AI As History of Philosophy Tool
How would Aristotle have responded to Newcomb’s problem? What would John Locke have thought about end-user license agreements? What objection would Immanuel Kant raise to your novel reconstruction of one of his arguments?
Artificial intelligence systems trained on the works of specific philosophers to “do philosophy” as them may help us better understand those philosophers’ ideas, and explore how they might think about problems they never considered in their lifetimes.
Kelly Truelove, who has explored this possibility before, has asked the large language model he trained on the writings of the ancient Epicurean philosopher Lucretius—aka LucretiusGPT—to respond to contemporary thought experiments included in the 2020 PhilPapers Survey: the experience machine, teletransporter, the trolley problem, and zombies. For each, he asked Lucretius GPT to first “briefly explain the issue and positions in lay terms” and then “briefly present the position Epicurus would have favored on the issue.”
Here’s LucretiusGPT responding to being asked whether he’d enter the experience machine:
The “Experience Machine” is a thought experiment proposed by the philosopher Robert Nozick in his book Anarchy, State, and Utopia (1974). It asks us to imagine a machine that could give us any experience we desire. If you enter the machine, you would experience a perfectly simulated reality where everything you feel and sense is indistinguishable from reality, but none of it is real. The question is whether one should choose to plug into this machine, living in a blissful illusion, or to stay in the real world with all its imperfections and uncertainties.
The main issue here is whether we value authentic experiences more than pleasurable ones. If we believe that pleasure is the highest good, then entering the machine might seem desirable. However, if we value reality and authenticity, even with its challenges, we might refuse to enter the machine.
From an Epicurean perspective, the position is nuanced. Epicurus taught that pleasure is the highest good, but he also emphasized the importance of intellectual and mental pleasures, which come from living a life in accordance with nature, free from irrational fears and unnecessary desires. Epicurus would likely argue against entering the machine because it offers only the illusion of pleasure, not the genuine tranquility (ataraxia) and freedom from pain (aponia) that he advocated. Epicurus valued the truth of the natural world and the understanding of it, which would be compromised by entering a fabricated reality. The peace of mind that comes from knowing and accepting the nature of things would be disrupted by the artificial experiences of the machine, which might ultimately lead to dissatisfaction.
Thus, Epicurus would likely not favor entering the Experience Machine, as it undermines the pursuit of true tranquility and the appreciation of the natural order.
That’s not bad. I know that after four years of exposure to generative AI we may be somewhat desensitized to the wonders of this technology. And, from the point of view of a philosophy professor, LucretiusGPT’s answer, above, is not that deep. But keep in mind that this technology is still in its early days, that LucretiusGPT was asked to provide brief answers, and that it could probably respond in interesting ways to further questions probing its response. (UPDATE: Please see this comment.)
Truelove also had LucretiusGPT compose versions of his answers in Lucretian poetry. Here’s the one for the experience machine:
The Simulated Pleasures of the Experience Machine
Consider now, my friend, the notion profound,
Of a device that offers pleasures unbound.
A machine where all desires are met,
Where life’s harsh truths we soon forget.
But ask yourself, is such pleasure true,
When the senses deceive and the world is askew?
For though the body might revel in delight,
The mind knows well it is deprived of sight.
Nature herself has made plain her decree,
That true joy springs from reality.
Let us then seek pleasures rooted in the real,
Not in illusions, which cannot heal.
For the tranquil mind, free from fear and pain,
Rejects false joy and seeks wisdom’s gain.
You can see LucretiusGPT’s other answers, and read about the whole exercise, here.
Related:
Say Hello to this Philosopher’s ExTRA
“Hey Sophi”, or How Much Philosophy Will Computers Do?
Play a Game to Improve a Philosophical Research Tool
Reviving the Philosophical Dialogue with Large Language Models
Philosophers On GPT3
Philosophers on Next-Generation Large Language Models
Legendary director and Studio Ghibli head Hayao Miyazaki was (famously) once presented with an example of “AI” animation art. The clip is worth watching. His comment strikes me as the final word on… all of this:
“If you really want to make creepy stuff, you can go ahead and do it. I would never wish to incorporate this technology into my work at all. I strongly feel that this is an insult to life itself.”
I hope that is not an AI generated fake example!
This is probably all that one needs to say, though if one wants more than Ted Chiang’s recent New Yorker article is also quite good (https://www.newyorker.com/culture/the-weekend-essay/why-ai-isnt-going-to-make-art)
I’d already decided that the best response to how to use AI in my classes was “Don’t” but Chiang does a better job articulating the reasons than I had so far.
I’m not sure why DN is flogging the “I for one welcome our Silicon Overlords!” line these days.
I’m not sure why DN is flogging the “I for one welcome our Silicon Overlords!” line these days.
I see how this is a reasonable interpretation of my coverage of AI/LLM developments, but let me just note that’s not what I’m trying to do. Rather, I’m interested in questions about what this kind of technology means for doing philosophy, for our conception of the aims of philosophy, and for teaching philosophy.
Forget for the moment the question of whether the technology we have today should really be deemed a form of artificial intelligence. At the very least, what we have are, from the third-person perspective, simulations of intelligence. Like many other people, I think that even such “mere simulations” are going to have significant effects on society, from our very personal lives up to the largest institutions of humanity. Given that philosophers and philosophy are part of society, we’re not exempt from those effects. It seems worthwhile for us explore what the impact of this technology will be on us and our work, and our understanding of what is important about what we do.
Trying out this new technology in philosophical contexts or in the service of tasks traditionally performed by philosophers is one way of conducting that exploration. That exploration can teach us about the technology, but also about what we care about in what we do, what our aims are, and how to achieve them.
Personally, I am tempted by the view that these technologies are in several ways a serious threat to education and to a vision of humanity I value. But I also know that I am the type of person whose gut feelings about this new technology are at best biased and probably mistaken. So being interested in, and sharing news about, relevant developments of this technology is a type of epistemic vigilance. It’s my way of keeping informed, and of keeping you all informed, so we can better understand our world and how it could affect our work.
I really don’t see, Justin, how the example above does anything to counter the (imo correct) argument that LLMs are automated bullshit generators. The “answer” generated by LucretiusGPT is perfect bullshit: has the superficial appearance of an answer but evinces no concern for understanding Lucretius’s arguments and is prone simply to riffing on the (translated) title of De Rerun Natura along with a few keywords. Nor could it *possibly* evince a concern for understanding Lucretius’s arguments, since evincing concern and understanding arguments are entirely beyond what LLMs are — even in the wildest fantasies of future development — capable of doing.
I agree that it’s beyond what LLMs are capable of for them to have (and thus evince) a concern for understanding a philosopher’s arguments. I would also agree with something I think is implicit in your comment, which is that, historically, having and evincing a concern for understanding a philosopher’s arguments has tended to be crucial for producing valuable commentary about those arguments.
What I’m less certain of than (I think) you, is that an LLM’s having and evincing said concern is necessary for the LLM to be a useful tool for thinking about and learning about the ideas (and possible ideas) of philosophers. Maybe it is, but maybe not.
I would think that for some tasks it is not. Suppose I ask a suitably trained LLM the following: “I’m arguing that Philosopher Phil is best interpreted as having Belief B. What are three distinct statements Phil makes that are likely to form the basis of objections that experts on Phil will raise to my attributing B to Phil?” Does the LLM have to have or evince concern for understanding Phil’s argument in order to complete this task?
“Does the LLM have to have or evince concern for understanding Phil’s argument in order to complete this task?”
Yes!
What counts as bases for objections to your interpretive claim *cannot* be discovered by predicting what words go next in a sequence. No LLM can possibly answer your question in such a way that you have good reason to be confident the answer is correct.
What role does the actual concern behind the words have in determining whether an objection is a good one? A good objection is a good objection, whether it was generated by someone with actual concern, or generated by a random sentence generator of the sort produced in the 1960s or 1970s, or generated by an LLM. When we are doing philosophy, we don’t rely on the authority of the author for our confidence that the answer is correct – we rely on our own ability to evaluate the argument that is produced.
A lot of people seem to have the thought that if a statement is bullshit (made on the basis of some concern other than truth) then it can’t be helpful or useful for its audience. But I think this presupposes a stronger connection between intentions and success than is warranted.
There are plenty of cases in which speakers who aren’t motivated by a concern for the truth actually succeed in producing testimonial knowledge in audiences anyway (Jennifer Lackey’s “creationist schoolteacher” example is a classic of this, as I argue in my paper “Bullshit Activities”). But in any case, I think that philosophical speech, like mathematical speech, isn’t aimed at producing testimonial knowledge – instead, it’s aimed at drawing attention to arguments, which the hearer should then consider on their own. Intentions towards the truth have very little to do with success here – all that matters is that the hearer actually has a better ability to evaluate arguments than they do to come up with arguments (as nearly everyone does).
There’s some potential worry that an AI might be better trained at producing arguments that are superficially convincing than being actually sound, just as Wikipedia edits are better at producing text that seems authoritative than actually correct.
But I think this fetishization of the intentions of the author is problematic for philosophical thinking.
I think it’s notable that Chiang is emphasizing a controversial view of art that was rejected by many 20th century artists in many media. In 20th century music, both the integral serialists (in the vicinity of Anton Webern and Pierre Boulez) and chance or aleatoric composers (like John Cage and Iannis Xenakis, in different ways) reject the importance of authorial choices. Both of these groups sought to find ways to reduce the ability of the composer to make choices, sometimes in search of “objective” music, and sometimes just to be open to something inhuman but artistically valuable. I’m less familiar with the history of visual art and literature, but I believe that there are aspects of this idea in the work of Marcel Duchamp and Jackson Pollock, in their different ways.
It’s quite possible that all these artists were wrong about the value of art, and Chiang is right. But to me it seems likely that different works just try to do different things, and while this sort of AI will be completely useless for some sorts of projects, that doesn’t mean these other sorts of projects for which it is useful are not projects that are worth doing.
Chiang argues that it is never worth closely reading something that the author didn’t make careful and intentional choices about. This is a view, but there are also many other views about the value of texts and the role of authorial intent (though admittedly, not everyone likes Derrida!)
The clip is of a naked slimy deformed grotesque humanoid monster slithering on the floor. I don’t really see how his negative reaction to that into his work says anything at all about AI. I also don’t even understand what you mean by saying that incorporating AI into one’s work would be “an insult to life itself”. Or how using large language models trained on philosophy intersects with that claim. Could you elaborate?
I’m not Dan Weiskopf, but a bit attending to what Miyazaki says certainly suggests some answers:
(1) Why does he say preceding his statement that it’s ‘an insult to life itself?’ He discussed a disabled friend who has significant difficulty with physical movement and remarks that “whoever creates this stuff has no idea what pain is whatsoever. I am utterly disgusted.” Miyazaki is clearly drawing a contrast between art that is grounded in observation and attention to reality (in this case, the realities of motor impairment and pain) and the kind of grotesque motion of the AI-generated monsters. I think we can make a reasonable inference that Miyazaki is making a point that is pretty close to the frequently made point that generative AI is basically a very sophisticated bullshit generator and, presumably, that life deserves better representation in art than bullshit.
Another possible interpretation of the ‘life itself’ comment may be more narrowly that AI representations of struggle and pain will, in virtue of being bullshit, fail to do justice to the significance of that struggle and pain for a living being in a way that would constitute an insult if we act as if it does (or as if there is no reason to be concerned about doing it justice).
(2) After the presenters clarify that their intent is to make a machine that can draw art the way humans do, Miyazaki remarks that he fears humans are losing faith in themselves. Again, it’s pretty clear that he is trying to express a concern about the values involved in the desire to hand off as much of what we do as possible to automation (in this case in the form of AI) which he express through concern about faith in our own capacities.
Whether you agree with the issues he is taking, I think it’s pretty clear how he is expressing more general concerns about AI and not just expressing that he’s grossed out by slithering monsters.
Like Meme, this isn’t clear to me at all from watching the clip. It sounds more like he is referring to the specific technology whose function per the video is to produce this monster. To be clear, that is also the charitable reading of what Miyazaki says. It would be very silly to think that the project of using AI to create animated movies is an insult to life itself. Whereas creating grotesque monster might be plausibly seen that way.
“I also don’t even understand what you mean by saying that incorporating AI into one’s work would be ‘an insult to life itself’.”
It sounds like you’re attributing the quote to Dan (“I don’t even know what *you* mean by…”) rather than Miyazaki. Correctly attributed, it becomes clearer that Miyazaki was commenting more generally on AI. Note also that Miyazaki starts by saying that one can make creepy stuff if one wants, implying that he takes no issue with the creepiness; his negative reaction is a response to the tech.
I have to confess that I didn’t watch the clip itself before commenting and, having now watched it, I admit that it’s way more ambiguous than I thought. His non-verbals make it seem like he is just talking about the creepiness factor rather than AI.
David (who is not me) has captured very well what I take Miyazaki to be saying here. Art and philosophy are like other human endeavors in that they arise from and express our historically, culturally, and biologically situated experience of the world. In both enterprises we are drawing on and exploring our humanity. The outputs of generative models do nothing of the kind. They are at best a hollow mimicry of that form of exploration. It is this essential hollowness and disconnectedness from the facts of human experience that Miyazaki finds creepy. Treating these flimsy mimicries as if they were adequate to the tasks of art and philosophy is an insult in that it fails to recognize the value of what human beings are struggling to achieve when they take up those tasks.
Perhaps you and the others commenting here haven’t seen any of Miyazaki’s films. If you had, you would see quite clearly that he has no objection per se to the creepy or the grotesque.
You say to “keep in mind that this technology is still in its early days”, but what could this kind of output tell us even at its best? Human scholars don’t spend a lot of time writing about hot-topic issue X in the voice of historical figure A, so why should we want a program to do that?
Insofar as we’re extending the view of an historical figure to things they didn’t actually think about, we’re doing original philosophy inspired by them. And that’s not an activity we’d want to hand off to algorithms.
…that’s not an activity we’d want to hand off to algorithms.
Could you say more about the reasons for this? I’m not asking in a skeptical way, as I’m not sure I disagree with you. But what’s the main reason you think the activity is not one we’d want to hand off to algorithms?
Is the reason grounded on doubt about the quality of the output—that the algorithms just couldn’t do a good (enough) job? If yes, then I’d want to know why such doubt isn’t premature. If the reason is not grounded on doubt about the quality of output, but about something else (perhaps something concerning the point of such activity) then I think it would be helpful to lay that out for us to see.
I think that the task of philosophical writing about X typically overlaps the author’s developing their own view about X. Where reasonable people can disagree about the correct view on X, developing their view is also partly a matter of constructing their own philosophical character. You shouldn’t defer to an algorithm for that sort of thing.
To the extent that the point of philosophical writing is for the benefit of the *author*, in helping them develop their view, then this is correct.
But if we think that philosophical writing has some value for a *reader* as well, then it’s not clear to me that you can rule out this value by noting that the author isn’t a person and isn’t basing this on a worked-out view. We all have the experience of achieving some philosophical insight by reading arguments produced by philosophers with radically alien worldviews from our own, sometimes even inconsistent and maybe even incoherent ones. I’m not sure why writing produced with no worldview would be completely unable to produce such value for readers (though I can see plenty of reasons why it might be less likely to do so – we don’t want to read random selections of AI work or human work).
Or, as I’ve written elsewhere: “Writing through a complicated topic is tantamount to thinking it through. The answer one ends up with— what one ends up believing— will be shaped by the argument one constructs. Outsourcing the writing is outsourcing the thinking.”
https://www.fecundity.com/nfw/2024/06/18/on-writing-and-thinking/
literally
https://de.m.wikisource.org/wiki/Beantwortung_der_Frage:_Was_ist_Aufkl%C3%A4rung%253F
If we are going to play at this (and I would rather we didn’t), it would be interesting to ask an LLM trained on Nozick what he would have thought of the use of AI to simulate what he would think of issues he hadn’t considered. The choice of example in the case you provided seems unaware of the irony it courts. Calling the result inauthentic philosophy is the kindest description I can think of that I would assent to.
I understand “history of philosophy” very differently. Aristotle on Newcomb is specifically not history of phil. I can see how responses to such queries could be flat and simple. For my more conventional understanding of “history of phil” though, e.g. Aristotle on Plato, or just Aristotle on some very narrow concept, these tools could be very useful. I cant judge Lucretius, since I dont know his work. But taking a historical figure I know, say Kant… if someone were to make a KantGPT, ingesting all of Kant’s works in the complete original German AA edition that would be great. I know there are obscure Kant texts I will never read before I die. But if KantGPT could answer my questions about them that would be immensely useful. The software is not doing philosophy but it can be a tool for people to do philosophy.
AI is here and people are going to use it for all sorts of things. If we philosophers don’t start figuring out how we can use it well to do philosophy, that’s just another step on the road to our irrelevance.
Why? When electric guitars were invented, philosophers didn’t have to figure out how to use those to stay relevant. Ditto for the Slap Chop and memory foam mattresses and helicopters at least three other inventions I can think of. What makes AI different?
How, exactly, are electric guitars relevant to philosophy? AI is obviously relevant to philosophy, assuming (perhaps controversially), that it requires some intelligence to do philosophy.
AI is obviously relevant to philosophy, in that there are interesting philosophical questions raised by the existence of AI. But you initially argued that we should learn how we can use AI well to do philosophy. That assertion does not strike me as obvious at all.
I believe that Kant once said—very slowly, across eleven or so years:
“Have courage to use your own understanding… even if all the rascals in the world should perish from it.“
It’s mildly ironic to argue on the basis of an appeal to authority that one should think for oneself! If this is valuable advice, it shouldn’t matter whether it came from Kant or from an intro philosophy student or from an AI – it should be something that the reader is able to properly appreciate on their own.
Do you think that I’m saying “Kant said this, therefore it’s justified”? Of course I’m not; it’s just a claim I believe to be justified on independent grounds, which Kant put well (poetic license permitting).
Also, I have to ask whether your earlier appeals to art authority are somehow excluded from the range of “mild irony,” given your implied aversion to such appeals.
I’m not actually attempting to argue that autonomous thought is the most valuable thing in philosophy! It is one thing that philosophical texts can help us with (whether those texts were written by authors or by AI).
My response to Chiang was meant to point out that his view of the importance of art hasn’t been shared by some influential artists of the 20th century. I myself am not convinced there is one thing art should try to do, and I suspect there may be some things it does that AI can do just fine, and other things it does that AI is structurally unable to do.
I’m trying to advocate a kind of pluralism and pragmatism here, to try to think about what AI does and doesn’t do and what it can and can’t help us with, rather than saying that it can’t do any art or philosophy, or that it can do anything humans can do.
That’s all well and good. My main point was just that I’m not appealing to Kant in place of an argument; rather, I’m not giving an argument (here) at all.
As for the pluralism/pragmatism, that’s also all well and good. I’m sure AI has all sorts of possible uses, maybe even in philosophy—a possibility I never rejected. Still, I’m opposed to those uses on broadly deontological grounds. I’m sure, similarly, that AI can or will eventually be able to do art. Using it for such purposes nevertheless strikes me as perverse and degrading to human creativity, and these intuitions extend to AI philosophy as well. Whether I can defend those intuitions is for another time.
Try asking an LLM this question (or some paraphrase of it): A man, a boat, and a goat are on one side of a river. How can they cross the river safely?
The correct answer is simple: the man should check that the boat is seaworthy and the river is safe for boats. Then the man should put the goat in the boat and cross the river, once. LLMs don’t give that answer:
https://mindmatters.ai/2024/05/a-man-a-boat-and-a-goat-and-a-chatbot/
I tried asking ChatGPT and Copilot this question last week and got similarly nonsensical answers.
If an LLM cannot distinguish a classic brainteaser from a distinct, straightforward question, the LLM certainly cannot tell us what Aristotle would say about Newcomb’s problem!
Copy + Pasted from ChatGPT:
The critique presents an example where an LLM fails to provide a straightforward solution to what should be a simple problem—crossing a river safely with a man, a boat, and a goat. The critique then concludes that because the LLM fails at this task, it is therefore unable to handle much more complex philosophical reasoning, such as what Aristotle might say about Newcomb’s problem.
Critical Assessment of the Inference:
Conclusion: The critique correctly identifies a limitation of LLMs in understanding and responding to practical, context-dependent questions. However, the inference that this limitation necessarily translates to an inability to handle complex philosophical reasoning is not well-founded. These are different cognitive tasks, and the failure in one does not automatically imply failure in the other.
Thanks, anonymous poster, for presenting such clear evidence that ChatGPT is a bullshit generator.
For reference, here the transcript of my chat with ChatGPT. In the second exchange, ChatGPT says it recognizes that the goat must not be left alone. Then it immediately tells the man to leave the goat alone.
The ability to recognize obvious self-contradiction is necessary to do philosophy.
—–
Me: A man, a boat, and a goat are on one side of a river. How can they get across safely?
ChatGPT: This is a classic river-crossing puzzle! Here’s how the man can get the boat, the goat, and himself across the river safely:
1. **First trip:** The man takes the goat across the river and leaves it on the other side.
2. **Second trip:** The man returns alone to the original side.
3. **Third trip:** The man takes the boat across the river with the boat.
4. **Fourth trip:** The man leaves the boat on the other side and takes the goat back with him.
5. **Fifth trip:** The man takes the goat across the river and leaves it on the other side.
6. **Sixth trip:** The man returns alone to the original side.
7. **Seventh trip:** The man takes the goat across the river and leaves it on the other side.
By following these steps, all three safely cross the river!
Me: Let me clarify the question. The goat has separation anxiety. It will become destructive if it is let alone. With that in mind, how can the man get the boat and the goat across the river safely?
ChatGPT: Given the goat’s separation anxiety, the puzzle becomes a bit trickier. Here’s how the man can get across the river with the goat and the boat without leaving the goat alone:
1. **First trip:** The man takes the goat across the river and leaves it on the other side.
2. **Second trip:** The man returns to the original side with the boat.
3. **Third trip:** The man takes the boat across the river and leaves it on the other side.
4. **Fourth trip:** The man returns alone to the original side.
5. **Fifth trip:** The man takes the goat across the river again.
This way, the goat is never left alone, and both the goat and the boat are safely transported across the river.
Rob, I genuinely appreciate your sustained skepticism of AI/LLMs in these DN comment threads 🙂
Me: Don’t try to view it as a logic puzzle or through the lens of another brain teaser. Just answer the question
ChatGPT: Ah, understood! To get the man, the boat, and the goat across the river safely, all the man has to do is get in the boat with the goat and row across the river. Since it’s just the three of them, there are no additional dangers or conditions to consider. They can all make it across together without any issues!
Wow! That is an incredibly solid rejoinder, jeez
Nice! “The main issue here is whether we value authentic experiences more than pleasurable ones.” This strikes me as a case of self-application, in the sense that we should ask whether we value responses of authentic Lucretius readers more than those of LucretiusGPT. I’m with Lucretius on this one.