# Enrico's Presentation: Generating Philosophy with AI Transcript of [[Enrico Terrone]] presenting the "Generating Philosophy with AI" paper, March 2026. Cleaned from rough speech-to-text transcription. --- ## Introduction: AI in Science, Art, and Philosophy AI is already playing an important role in science. Philosophers are somehow pillars of human culture, and there is controversy over whether what AI produces — images or text that seem to have interesting resemblance to work so far — can count as real art or not. In science, on the other hand, it seems there are important cases in which AI is already giving relevant contributions to research. It's also easier in science, in comparison with both art and philosophy, to check whether what the AI is doing is an important achievement or not, because science has quite clear and uncontroversial success conditions. It's not so difficult to evaluate AI's outcome in a certain domain and establish whether it counts as an important contribution or not. In philosophy, this can be done, but it requires clarification on what success in philosophy is — and therefore what philosophy itself is. Whether we can count contributions of LLMs and generative AI more generally as interesting philosophy depends, first of all, on an important division in conceptions of philosophy. --- ## Two Approaches to Philosophy We have inherited two approaches to philosophy. ### The Text-Focused Approach (Functional) According to the text-focused approach, philosophy is located in the text. A contribution is a text exhibiting certain properties: coherence, explanatory power, integration. This is "functional" in the sense that what makes a text philosophy — and possibly good philosophy — is performing a certain function related to explanatory power. The function of enabling understanding, shedding light on a certain phenomenon, is something that texts themselves can exert. We read the text and understand more about what the text is about, if the text is good. ### The Practitioner-Focused Approach (Procedural) On the other end, there is a conception of philosophy as something in which the text is also important — we have to read something to engage with philosophy — but the text is just the endpoint of a process whose protagonist is the philosopher. This conception has to do with self-transformation, therapy, biography, phenomenological observation, or introspection. This account of philosophy has important precedents especially in the 20th century and late 19th century. [[Søren Kierkegaard]] and [[Martin Heidegger]] are examples of philosophers whose texts seem to be a sort of extension of their biography. --- ## The Science/Art Divide This division seems to track the science versus art divide. In science, usually we don't care about the scientist. We can be interested in the life of [[Albert Einstein|Einstein]], but to really understand general relativity, we don't need to know Einstein's biography. We don't even need to know who Einstein was — we can completely ignore Einstein and nevertheless afford to trust general relativity, if we engage with the physics text properly. Art seems to be a strongly procedure-based and practitioner-focused practice. Even though [[Caravaggio]]'s paintings are beautiful to see, impressive, if we don't know anything about Caravaggio — we just see this painting while ignoring when it was made, by whom, and in which way — this seems not to be the proper way to engage with artworks. Similarly with poems: without hearing anything about [[Eugenio Montale|Montale]], not knowing if he lived in the 20th century or in the 17th century, it seems we are not really getting the point of the poem. It's obvious and intuitive to say that philosophy is on the science side and not on the art side. But I can see there is a tension: in some traditions, especially in the so-called Continental tradition, philosophy is more considered as a kind of literature — a sort of literary way of writing, not novels or poems, but philosophical texts. In that case, if it's a kind of literature, the connection with the writer seems somehow constitutive of the work. --- ## The Core Claim The idea of this talk is that if we adopt the practitioner-focused approach, it seems very hard to say that LLMs can do philosophy — because they are not human, they don't have a relevant biography, they don't feel anything (as far as we know). We adopted the text-focused approach. We're trying to see whether LLMs can do philosophy within this framework, which seems more suitable to the way in which LLMs function. There's also a practice in contemporary philosophy — blind review — which seems to suggest that philosophy is about texts and not about philosophers. Otherwise it would be very odd to evaluate texts and select which to publish on the basis of the text alone, while it is mandatory to ignore the name of the philosopher. If the philosopher's life is so important for the identity of the work, blind review would be strange. --- ## Discovery vs Argumentation In this sense, if you adopt the text-oriented approach, philosophy seems much closer to science than to art. But there is also an important difference. Science is often about discoveries — finding something that was already there, but discovering it and bringing it to light. There is some counterpart in the objective world that is the target of science. According to at least one famous view — Quine, by way of [[Susan Haack|Haack]] — philosophy is more about making arguments. There is no prior reality the paper reports. The work is more to do with articulating concepts, offering reasons and arguments. So this introduces an interesting difference once we have the common text-based account. --- ## Likeliness and Loveliness On the other hand, there is another important analogy. There are two virtues of works in both science and philosophy. One is **likeliness**: the probability of tracking the truth, a sort of accuracy. An explanation is capable of tracking real facts. But there is also the **loveliness** feature, which is more related to virtues such as elegance, order, and capacity of providing a deeper understanding *if true*. It's conditional: in principle every theory might be true, maybe not in our world but in another world. And if we assume that it's true, and it would also provide a deeper understanding, that's a further feature of the theory besides being true. We can assess loveliness independently of whether an explanation is correct. In science, likeliness seems to be even more important than loveliness because there are experiments, measurements, and so on — that's the main test for the theory. But in philosophy, it's much harder to have an empirical test of a theory, and so it's often loveliness which is the relevant virtue rather than likeliness. That's a sort of consequence of this distinction between discovery and argument. --- ## Theoretical Virtues What makes the value of a work of philosophy are features like elegance and unity. A good theory explains a lot of things with a small set of concepts and principles — being non-ad-hoc, not having patches just to solve particular problems that have little connection with the other parts of the theory. Having a small set of well-connected principles that can explain the whole phenomenon, and not just parts where you have to add further parts to explain missing aspects. The capacity to fend off objections and draw distinctions at the right places makes our understanding of the subject matter deeper. In this sense, there may be — as something [[Timothy Williamson|Williamson]] and [[Luciano Floridi|Floridi]] talk about — progress in philosophy. We can try to do better philosophy than our predecessors. What is important: if philosophy is a sort of game of giving and asking for reasons, of finding the best arguments or the best objections, and LLMs are good at that, it's hard to deny that they are doing good philosophy — just like it would be hard to deny that [[Deep Blue]] is playing good chess just because it lacks biography, consciousness, and mental states. So far so good, in the sense that all of this suggests that just as we can make science with AI, we can also make philosophy with AI. --- ## Two Challenges But there are two challenges, and that's what we're going to consider in the remainder of the talk. ### Challenge 1: Abduction The first challenge has to do with **abduction** — the kind of epistemic operation that seems crucial to philosophy. Not induction from a series of observations, not deduction from general principles, but rather: finding the best explanation given a certain basis of evidence and certain analogies to other domains. Trying to find the principle that makes all of this explained in the most effective way. As a person who was mentioned before — and who has insisted a lot about this in contemporary philosophy, especially metaphilosophy or philosophy that has philosophy itself as its own topic — the importance of abduction has been highlighted by [[Timothy Williamson|Williamson]], and more recently [[Luciano Floridi|Floridi]] and others. The question is whether LLMs can do abduction. They seem to do something like abduction because they are not inductive in the way old artificial intelligence was — deducing from principles and drawing conclusions. LLMs don't work this way. They're not collecting cases and looking for general laws either — they're doing something more based on recurrences of patterns. So they are somehow abductive in what they do. But Floridi and others argue that they are not. They call it "zeroth-order abduction" or a sort of "inversion" of abduction, because what LLMs do is just look for the most probable next token. There is no evaluation of different systems of principles, trying to see which best explains the phenomenon. LLMs don't care about systems of principles; they just care about finding the next token that is most probable given the previous tokens. The model does not understand what an explanation is. Nevertheless, it produces text that follows the typical phrasing and structure of explanations. But again, this seems to mistake the centrality of text for the centrality of the procedure through which the text is produced. Even though the process is different, if the text has the same abductive virtues, the same explanatory virtues, that a text produced by human abduction would have, that seems enough to have a good piece of philosophy. It's very important to distinguish the process in which the text is produced from the epistemic import of the text. ### Challenge 2: Embodied Simulation (Zahavy) There is another objection we find more compelling than the Floridi one. The Floridi objection is softer and can still be resisted by relying on the distinction between text and procedure. Whereas this objection — raised against the possibility of LLMs doing innovative science — has a phenomenological core. [[Tom Zahavy|Zahavy]] says: Einstein imagined himself inside a falling elevator. He simulated the sensations and abduced that gravity and acceleration must be the same — the equivalence principle. The famous thought experiment through which Einstein arrived at his groundbreaking theory was not a logical exercise conducted in symbols but essentially one conducted in imagination. Even though the theory at the end is a matter of symbols and equations, the way in which Einstein could make this theory was by imagining himself in a sensory way — as a person, relying on sensory imagination, thinking of himself as being in an elevator. That was the element that led him to conceive the laws of motion in a completely new way. That's something LLMs cannot do because they don't have sensations. They don't have sensory imagination. They just manipulate symbols. The conclusion Zahavy draws is that AI can maybe do some physics, but cannot do *innovative* physics — because certain kinds of innovative science require this sort of sensory experience beyond the reach of artificial intelligence. His argument is about physics, but it seems to be even more relevant for philosophy. In certain areas of philosophy, it seems very important to consider what we feel — for instance, philosophy of mind is a lot about introspection, sensation, what it's like to see colours, what it's like to compare. Even aesthetics: when we discuss intuitions about what looks like good art to us, what doesn't seem beautiful at all. If we accept for the sake of argument that the value of philosophy is not in the process of production but in the text, it seems that without certain processes available only to conscious beings, certain texts cannot be produced. This is a stronger objection because it cannot be addressed just by relying on the text/process distinction. --- ## Response to the Challenges We think we can resist these objections. ### Response to Floridi First, we can find the relevant features in the text itself. Even if the LLM arrives at elegance and unification through statistical processes and not through real abductive processes, the abductive features of the text are enough to make it a good piece of philosophy. What seems more relevant to us is that the LLM can achieve this by relying on an accumulated corpus of text. Even though the LLM lacks abductive skills — because having abductive skills means being capable of comparing different theories and looking for the theory that best explains the phenomena (the LLM is not doing that, it's just doing next-token prediction) — it's doing next-token prediction by relying on a corpus of text that was produced through abductive processes and has a lot of other virtues. So the LLM can somehow find these explanatorily relevant patterns because it can do statistical prediction on the basis of texts that exhibit these features. Working on this corpus of selected text, the LLM can acquire a distribution of good explanations — of good abductive explanations. ### Response to Zahavy The same thing can be done for the problem of the lack of first-person experience. It's true that the LLM doesn't feel anything. An LLM has never tried anything, never took a debate, never felt anything in an elevator or any other situation. It has never had an experience of pain or pleasure, of fear. But there is a huge corpus of text in which human beings have described, recorded their subjective experience. The LLM can have, as it were, *second-hand subjective experience* — which can be as good as first-hand subjective experience, through the patterns of building theory that take into account the subjective dimensions. That's basically our idea: even if this is the most challenging argument against the possibility of LLMs doing philosophy, considering the fact that an LLM can rely on a data set in which subjective experience is sedimented — that will be enough to overcome this problem. --- ## Conclusion So that's basically it. Our conclusion. Thank you. --- ## Q&A Session ### Q1: What about training an LLM specifically for philosophy? **Question:** What would you say if an LLM was trained every day — specially trained to do philosophy? **Nick:** I wouldn't put it in terms of a philosopher training it every day. But I do wonder whether you could make an analytic philosophy specialised LLM by being very selective in the corpus and in exactly how you're doing reinforcement learning afterwards as well. **Enrico:** It's what we're already doing, in a sense. When we use an LLM, we upload our papers and try to train it to do philosophy the way we do. They seem to have strong potential for this. Even though they sometimes confabulate, there is a way of forcing them to focus only on certain specific texts, especially within a conversation. So if a philosopher every day uses ChatGPT, uploads the papers they're reading, the papers they're writing — in a sense it's sort of creating an avatar of oneself who is doing philosophy more and more in the way that person does. ### Q2: What about AI peer reviewing? **Question:** Peer reviewing is increasingly made by AI today. What do you think about that? Is it a deception, a fraud? **Nick:** Ultimately, the person who is supposed to be doing the peer review is responsible for what comes back. So I wouldn't say let's just put that all over to the LLMs. On the other hand, an LLM is perfectly capable — if prompted right — to give somebody very good feedback on their paper, in my opinion. **Enrico:** In principle, the final scenario may be one in which the LLM both writes the papers and makes the reviews — but it seems apocalyptic. On the other hand, it's just like a very good LLM playing chess against another. **Nick:** Going back to what Enrico said at the beginning: if you do subscribe to this text-based version of philosophy — and I think most of us here, at least the analytic philosophers, do — and if you also believe that we can make LLMs do philosophy well, why not, just for the sake of philosophical progress, turn it all over to the LLMs eventually? [This was posed as a provocation.] ### Q3: What about bad philosophy in the training data? **Question:** If in the data set there's so much bad philosophy — student essays, naive work, not elaborated enough — and the good philosophy is a minority, how can the explanatory abductive virtues be preserved? **Nick:** First, a little bit of a cheat: these days, apparently, the people who make LLMs are very selective in the corpus that is fed into the LLM in the initial stages. I'm pretty sure something like a crappy undergraduate student essay isn't going to be in the corpus. Second, a better way of thinking about this: the LLM, by being exposed to all this data and being trained up, is being exposed not just to a load of text but being exposed to a load of theoretical virtues. It's learning that when particular words in a philosophy context are used in a particular order, more often than not, other specific words come after them. What I'm trying to get at: we don't need to train on 10 million philosophers so much as train them to be respecting the way philosophical arguments go — or even logical arguments go. What would be written in the literature as a good philosophical move versus a bad one, versus a particular sort of counter you might make to a philosophical move. If you're thinking about training for things more specific like that, then I don't think we need to worry about undergraduates. ### Q4: What about justification and reliabilism? **Question (online):** When you presented the Floridi objection, you said something about focusing on the final output rather than the process. But the process might still be relevant because sometimes the way we get an epistemically valuable output is relevant for justification purposes — that's the point of reliabilism in epistemology. It doesn't just matter the final belief that is produced, because that might just happen to be knowledge by luck; what matters is also the way we get there. **Nick:** Something we should have been clearer about: the example we give at the beginning with Deep Thought and "42" being the answer — I wasn't envisaging it like that. I was envisaging getting a whole philosophical argument from these things. You're not just getting an answer, you're getting the steps of an answer — the reasoning up to it. When I'm using LLMs to do philosophy, I'll say things like "reason before you give me a final answer," which forces it to go step by step. **Questioner's follow-up:** I was wondering whether one couldn't push back and say: yes, the AI is capable of performing abduction, but at the wrong level. What they're able to do is abduction on a corpus to produce a text, but what you would need is abduction on the *content* that the text talks about. That transition isn't justified if you're basically locating abduction at the wrong level — textual abduction versus content abduction. **Nick:** A problem I've been thinking about — not the same as the one we've been thinking about today — is about physics and "folk physics" and LLMs' understanding of folk physics. A lot of people will say LLMs lack something called a "world model" — they're not interfacing with the external world itself. All they're ever doing is text. They've never seen a ball fall to the ground and bounce, but they've got loads of text about balls falling to the ground and bouncing. I think, in that case for physics, if you wanted to talk about the fine-grainedness of observations — exactly how balls bounce across the pavement — you probably *would* need a world model, being able to represent physical objects moving through space. But with philosophy, I would say a large proportion of philosophy doesn't need that sort of fine-grainedness. **Example from auditory perception:** I used to do stuff on auditory perception. I was always struck by the fact that philosophers interested in sounds always disagree about whether sounds are particulars or properties of objects or events. This was weird compared to colours — although you can disagree what colour is philosophically, everyone agrees that colours at least *look like* they're along the surfaces of objects. It seems to me that an LLM would be able to know, or at least give answers entirely consistent with, colours being along the surfaces of objects — despite never actually having seen an object with colours on its surface — purely because of the regularities in the text. But there isn't that clarity in the text between sound-as-property and sound-as-particular. I suspect it would be quite easy to confuse the LLM on this. ### Q5: What about phenomenological approaches? **Question:** What about a phenomenological approach? It doesn't really seem like you can get at the content that way — just at text. **Enrico:** I'm not sure I understood the whole thing. But there's a point that — even if you accept what we have said — there can be two conclusions to draw from our argument. **Weak conclusion:** LLMs can be good partners for philosophers. We can collaborate; we can do philosophy *with* the LLM. They can be very effective tools to do better philosophy. This is the idea of philosophy as "growing the image" — using art as the gardener uses nature, as a process that can be aesthetically encouraged and results in the garden. The philosopher can use the LLM to grow not images but philosophy papers, books, and arguments. Just that some crucial abductive passages are made by the LLM — a sort of outsourcing of parts of philosophical research. **Strong conclusion:** LLMs do most of the philosophical work. It's not "I have an idea for a theory about the mind-body problem, I have a sketch of the argument, I put it in the LLM, it makes comparisons with existing theories, gives me suggestions, I develop the argument, I put the text in" — that's more like growing the paper. If I just write a prompt saying "please try to solve the mind-body problem" and it produces a 20-page paper with a new theory of the mind-body problem — that's not really growing. My contribution is just a very obvious traditional philosophical question. In that case, the LLM is not just a tool for the philosopher but a full contributor. This may sound somewhat apocalyptic, somewhat inhuman. But again, this seems to have to do with the assumption that philosophy is in the process. If we think philosophy is just about knowing more about the world, and the LLM can produce better texts than humans — and we humans remain those who can read the texts and enjoy them — understanding of the world can be produced by LLMs, but understanding remains a mental, cognitive operation that as far as we know only humans can perform. LLMs can produce texts that enhance understanding, but understanding remains up to us. That's what makes all of this worthwhile. ### Q6: Corpus decay over time? **Question (online):** Suppose AI works based on that corpus. Here's a dilemma: do you put the AI-produced work into the corpus or not? If you do put it in, won't there eventually be a risk of decay of virtue with each new generation? But if you don't put it in, then you can have no progress — or you need humans to add more good quality stuff to make the corpus progress. **Nick:** My intuition would be: yes, you've got to put it in the corpus because it's going to be good philosophy like the rest of good philosophy. Then I'd probably say the onus is on you to tell me why that's going to definitely cause trouble. If you were doing this with literature — trying to make a literature-creating LLM — I can understand that if you have this sort of resolving-to-the-baseline thing, you get very bland, very boring literature. But bland, boring philosophy is sometimes the best, most correct philosophy. So I don't know — I understand the worry because people do talk about decay over time, but I wonder whether it would work the same way for philosophy. One more example: there are also LLMs used to produce mathematical proofs. And they've solved things — they must almost certainly be putting those proofs back into the training data. ### Q7: Final comment on artistic research **Question:** [Largely inaudible — about artistic research not being about traces or objects but about the progression of artistic practice, and how this might apply to philosophy.] **Enrico:** It seems to me that this is still an objection about the process. Our point is that — whether we are arguing for the weak or strong conclusion — the final assumption is that what makes a piece of philosophy good or bad is the text itself. Whether it's about justice, or knowledge, we take the text (as a reviewer does in blind review) and evaluate it by considering whether it enhances our understanding of the relevant phenomenon. That's what makes it good or bad. Sure, there is progress in this, and you have to consider what is already done at a certain moment. And this is the way science works: if a new theory is proposed, the key point is first to see whether it says something explanatorily relevant about the phenomena, and then consider whether other theories also say the same thing. So you need knowledge of the landscape of existing theories. But the issue is whether we have to take into account the subjectivity of the philosopher or not — and I don't see how this impacts our argument. --- ## Closing **Nick:** By the way, I should mention that this paper — at least my part of this paper — was very much grown with AI as well. **Enrico:** A little radical paper with the gardener approach. **Chair:** This is the last talk of the day. I want to thank you all. I hope there will be future occasions for us to meet and continue the conversation. Much success. **Nick:** How do you think it went? **Enrico:** Pretty well, I think. The last talk of the day, and despite a week of presenting, we presented very well.