtue in a discipline. He recalled Schopenhauer's remark that truth is granted only a brief celebration of victory between the long stretch in which it is condemned as paradoxical and the long stretch in which it is dismissed as trivial.

What the argument is really about#

On the surface this is an argument about technology. Underneath it lies a classical division about what sort of activity philosophy is.

Make philosophy resemble science and method becomes instrumental: if the result is true, how one reached it is secondary. Make it resemble one of the humanities and method becomes part of the result: how a thought is constructed determines what that thought is.

The division is not new. As we have reported before, Bernard Williams argued years ago that philosophy is closer to history than to science. Kaufman's objection puts that old thesis to the test against a new technology.

Another side of the argument was covered in these pages a few weeks ago: a philosophy journal knowingly publishing an article written largely by artificial intelligence was the same question in concrete form.

A note#

There is an assumption that never appears by name in Kaufman's text but quietly governs the whole dispute: that a philosophical text has an author.

The use of artificial intelligence does not attack that assumption directly; but it does open the question of how much of authorship can be handed over to an instrument. If the footnotes, the bibliography, the map of the argument and the first draft can all be delegated, what exactly is the “distinctive voice” that remains?

The argument continues. It is fair to say that whichever side has its answer ready has probably not taken the question seriously enough.