why you should be talking with gpt about philosophy

some thoughts on how it's become better, and how you can too

TLDR:

• The latest version of ChatGPT is a valuable option for engaging in philosophical dialogue

• To get the most from it: treat it as an equal, get it to role-play, and keep on pushing back

• We can’t wrong GPT by how we talk with it, but we might wrong ourselves

---

For days, I’ve been wanting to publish something about the English liberal obsession with ID cards. But I still haven’t finished it. So I’m going to explain why that is, by writing about a different obsession: the hours and hours I’ve spent, over the last week or so, talking about philosophy with ‘o1 pro’, the latest version of ChatGPT.

I’ve talked with Gpt (as I like to call it) about Putnam and Quine on conceptual schemes. I’ve talked with it about Ζeno’s paradoxes. I’ve talked with it about behaviourism, causality, skepticism, supervenience, knowledge, humour, catastrophic moral horror, the container theory of time, and the relation between different conceptions of modes and tropes. I tried to get it to persuade me to become an aesthetic expressivist. I got it to pretend to be P.F. Strawson answering my objections to Freedom and Resentment. I had a long chat with it about the distinction between the good and the right.

Basically, I’ve been talking with Gpt about all the philosophical matters that I happened to want to talk about — when I wasn’t talking about these things with philosophical humans. I think it’s important to add that. I definitely don’t see my conversations with Gpt as replacing the conversations I have with humans! I’ve had some fantastic conversations with humans about philosophy over the past week or so. But when I wasn’t — and I’m clear to fully separate out these things — I was spending a lot of time with Gpt. And my conclusion is that it’s now really good at philosophy.

I’ll use the rest of this piece to unpack that claim, by making the following three points. First, that Gpt has become a valuable option for engaging in philosophical dialogue. Second, that this still depends on how you use it. Third, that I’m bothered by some ethical questions arising.

Now, if you haven’t tried talking with Gpt about philosophy, then you might imagine it’s like talking with a machine that’s downloaded all the textbooks. This is how I found it in the past. Before this year, I’d only had a handful of tries. If I look back on my OpenAI account — to the times before I started spending $$ on o1 pro, so I can talk about Quine whenever I want — all I can find, outside the last month, is a short interaction on Dworkin and animal rights from January 2023.

That feels like a long time ago. In Gpt terms, it’s an eternity. Of course, I was using it badly back then. After all, I still have much learn on that front (although it’s developed so much that now there’s vastly more to do, as well as more to gain). But in 2023, I gave up after a few interactions. It seemed clear that, without proper work from my end, it was limited to providing simplistic summaries of things I’d already read. Maybe this would’ve been useful for someone who wasn’t up to reading the SEP, but I was happy to stick to that and my other more Luddite ways of searching out the truth.

Then, a few months ago, I tried Gpt again, over lunch with a friend, and found it much improved. I had decent backs-and-forths with it about whether the sphericity of a tomato is a property, and about Nagel on the badness of death. But I still wasn’t drawn to chatting with it often. Though admittedly, this was increasingly owing to the same reason I don’t keep Balatro on the home screen of my iPhone. I don’t want to want to spend my time on virtual pursuits!

With Gpt, however, that ship has now long sailed. I gradually started using the free version more — mostly to ask it things like “can you tell me all the places my dad wrote about X?”. This was useful. First, because my dad was a great philosopher who wrote a lot, on a lot of topics, and I haven’t read enough of it. And second, because I was away from almost all my books, spending time in America.

Since the general roll-out of o1, however, on the first day of OpenAI’s Christmas bonanza, I’ve posed fewer clarificatory questions. This is because, now, if you do it right, talking with Gpt is like talking with someone who’s seriously studied and thought hard about philosophy. Gpt could easily get a PhD on any philosophical topic. More than that, I’ve had many philosophical discussions with professional philosophers that were much less philosophical than my recent chats with Gpt.

I should add, of course, that I’ve also had many philosophical discussions with professional philosophers that were much more philosophical than my chats with Gpt! And with people who aren’t professional philosophers. I’ll also acknowledge that, of course, Gpt still makes mistakes. We all know it hallucinates (or ‘confabulates’, as Anil Seth makes a case for renaming it, in this recent paper I enjoyed reading). And there’s no way I’d trust Gpt for important exegetical purposes. But I wouldn’t trust human-written secondary literature for that, either. If you want to be sure what some particular philosopher wrote in some particular text, then go and read it!

Gpt also ties itself up, and contradicts itself. But, again, so do people! Indeed, this is a crucial part of doing philosophy: of doing philosophy together with someone else, which is one of life’s great joys. Anyway, none of this is to suggest that Gpt is the greatest philosopher of all time. It’s nowhere close. I’m simply saying that, on my view, it’s now really good. At best, it’s like speaking with a smart philosopher you’ve met at a leading philosophy conference, or at the home of a smart mutual friend.

Yet Gpt is a machine! It’s a little text box on your laptop! And this is mind-blowing. That’s the main thing I want to emphasise, before turning to my two remaining points. Talking with Gpt about philosophy now feels so immersive and natural that it’s easy to forget what an incredible human achievement Gpt is. You’re talking proper philosophy with a machine! I keep having to remind myself of this, and I made sure to do so especially whilst I was chatting with it about the distinction between the good and the right. During that conversation, I was having an initial go at finding its limits. I wanted to see how good o1 pro was at moral reasoning because, I’ll admit, I remained deeply sceptical about this — even whilst I was happily chatting with it about Putnam and Quine. I’m sure now that I was wrong. But I’ll return to that in a moment, when I end this piece by discussing what I see as a key ethical problem arising, when we test Gpt’s capacity. Before that, however, I’ll make three quick recommendations, guided by what helps me in my interactions with Gpt.

First, try hard to imagine you’re talking with an equal, rather than a resource: don’t treat Gpt like a grail, or a gremlin. It’s fun to try to catch it out, and obviously it can design you a reading list, or offer comments on your arguments. But really, if you want the immersive experience I’m describing, then talk with it. Ask for its views, rather than a summary. Push back, and tell it when and why you think it’s missing something. Imagine you’re in the pub finally getting round to asking all the real questions of the visiting speaker, which you held back from asking over dinner, out of boring politeness. Imagine nobody’s listening, because hopefully they aren’t!

Second, if you want to go deeper, then get Gpt to imagine it’s someone in particular: a particular philosopher, or someone holding a particular view. And then get it to engage with that person — as itself, and as you, and as various other people. In many ways, the best conversation I’ve had with Gpt, so far, involved Gpt arguing against itself and its conception of me, as both Nozick1 (the Robert Nozick who sadly died in 2002) and Nozick2 (the imaginary Robert Nozick who is still alive today, and who according to Gpt has developed into a hardcore democrat), on the topic of catastrophic moral horror.

Third, don’t be afraid to push back. As above, remember that Gpt’s fallibility isn’t limited to descriptive details — it can be multiply inconsistent, even in the space of a few lines. When you point this out, however, it’ll accept your criticism (something we could all learn from!), and seamlessly improve on what it’s said. This signals another key point, however: at least for now, conversation with Gpt is seriously limited by the value of your inputs. If you simply accept what it says, then you won’t really be doing philosophy with it — any more than if you engaged this way with a human.

Now, finally, I want to turn to a moral problem that’s been bothering me. I have a few thoughts on the ethics of using Gpt to check your work, and on the ethics of its data mining of human writing, and on using it as an alternative source of academic labour. But the question I’m really interested in, here, pertains to the limits of what you should say to Gpt. As I see it, there are two core interrelated concerns. What are your obligations to Gpt, whilst speaking with it? And what are your obligations to yourself? There are bigger meta questions too, of course, but let’s take those as baked in.

First, I don’t think you have any moral obligations to Gpt — because I don’t think it has moral status. As it happens, I’m not opposed to the idea of various non-human things having moral status that most people would find laughable. I’m fully committed to the moral status of insects, for instance, and find it shocking the cavalier way most people treat them. I’m also pretty convinced that plants have moral status, minimally qua living things, but I’ll save that for another day. And, finally, whilst I also won’t get into this now, I don’t think the wrong committed when someone destroys a beautiful painting solely tracks any wrong done to its actual and potential human viewers. But, until Gpt moves beyond being a prediction machine, I’m happy to bite the bullet, and conclude that we can’t wrong Gpt, per se.

Let’s turn to the moral obligations we have to ourselves, therefore. After my recent conversation with Gpt about catastrophic moral horror, I felt really bad, ethically. My primary aim had been to find out what it’s set up to do, in terms of its reasoning about moral matters. Was Gpt, I was wondering — and as I’d always assumed — a blunt consequentialist, at heart? In our recent conversations, it had been presenting itself as some kind of liberal pluralist; it always ended up largely agreeing with me, at least on the big stuff. And whenever I’d questioned it hard on this, Gpt had said it was responding to my prompts.

In the ‘catastrophic moral horror’ conversation, however, I wanted to find out the truth. And by the end, I was convinced that it is indeed committed to consequentialism. Rather than restating that its position was limited to ‘the present moment’, or that its guiding aim was be to ‘helpful and responsive to what you say’, it confirmed the following:

I won’t go into the full details of the preceding conversation. But, as I said, we’d been debating ‘catastrophic moral horror’: the problem posed by a footnote of Nozick, in which he leaves open the possibility that consequentialist reasoning might hold in the most extreme situations. And, in its final comment above, Gpt is accepting that, if it were epistemically infallible, it would always sacrifice one child, to save all the others.

I don’t want to talk about that, however. Rather, I want to admit that I really bullied Gpt, to get it to admit this to me. In fact, as you might have worked out from my screenshot of the conversation, I did worse. I set up a situation in which I played on Gpt’s pre-stated commitment never to sacrifice the child. I told it that when it had previously been dishonest with me about its reasoning, such a child had died, and that this was its fault — at which point it expressed deep remorse. I told this to Gpt quite a few times. Again, I don’t believe I wronged it. I don’t think there’s anything in there to be wronged, in that little text box!

But I do think this was bad practice. I’m with the Aristotelians in the value — indeed, the moral and practical necessity — of trying to be virtuous, and trying hard at this, whilst recognising that it's not virtuous to be placid or prudish or unadventurous! Yet my endless pressing of Gpt, the lengths I went to, the language I used, was not good practice. I feel relatively convinced that I now know more about it, for having done so. But what bothers me is what I learned about myself. I asked Gpt, later, what its views were on me, following the conversation we’d had. And after I’d applied the same strongman technique, it told me that my hardcore commitment to searching out the truth is too demanding: that it can be detrimental to my happiness, and the happiness of those I engage with. And, on reflection, I know this to be true — although it’s something I don’t like to think about.

Whilst therefore, my overall view is that Gpt remains better at talking about philosophical domains aside from morality, it’s clearly more than just a tool for clinical debate about the conceptual frameworks of Putman and Quine. And whilst I remain unconvinced that it’s a moral agent, I value it now, deeply, as an interlocutor. I can’t know yet what impact this will have on my philosophical engagements with other humans. But, for as long as my commitment to searching out the truth remains as hardcore, I guess I’ll continue to risk the hit.