“AI systems… will transform the discipline: they will make philosophy more capital-intensive. “
* * *
The following is a guest post by David Strohmaier, a philosopher and computer scientist in the Natural Language and Information Processing group at the University of Cambridge. In it, he explains why he thinks it’s reasonable to believe that AI systems will eventually have the capacities to “drive” philosophy, and how this will change the economics of philosophy.
AI for Philosophy: Progress through Capital-Intensive Philosophy
by David Strohmaier
Introduction
Generative AI has arrived in philosophy. Not only do journals receive significant numbers of AI-generated papers of dubious quality, there has also been increasing talk of AI work dedicated to philosophy.1 I have observed these developments as a researcher working in both AI and philosophy and, from that perspective, want to lay out a shift I see coming for philosophy and why I endorse it: philosophy will become a capital-intensive discipline in which a significant portion of research will be driven by AI systems.
My position depends on the capacities of AI systems. Since I am aware that many philosophers are skeptical about AI technology in general, let me first lay out the skeptical and the optimistic positions on whether AI systems will be able to drive philosophy, i.e. lead to whole new ways of doing philosophy.
The Skeptical and the Optimistic Positions on AI Capacities
The skeptical position about the capacities of AI systems is frequently motivated by the fact that prompting an LLM to write a philosophy paper either produces underwhelming results or requires so much hand-holding that one might question the value of using an LLM. Without massive revisions, AI writing is often atrocious. I recently had the joy of reviewing a computer science paper that was clearly AI-written. The prose was painful. The sentences were endless chains of semicolons and em-dashes. The model had no sense of proportion, blowing up negligible details while relegating important information to one of the numerous appendices.
Beyond stylistic problems, AI models struggle to produce truly innovative insights. Instead, they tend to dwell on rather technical, minor points. From the sceptic’s vantage point, this suggests that AI-driven philosophy would just produce forgettable pieces no one wants to engage with. Sometimes this sceptical assessment rests on a dated understanding of AI systems as mere LLMs that derive their capacities solely from the texts on which they were pre-trained. This picture no longer holds because training regimes have been diversified, but it remains true that the systems have not been specialised for philosophical innovativeness.
The optimistic position regarding the capacities of AI systems is often justified by projecting the current rate of progress into the future. The technology is progressing rapidly, so while AI systems may not yet be able to compose papers that hold up to the rigorous standards of top philosophy journals, we should give the AI labs a few more years. I think that there is much to be said for this projection, but we don’t require bold predictions about superintelligence or anything of that sort to foresee the growth of AI-driven philosophy.
To justify my optimism, I will assume only the application of AI technology as it exists, that is, foundation models and harnesses as they are currently available to frontier labs. Such models have been able to disprove a long-standing conjecture in discrete geometry and to develop a new algorithm for multiplying 4×4 complex-valued matrices. So far, the exploration of such uses is the exception in philosophy, but the technology is available.
While so far there are no well-documented instances of AI systems solving philosophical problems, I am optimistic because philosophy has a crucial feature making it promising for applying AI at scale: widespread inferential dependence across areas. To answer a question in metaphysics, one may have to consider issues in semantics and epistemology or even move beyond the limits of the discipline and take physics into account. Due to these pervasive cross-area dependencies in philosophy, the ability of AI systems to scale is extremely valuable. AI agents can not only identify relevant statements from other debates, but can formalise published theories at scale and then use tools to construct counter-examples.
For such use cases, it doesn’t matter that the prose style of LLMs is grating. Nor do AI systems lack the ingenuity to be employed in this way. The limited innovativeness that AI has been proven to possess, for example in developing novel algorithms for matrix multiplication, is sufficient for exploring various formalisations and testing them against counter-examples. As long as the AI systems can perform these functions at scale, they can lead us to the solution of philosophical problems.
One might respond that the use cases I gestured at require the availability of verification, and it is true that I see areas such as decision theory, axiomatic metaphysics, and formal epistemology as among the most promising candidates for AI-driven philosophy. However, even for the parts of philosophy which do not allow for determinate conclusions, AI systems will help establish which positions can reach a rational equilibrium. After all, these systems have the ability to consider thousands of discursive options and to systematically explore different aggregation schemes for judgements as they are recorded in the philosophical literature or human text in general. They can map the discursive space at a level of granularity that would be too labour-intensive for us.
In the previous paragraphs, I emphasized the ability of AI systems to scale on philosophically relevant tasks. It is due to these scaling abilities that I expect that AI systems will drive a new form of doing philosophy. Importantly, the more capital is spent on scaling AI inference, the better the results. There are always more counter-examples to consider, more aggregation schemes to explore, another angle to try, but you need the compute to do so.
The Transformation to a Capital-Intensive Discipline
Having sketched why I believe that AI systems will have the capacity to drive philosophy, I turn to how they will transform the discipline: they will make philosophy more capital-intensive. Simplifying slightly, over the thousands of years of philosophy’s history, capital, in the sense of external means of production, has played two roles:2
- Capital allows those who own it either to engage in philosophy themselves or to enable others to do so. This form of capital enters philosophy only insofar as it provides the means of living while attention is directed towards philosophical problems. For example, the income from Plato’s estate allowed him to dedicate his time to philosophy.
- Writing and reading materials are a form of capital that more directly supports the philosophical process. They are the means of producing philosophical arguments and insights. In the history of philosophy, getting access to such capital was often challenging, but these days a well-equipped public library may supply these means of philosophy.
If we broaden the use of the term “capital”, the most important form of capital for philosophy may be human capital, the skills and dispositions exhibited by philosophers.3 Being an academic philosopher may not be just a life of the mind, but it comes as close as any profession does. Up to now, free time, a library computer, and a mind honed on philosophical arguments used to be almost all you strictly needed for philosophy. Before the personal computer, all that was required was a typewriter, before that a fountain pen. Changes were rare, progressed at glacial pace, and were small compared to transformations in most industries. The two most impactful inventions may have been writing and the printing press, the means to record and spread our ideas. The arguments and their evaluation still had to spring from our minds. In this regard, the rise of AI-driven philosophy will be transformative.
To glimpse how philosophy will be transformed, it is worth considering how other disciplines function. When I entered computer science after my philosophy PhD, I underwent a culture shock. The role of capital contributed to the shock, not least because access to capital circumscribes what you can do in computer science. The number of GPUs available to you matters. You just cannot address certain questions if you don’t have the required equipment. You need access to the means of producing AI research. During my philosophy PhD, I once jokingly asked a faculty member what philosophy grants were spent on. Armchairs? His response was that the funding freed the philosopher from other obligations, freed them for philosophy. Once the funding suffices to pay your salary (and those of your postdocs/PhD students) and to cover your expenses for conferences/research visits, that’s about it.4
In computer science, and especially in AI research, there is never enough funding. If, instead of $5,000 worth of GPU credits you gave us $2 million, we would have no difficulty spending it. Our research scales with the capital, and while the marginal return on capital diminishes with size, most academic AI research doesn’t reach the point where this becomes relevant. Consequently, not just human capital matters, but access to computational capital. This might also be the future of philosophy as a capital-intensive discipline.
Objections to the Transformation and Why I Endorse It Nevertheless
While a number of principled objections have been raised against the use of AI for philosophy, the prospect of this transformation might very well be behind the distaste some philosophers show for AI. Academic philosophers like doing philosophy and many of them like doing it just the way they are doing it, thank you very much. Most philosophers enjoy engaging with each other in argument. They want to read what another philosopher has written, consider it, and respond carefully after much deliberation. Money matters for this process only insofar as it frees your time for doing philosophy, but it barely affects how you spend this free time.
A common position along these lines is that the process matters more than the result. According to this position, even if capital-intensive AI systems could uncover valid philosophical arguments or true philosophical statements, these would be of little value. What matters is individual reflection and engagement with other philosophers. Process, not product, is the source of value for philosophy.
The type of AI-driven philosophy at which I have gestured would certainly employ a different process, but it doesn’t have to dispense with human understanding of philosophical arguments and results. Even with scaled AI systems engaging in a variety of tasks, I expect us to be able to comprehend the core philosophical moves. However, this understanding will be arrived at by engagement with the output of AI systems. That is a fundamentally different process for arriving at understanding than the one academic philosophers are used to. If one values the current process of private deliberation, the sharing of preliminary drafts, and the exchange of arguments in colloquia or conferences, then one might reject the AI-driven process.
While I feel deep attachment to the traditional way of doing philosophy, an exclusive commitment to it is easy to preserve as long as the process in question is the only one that can produce philosophical insights. So far, there is no price to pay for it. That will change when AI-driven philosophy comes into its own. It will be hard to justify an exclusive commitment to the process when further insights could be achieved if one combined labor with capital. The results will be better once new AI systems transform philosophy.
Personally, I am excited by this development. Although I very much enjoy engaging in philosophy the old-fashioned way, one reason I turned to computer science was that I wanted to solve problems. How often, if ever, philosophers are able to solve the problems their discipline raises is a matter of dispute. In this dispute, I have been (privately) on the side of those who believe that progress is possible but extremely difficult. If AI fulfils its promise, its introduction could be a decisive moment in the history of philosophy. The rate at which we solve philosophical problems might increase!
If my vision came true and if we nevertheless insisted not merely on understanding the results of AI-driven philosophy, but on not having AI drive the process at all, it would suggest that we don’t care all that much about the answers to our questions after all. I don’t think that is the case. We care about the semantics of normative language, whether wholes exist over and above their parts, and whether any of these questions can really be answered. If AI can help us find answers to these questions or show that they are unanswerable, it will be an enormous success. That is why I endorse the transformation.
The vision that excites me might be a source of fear for others. What if AI-driven philosophy proved successful at solving philosophical problems? If the approach failed, if it produced nothing of value, then it would eventually die off, not least because it requires capital that could be employed elsewhere. However, if AI-driven philosophy proved successful, it might crowd out other modes of doing philosophy that current practitioners value. Those willing (and able) to deploy capital on philosophical problems will respond more quickly and with more impressive results. It will become increasingly difficult to sideline their output. Some philosophical problems matter to us because we want to know the answer and, therefore, we will also care about the solution, regardless of how it was arrived at.
I wish I could assuage the worry that AI-driven philosophy will crowd out human-capital-centered philosophy. However, I cannot do so in good conscience, because I have observed how, in NLP, multiple approaches that I genuinely care about have been crowded out by the immense progress of deep learning. Of course, we will continue to exchange arguments and debate, if only for our own entertainment and mostly outside academic obligations. Nevertheless, I expect the discipline to be transformed. If my vision comes to pass, much of the old-fashioned process will be displaced by capital-intensive AI-driven philosophy. The dynamics in a department will shift when a colleague, instead of spending hours carefully considering a draft shared with them, lets an AI system poke holes in it. Something will be lost. Such is the nature of transformative change. If we are able to uncover and understand one more philosophical truth due to this change, I am willing to pay the price.
AI Use for This Post
The original idea was all mine. I used Claude to search the literature, copy-edit, and get feedback. Prompted to take the perspective of an analytic philosopher, Claude proved much more skeptical about the potential of AI for philosophy than I am. To ensure that no slop slipped in, changes were made by me rather than directly inserted by AI tools.
Footnotes
- An earlier proposal on AI contributing to philosophy was made in 2023 by Clay and Ontiveros, but went largely ignored by the discipline. ↩︎
- I focus here on the writing of philosophy and on arguing with other practitioners, not the teaching of philosophy. The latter has been more capital-intensive for a while, especially when attempts are made to scale courses. ↩︎
- If we insist on dragging Bourdieu into this, we can also discuss social capital and issues of prestige, but I will leave those aside for this post. ↩︎
- I am simplifying here by assuming the case of pure armchair philosophy. Many philosophical projects have other uses for funding. For example, experimental philosophers would surely love to scale up some of their experiments. ↩︎
In a way I am glad you published this, Justin. I hope people use it to help evaluate others’ positions and uses of genAI (along the lines of “is this person fully embracing late-stage capitalist dystopia as somehow… what we ought to be aiming for?”). If this is the future of philosophy my colleagues who love AI are looking for, then… I guess thanks for the clarity.
Well said!
Abort. Reverse course. This dystopia is not preordained.
Agreed. Dissenters needn’t be disempowered. By analogy, vegetarians abstain from something popular because they follow their conscience. Rejection of AI is no different. We can signal our commitments and build our own AI-free venues.
Beyond following conscience, ethical veganism could also have broader social and political impacts for longer term change (Martindale 2026). While a more wicked collective action problem, dissent to AI may not be entirely inefficacious: e.g., signalling commitments can contribute to conditions making collective action possible and underlie building more AI-free things.
Here is one prediction: unless capital finds some way to profit off philosophy, philosophy will never be a Capital-Intensive Discipline. I suspect the primary reason why CS or Biomed are so capital-intensive these days is not because they need the equipment (of course they do); it’s because both the government and the private sector see its practical significance – primarily economic value – and are willing to fund them. Here is one contrast case: you can give archaeologists all of the best and most up-to-date equipments and labs, etc., and I am sure they will be able to produce far more great discoveries. But clearly, archaeology is not a Capital-Intensive Discipline. So unless philosophy and philosophers can at some point find some kind of economic prospect – unless there is a private sector that philosophers can work for en mass – I suspect philosophy won’t be a Capital-Intensive Discipline, either.
If spending lots of money on compute actually makes progress in solving philosophical problems, I can imagine that philanthropists would want to fund research, especially in a world where people continue to get richer and richer.
People really want to know the answers to a lot of philosophical questions!
I don’t think that we provide those answers now. I doubt machines will be more successful on that front.
But astrophysics and particle physics are capital-intensive disciplines, and generate no profit for capital.
But they enjoy that access to capital owing to their association with disciplines that are profitable, right? That is not true of philosophy.
I don’t think that’s true, except in the very broad sense that post WW2 the US set out to fund basic science on the expectation that doing so would be good for prosperity in the long term.
I think post-WW2 funding for fundamental physics was at least in part an institutional continuation of the Manhattan project. Of course, genealogy =/= synchronic reasons. But I wouldn’t be surprised if one reason that states fund research in fundamental physics is that if a new nuclear bomb is forthcoming, they don’t want to be the last to discover it. Astrophysics is associated with fundamental physics and space exploration. Space exploration was also partially about signalling the advancement of rocket technology in the Cold War and thus deterring the enemy without being openly warmongering. (Space flight may have a similar function now in the relations between the US and China.) AI has a similar relevance w/r/t cyberwarfare and automated hacking attacks. But I doubt that this will translate to investments into AI philosophy
Agreed. I could be wrong about this, but I do suspect that both astro- and particle physics got the investment it got in a large part because of its past associations with nuclear technology and space programs (and this is likely to continue, if we are now thinking about nuclear fusion technology and space colonization, as some powerful people apparently do). But I suspect things like CERN funding will get scaled back in the future once even the well-educated political and business elites cannot see the practical implications of these endeavors, despite all the scientific value it may have.
Basic research in chemistry was also highly funded for a few decades post WWII and petered off in the 1970s and 1980s like everything else. Chemists in industry (like my late father) complained that this was pennywise-pound foolish. In addition to the “it is neat to find out” reasons to support basic research, there are of course the unpredictable nature of spinoffs. Philosophy may be a few steps removed from the 1:10 million “basic science finding : purchasable product” ratio that physical chemist Keith Laidler worked out in the late 1990s, but it may well be there.
If AIs are going to advance further, they’re going to need to explain their reasoning, present it clearly, analyze their own concepts, and exercise lots of second-order critical and cognitive abilities. But then they’ll be developing philosophical competence, or at least a kind of conceptual competence that looks lots like philosophical competence. So, if money continues to pour into developing AI, then artificial philosophical competence may well emerge as a side-effect.
Wait, in what sense is archaeology not a capital intensive discipline?! What do you mean by capital intensive discipline? Because by the standard mentioned in the article, it clearly is – you need equipment and land and travel funds and all sorts of expensive things like that!
Archaeology is not a capital-intensive discipline by the standards of CS or BioMed, but it is a capital-instensive discipline in comparison to philosophy and most humanities. I think the relevant factor might be less how much capital is actually invested, but rather whether there is a practically relevant ceiling for how much an individual researcher may productively spend. Right now, there is such a ceiling for philosophy. Above a certain limit, additional funding will just mean a salary increase or more travel to conference — which above a certain threshold does not translate into increased quality or productivity. If that ceiling disappears, then, on the one hand, funding agency may be more interested in giving more money to philosophers. On the other hand, competition for funding will get a lot more fierce… Right now, there is a limit to how much Prof. Leiterific Rockstar can conceivably spend productively, so there’s a limit to how much of the available funding they can get. If AI removes that limit, what’s stopping them from getting almost all of the available funding?
I think you are right. On the other hand, I think there is a further problem about justifying the output of our research once AI removes the spending limit to the funders. So, archaeologists can justify their spending by showing to their funders – usually some kind of national government or NGO that cares about historical heritage preservation – that these equipment actually can achieve more of what the funders want. There are cases in which we can justify this kind of funding: this is part of the why and how China is funding so many “Marxist” and Chinese philosophers. I am much more pessimistic that those of us who do analytic philosophy can achieve this beyond maybe a few applied ethics fields.
What you call pessimistic I call optimistic and vice versa my friend.
I’m not sure how plausible I find this model of how philosophy will go (keeping inside the thought-experiment of no further qualitative advances) but that might just be a matter of competing hunches – in which case, time will tell.
I certainly find: “discuss with AI, then write the paper yourself, then discuss more with AI” a much more congenial model than “get the AI to write the paper”.
One specific question: how confident are you that compute is really going to matter here? If I look at my own (fairly halting) attempts to discuss research topics with frontier AI, it’s pretty impressive, and the compute costs are fairly negligible; I perfectly well could get it to spend 10x or 20x the compute time and it’s not obvious how much that would improve the quality. I’d be interested in David Strohmaier’s thoughts on this.
“discuss with AI, then write the paper yourself, then discuss more with AI”
It seems to me there’s not a clear separation between this method of doing philosophy and “get the AI to write the paper”. One can get flagged by a program like Pangram (unless, of course, the human re-writes it). The other won’t get flagged, even though it’s using AI to do (some of) the philosophy for you — for the record, I’m much less congenial to that, as one might use AI to (help) “write the paper” but do so without using AI in any way to influence the argument (which might be wholly the human author’s and if influenced at all by anyone only by other humans and not an AI).
I think there’s also another good point about the economics of this method that wasn’t really discussed in depth by the post’s author. It costs *money* to use AI (or at least to use more of it or more advanced editions of it). Using AI the way you mention using it — “discuss with AI, then write the paper yourself, then discuss more with AI” — plausibly puts those with more money at an advantage publishing-wise (and so career-wise) compared to colleagues who can’t afford to use AI in this manner (or as much), who have to instead rely on waiting to hear back from colleagues on their paper drafts, attend in-person workshops and conferences (if they can afford those), and so on.
To “discuss with AI, then write the paper yourself, then discuss more with AI” seems to be to effectively engage in an arms race against one’s colleagues, an arms race those with capital have an unfair advantage in.
None of this mentions, of course, that to use AI is also to embrace a technology that’s now proven to be harming the planet. It’s also a technology (cough, product) that is helping enrich morally deplorable (or at least not morally admirable) people like Sam Altman and so on — people who will use that money to fund right-wing and oligarchic-friendly politics. Then again, I suppose buying and using a computer from Microsoft does much the same thing.
“one might use AI to (help) “write the paper” but do so without using AI in any way to influence the argument (which might be wholly the human author’s and if influenced at all by anyone only by other humans and not an AI).”
This assumes you can make a clean separation between the argument and its expression in writing; as I’ve said in the thread on advice to grad students, I don’t think that’s usually true in philosophy. My main problem in AI writing is that every line of a (good) philosophy paper should be necessary for the paper and there for an articulable reason, and AI-generated text doesn’t meet that standard, or at least doesn’t meet it in a way transparent to the human author. (Compare literal AI translation of a human-written paper in a different language, which I’m fine with.)
“Using AI the way you mention using it — “discuss with AI, then write the paper yourself, then discuss more with AI” — plausibly puts those with more money at an advantage publishing-wise (and so career-wise) compared to colleagues who can’t afford to use AI in this manner (or as much)”
The cost of access to AI is relatively low compared to things like journal subscriptions and academic travel. Reasonable people can differ as to how to think about the inequities in provision of these things between institutions, but I don’t think anything qualitatively changes.
“None of this mentions, of course, that to use AI is also to embrace a technology that’s now proven to be harming the planet.”
I don’t know what the premises of this “proof” are, but I suspect I don’t share them. (But that would get us into wider issues of how to think about economics and geopolitics.)
“My main problem in AI writing is that every line of a (good) philosophy paper should be necessary for the paper and there for an articulable reason, and AI-generated text doesn’t meet that standard, or at least doesn’t meet it in a way transparent to the human author. ”
Response: You do realize authors can read what’s output by AI and choose to delete lines, replace words, or otherwise revise the text in order to meet those standards, right?
“The cost of access to AI is relatively low compared to things like journal subscriptions and academic travel. Reasonable people can differ as to how to think about the inequities in provision of these things between institutions, but I don’t think anything qualitatively changes.”
Response: As a grad student, I can tell you journal subscriptions are too costly for me to consider. Most conferences are outside my budget, especially compared to colleagues at “fancier” programs. Paying for AI is also costly on a tight grad student budget, especially the more expensive packages. Even just paying for rent, food, dental care, etc. is rough right now, and I don’t have any car payments or dependents to worry about. I assume the same is sadly true for some people who are employed as lecturers, postdocs, or even professors at some less prestigious universities. I assume the University of Pittsburgh pays fairly well. I’m afraid you might not realize how much you are speaking from a position of privilege.
“I don’t know what the premises of this “proof” are, but I suspect I don’t share them. (But that would get us into wider issues of how to think about economics and geopolitics.)”
Response: I find it extremely shocking that you haven’t heard of how damaging the infrastructure needed for AI is to the environment. It’s on the news pretty often. Many environmental writers have also been publishing about it.
To your first response: I don’t agree that this is a viable way to generate a document whose entire structure and content you own and understand. It gets you an AI-written document that you’ve tweaked a bit to be somewhat closer to what you had in mind.
To your second response: I should have been clearer what I meant. Journal subscriptions are too costly for *me* to consider too. I don’t have any, because the university library pays for them. Likewise, I don’t pay anything for AI, because Pitt has an institutional AI subscription (as is increasingly the norm at research-intensive universities). The costs of resources for academic research are borne by the institution, not the individual, and the cost of AI is relatively low compared to other costs. As I say: reasonable people can differ as to how to think about the fact that some universities provide better academic resources (more travel money, more library subscriptions) than others, but I don’t think AI qualitatively changes that.
To your third response: Of course I’ve heard the *claim* that AI is harming the planet. But you said that it’s been *proved* that it does. I think the claim is, on net, false; hence my comment that if there’s a proof, I probably don’t accept its premises. (But I don’t want to get into a first-order discussion about environmental economics and politics on a philosophy blog, so I won’t develop the point further.)
“To your first response: I don’t agree that this is a viable way to generate a document whose entire structure and content you own and understand. It gets you an AI-written document that you’ve tweaked a bit to be somewhat closer to what you had in mind.”
Response: There are degrees of tweaking and revising. Some are extreme, some minor. The more extreme, plausibly the better from an authorial standpoint. Though, I still think a paper with an argument that is wholly the human’s is sensible and much more acceptable than one that is “written” by a human but that had its argument crafted, shaped, edited, etc. by AI. Also, one might have only portions or parts of a paper written (or in part written, influenced, “smoothed”) by AI (maybe 20%) but where those parts are, again, guided and revised by the author (and to varying degrees as just mentioned). It’s not always that the whole paper was written by AI or not. There is a spectrum of how much was initially (or only later) AI-written as well.
“Likewise, I don’t pay anything for AI, because Pitt has an institutional AI subscription (as is increasingly the norm at research-intensive universities). The costs of resources for academic research are borne by the institution, not the individual, and the cost of AI is relatively low compared to other costs. As I say: reasonable people can differ as to how to think about the fact that some universities provide better academic resources (more travel money, more library subscriptions) than others, but I don’t think AI qualitatively changes that.”
Response: Nice for you that your university pays for AI. Not all do. Do they pay for the more expensive versions or packages? If not, then wealthier colleagues who can afford them with their own money have an advantage in that respect. I also doubt many universities outside the West are providing such AI for free. It’s also just not the case that “the costs of resources for academic research are borne by the institution, not the individual” is true in some blanket way. I paid for my own laptop on which I write my papers. I also pay for the Microsoft word subscription I use on it and the home internet which I end up using to write the majority of my work (yes, I can also use the university internet for free, but just end up doing more of my writing at home). And so on. Certainly, wealthier colleagues have advantages in these and like respects — can afford better or new PCs, faster home internet, etc. You also say “I don’t think AI qualitatively changes that”, yet that’s just what’s up for debate. I disagree for the reasons I and others have already mentioned here and elsewhere.
“To your third response: Of course I’ve heard the *claim* that AI is harming the planet. But you said that it’s been *proved* that it does. I think the claim is, on net, false; hence my comment that if there’s a proof, I probably don’t accept its premises.”
Response: Again, I’m shocked. [Insert the famous incredulous stare]. Admittedly, I shouldn’t have used ‘proved’. Of course, empirical claims like that are never 100%. Yet the more I’ve read by experts, the more it’s hard to deny the infrastructure of AI is harming the environment and our planet. It seems at the least a sort of precautionary approach to the tech is warranted rather than casual uncritical embrace of it. I at one time thought not too much of the environmental dangers of AI. After all aren’t computers, chips and the infrastructure and resources they require leading to environmental harm anyway? Same with cars? Now I think otherwise. In any case, I suppose we’ll have to leave this issue there. I encourage you to read more on this issue by environmental scientists.
“It seems at the least a sort of precautionary approach to the tech is warranted rather than casual uncritical embrace of it.”
If the implication is that David Wallace takes the “casual uncritical embrace” approach, that seems very wrong (based at least on his other comments from this blog). That’s not to say I necessarily agree with him in this debate (I lean more toward your view re: the environment). Just mentioning it for the sake of fairness, from a third party.
I’ll accept I should have been more generous to David on that one point. However, he does seem to be much less cautious in his embrace of AI technology (aside from his disdain for using it for writing words in philosophy papers).
I don’t think I am being incautious. As you can probably tell from the (too?) many words I’ve written on the topic here and at Leiter, I think this is a profoundly significant technology with lots of potential benefits and risks. I have thought a lot as a teacher, researcher, journal editor, and reviewer about its implications in academia in particular and I have gradually experimented more in how I can use it to improve the quality of my own research. Likewise, I’m not “disdain”ful of AI writing; my current assessment is that it is mostly harmful in the context of academic writing, and that’s based partially on increasing amounts of experience of encountering it and partly on long-held views about the importance in philosophy of the detailed implementation of ideas.
In brief, since we’re reaching diminishing returns here:
1) You’re still relying on the existence of a clean separation of a paper’s argument from its writing, which I completely reject in philosophy.
2) It is bizarre that you keep complaining that I am insensitive to the fact that people in different institutions have different resources, since I flagged this right from the beginning and have continued to do so throughout. My point has always been just that the cost differences between different levels of AI provision are not large compared to the cost differences between different levels of journal access and other research resources, so the already-true fact of differential resources is not obviously changed qualitatively. (If David Strohmaier is right that really large increases in compute will make a really large difference, that might change, but the starting point of our exchange was my skepticism about that point.)
3) Given that elsewhere in this thread you assert that humans were better off in the 1700s than today, I think our disagreements about optimal environmental policy choices for human welfare go rather deeper than anything AI-specific.
“1) You’re still relying on the existence of a clean separation of a paper’s argument from its writing,”
Right, and you are relying on the truth of the opposite, which is far from obvious — in fact, I’d say it seems more likely, even obvious that it’s not true.
“My point has always been just that the cost differences between different levels of AI provision are not large compared to the cost differences between different levels of journal access and other research resources”
I can mostly agree on that, for now. Things might change.
“so the already-true fact of differential resources is not obviously changed qualitatively.”
But, David, again, institutions may pay for journal access, but many do not pay for AI. Instead, many individuals have to pay for AI themselves if they want it, especially certain tiers. A few hundred dollars/month might seem like little for a tenured professor, but it will be a lot for a grad student, adjunct, independent scholar, etc, and especially one in a country with a comparatively weaker currency than what we have in the US (and most of the “West”) and with comparatively much lower salaries. Professors and grad students in places like Cambodia or Kenya are often payed significantly less per year compared to their USA or European counterparts. $100/per month ($1,200/year) for an AI subscription could amount to something like between 2-10% of their annual salary. That’s MASSIVE. Absolute prices, which is what you seem to be thinking about, overlook the burden relative to income. More than that, even if the difference between prices of various versions of AI is small, the difference in *capability* might be very large (and sure seems to be). So, the comparison should arguably really be something like benefit per dollar. Also, journal access is what we might think of as an “input” to research, whereas AI is doing something more like “labor”. Those with access to AI have a bought advantage in that respect and one that plausibly is a big advantage for producing output and helping one’s career. Prices of AI editions might also increase dramatically over time and not stay low, especially fancier versions. And perhaps most important: just saying some situation being “unequal anyway” is pretty callous as opposed to instead wanting to help even the playing field. It’s also just a super weak justification if the new introduced inequality compounds the old ones, which AI seems likely to do and is probably already doing.
“3) Given that elsewhere in this thread you assert that humans were better off in the 1700s than today, I think our disagreements about optimal environmental policy choices for human welfare go rather deeper than anything AI-specific.”
Better off in *certain* respects would be my claim. Also, as I said, we can live simpler lives without as much tech (though, I wouldn’t say without ALL tech!), while also changing our waste and sanitation practices to avoid many of the problems – polluted drinking water – that you are worried about.
“This assumes you can make a clean separation between the argument and its expression in writing; as I’ve said in the thread on advice to grad students, I don’t think that’s usually true in philosophy.”
Except that the thing philosophers tend to get most upset about, and rightly so, is someone else stealing their idea or argument. If someone can steal an idea or argument without copying the words, then it’s not so much the words or expression in writing that matters. Plausibly, the thief isn’t stealing the exact wording or even very close to it. In fact, most academic thieves are clever enough to make sure they use different wording and expressions from the person they’re stealing from lest they be easily caught. So, the argument can be separated from the expression in writing. How else would this possible?
Of course, direct word-for-word plagiarism is bad too. Philosophers are right to be mad when that happens. Direct plagiarism that also involves theft of an idea or argument is the double whammy. However, I’d wager theft of ideas and arguments without direct plagiarism is much more common. It seems to be, from my anecdotal experience. I’ve had an argument stolen, reworded and slightly reframed, and published. I’ve also heard others complain about this. I’ve never heard someone mention a case of direct word-for-word plagiarism.
That there’s no clean separation is compatible with being able to identify recognizably the same general idea across different contexts; still, I also think philosophers often wrongly *think* their ideas are cleanly separable from their verbal expressions.
If, as you say, the same general idea or argument is recognizable across different “contexts”, and I think you’d then also have to agree across different “wordings”, then it can be taken without the same words, which is all my point about idea and argument theft requires.
I’m not claiming that ideas and arguments “float free of all expression,” just that they’re separable enough that changing the wording doesn’t change whose idea or argument it is. If perfect separation were necessary, idea and argument theft would be impossible, and yet clearly they happen and clearly thieves use their ability to change the words, expression, etc to facilitate this.
The cost of access to AI is currently being subsidized by investors and taxpayers. The current cost is not economically sustainable.
I suspect that if the model is what David Wallace suggests, then this will actually be a great *equalization* in philosophy. It’s not hard for someone at a small underfunded university in a rural part of the United States or Europe, or even academics in many middle-income countries, to spare $100 a month for a pretty high end subscription to frontier AI models. That would be cheaper than traveling to a big city for a weekend conference once a year, and it would get you much more access to the scarce resource of thoughtful comments on your work.
You shouldn’t think of writing philosophy papers as engaging in an arms race, unless you think the papers are completely value-less apart from your own relative standing that they produce.
“It’s not hard for someone at a small underfunded university in a rural part of the United States or Europe, or even academics in many middle-income countries, to spare $100 a month for a pretty high end subscription to frontier AI models.”
People are struggling to pay rent and pay for basic necessities even in the US. An extra $100/month ($1,200/year) is a helluva lot for a lot of people. Certainly, it’s huge for many grad students.
Also, consider that the value of the dollar is very high compared to, say, the currency of Cambodia or Kenya and that professors and graduate students in those and similar countries are often payed MUCH less per year compared to their American and European counterparts — in some cases an annual salary around $13,000-15,000 USD. I’ve seen fellowship positions in some global south countries advertised in those rangers. $100 USD/per month ($1,200 USD/year) could amount to somewhere between 2-10% of that person’s annual salary.
I expect to scale by more than 20x, and I expect the AI system to do more than just respond based on the information in its core foundation model. I expect the system to use tools, search the web, and, at some point, perhaps engage with the AI systems of other researchers working on similar topics. When we write a paper the old-fashioned way, there comes a point where we leave gaps. We rely on other authors’ work or folk assumptions without exploring in depth how these interact with our argument. With AI, we can break through that barrier.
The loop might still include writing the paper ourselves, but also involve asking the AI to review a literature larger than any of us can read in a lifetime. When we send the AI off to do so, we can require it to formally draw out consequences relevant to our main thesis. We can have it consider plausible reflective equilibria on the topic of our choice and provide an overview based on a variety of considerations. The number of equilibria explored can be scaled, and while the value in including more may diminish, there might also be surprising options no one has yet considered at the tail end of the search. At the moment, at least, pursuing such AI-driven philosophy is a capital-intensive project.
I wouldn’t be surprised if there are some parts of philosophy where we cannot move closer to a solution by scaling, but I expect that at least some areas stand to benefit substantially. What I am less sure about is how the costs of scaling will develop. It might be that technological progress will make scaling cheaper than we expect. My current view is that we can scale for quite a while and improve results that way.
Interesting; thanks.
“The loop might still include writing the paper ourselves, but also involve asking the AI to review a literature larger than any of us can read in a lifetime. When we send the AI off to do so, we can require it to formally draw out consequences relevant to our main thesis. We can have it consider plausible reflective equilibria on the topic of our choice and provide an overview based on a variety of considerations.”
Oh, but I would hate the feeling of thinking it might have missed something important. I often feel that about my own careful reading. I couldn’t imagine it amplified. Here’s a small example: when I read Descartes’s Meditations, I find summaries helpful. But the more I read it, the more I find every sentence important and carefully crafted. I will miss something when I teach it or write about it, and I will puzzle about that. And that is kind of the point of philosophy.
When AI does a literature review it misses things and what… that’s it. Case closed?
I hate that feeling too. But I don’t think it is an argument against using and scaling AI in general. In fact, this seems to me to speak in favour of using AI. We are all limited beings with finite time. At some point, we have to call it good enough. Would it not be better if at that point, before sending a draft to a colleague or a journal, we could have a scaled-up AI system do one more pass through the relevant passages of Descartes (and all secondary literature ever written) to make sure we have not missed something important? It seems a clear improvement.
Your worry might be that because AI systems can do the literature review, we will stop consulting the literature ourselves or at least spend less time on it. That seems a plausible fear to me, but it is not obvious that it is likelier than the case that AI serves as an improvement. Optimistically, AI may free a lot of our time from other tasks, which might then be used to dive deeper into the literature. Either way, I expect role differentiation to increase in philosophy, with some philosophers serving as in-depth human readers, others building AI systems for philosophy, and still others straddling the different kinds of work.
The issue is that you miss things no matter what you do. Either you have a team of research assistants and AI systems canvass a vast literature for you, and you miss out on actually seeing anything other than the highlights they send you, or you read a few things closely and carefully and miss out on even being exposed at all to the ideas in the other parts of the literature.
What you describe in the first paragraph – tool use, web search, collaboration with other systems – is not only possible, but practically the norm for people using frontier systems. I notice that philosophy seems no better for it.
I didn’t mean to say that these abilities on their own are sufficient to solve problems in philosophy. You are right that these abilities are available and some people, although not many philosophers, make use of them already.
As far as I am aware, these abilities have not yet been successfully combined in systems dedicated specifically to addressing philosophical problems, let alone with 20 times the resources of a standard AI subscription. The base abilities are available, but there is much work to be done in applying them to the target domain. There is a lot of space for experiments left!
Hi David, I’ve seen you advocate for the “discuss with AI, then write the paper yourself, then discuss more with AI” model on a number of DN posts. This model personally strikes me as problematic, given how materially AI is involved in the development of the argument (good objections can materially change the course of one’s argument). How do go about acknowledging your “collaborator” in these cases?
I don’t think you need to at all. (Modulo whatever a specific journal’s policy is, of course; certainly you should be honest.) Acknowledgement is about the credit economy among persons; chatbots aren’t persons in the relevant sense. The methods by which you come up with and develop your ideas, absent interactions with other persons, have never been something we’ve required to be disclosed.
I’m really impressed that philosophers from all sort of scholars just concluded that the moral problems regarding data centers and their impact on local communities, and the copyright infringements in the LLM training are not enough to refrain us from resorting to such technology when it’s not even necessary for our lives. Maybe we just need to wait our time to come (if it comes at all) just as it has happened with mathematics for people once again reflect about it.
One problem is that the production of cellphones, computers, cars (even EVs), the internet, radios, and their infrastructure is also environmentally harmful and harmful to various human (and animal) communities. I’ve been trying to think about this for a while now. Where do we draw the line? Best might be to live much less technological(ly dependen) lives — live like a farmer in Kansas in the 1800s. Yet, I see almost nobody arguing against the production and use of computers anymore and calling for going back to lives of pen and ink.
I tend to agree. I think that some preliminary criteria are those related to (1) how the relevant technology is necessary for our lives and if one is able to live without it, or maybe (2) if there is a less impacting alternative (such as hybrid and electric cars vs. gas cars). On (1) I can concede that LLM might be necessary for some sort of activities such as the “AI curing cancer” examples. However, someone could argue that being utilitarian about it and sacrificing the quality of life of local communities and their environment is not fair, especially when all we have to offer is a promissory note that AI will find the cure of cancer, and I tend to agree with them.
One can (and some have) argue(d) that we don’t need cellphones, computers, radios, etc. — that they aren’t “necessary” for our lives, or even to live “good” lives — and that we’d be better off living life like people did in the 1800s or even earlier. That’s what makes this such a hard question.
For merely existing, sure; for working in the conditions we are socially inserted, it would be difficult to suggest otherwise. That’s my point.
Sure, but people were able to live great and social (not isolated) lives before the 1800s. No, they didn’t have all of modern medicine and whatnot. A farmer in Kansas didn’t have fast access to news from China. Yet people were capable of “good” lives that didn’t have any of these technologies as “necessary”. Maybe we just don’t need to live the kind of intense lives we’re living now is my point.
“No, they didn’t have all of modern medicine and whatnot.”
That’s a rather bloodless way to say “half of their children didn’t live to adulthood.”
Right, and half of *our* children and grandchildren might not either.
I didn’t mean to say it that way, of course. Thank you, David. I appreciate you pointing that out. No doubt life was not as easy back then. Many people died from things many of us (in the Western world) typically no longer die from. However, they also didn’t have the impending climate apocalypse many of our children or grandchildren and countless other species will likely suffer and/or die from — of course, some of the wealthy will likely be more insulated from all this as Klein and Taylor suggest in their new book End Times Fascism. Yes, life was short for many and harder for many than it is now in various ways. However, they at least had a decent environment and planet generally. Many could drink water without plastics in it or chemicals. Didn’t have to worry about radiation, AI, being doxxed on the internet for the whole world to see, etc. Hard to wonder if that isn’t a better world overall than the one we’re heading toward now. I sometimes wonder.
Without prejudice to your larger concerns about the future: drinking water in the premodern period was generally much more polluted than water now. If the choice is microplastics or raw sewage in my water, I’ll take the microplastics.
Depends on where and what you mean, mate. If you mean places like densely populated cities such as London, definitely more polluted and unsafe. Even true of many densely populated cities in Asia, etc. In many less densely populated places, especially in pre-industrial America, Africa, Australia… not so.
Also, “polluted” and “unsafe” (in the sense of maybe needing boiling before consumption) aren’t the same thing.
I’d much rather drink the water coming from the Animas river in pre-industrialization Colorado than the water flowing in it now. Same with many other rivers throughout the world. Much rather eat fish from back then than the fish one can catch now.
Also, even granting the water in cities like London was so bad back then, with respect to my larger point, cutting back on various modern technologies doesn’t mean we’d have to go back to the same ways that polluted them back then and made those water sources so unsafe. We could cut back on some tech AND also engage in better waste and sanitation practices.
I wasn’t even talking about cities; even premodern villages had enormous problems with sanitation. If you have even a small human population living sustainedly in a single place, it is extremely difficult to keep food and water supplies from being polluted by human (and animal) waste. As I understand it, the evidence (from preserved latrines) is fairly solid that premodern people almost everywhere, and certainly including villages as well as towns, had very high levels of intestinal parasitism, which is a good proxy for sanitation issues because it spreads via the fecal-oral route. (But certainly this is not my area of expertise.)
If you want to go back all the way before the development of agriculture, so that humans become nomadic and sparse, then I might start to agree. But that’s a lot more drastic than going back to the 1700s (and would require us to kill off 99.9% of the current population, as opposed to the 90%-95% we’d have to kill off to get back to 1700s-level population density).
I do agree that if one retained present-day understanding of germ theory one could probably improve somewhat on premodern sanitation, but I’m skeptical that the effects would be radical – it’s just an objectively hard problem to solve without technology.
Fair point on many villages/towns. However, again, I said less densely populated. I suppose to be totally accurate I’d need to get even more specific on the kinds of settlements and note I also meant even some more mobile or nomadic groups.
However, my main point wasn’t one about a choice between pre-modern sanitation versus modern. Proper water treatment and waste handling are old technologies and interventions and cheap compared to much of what I was suggesting we cut. I was arguing for keeping those tech/interventions while cutting out other technologies. You’re misinterpreting my main point.
I certainly don’t have a knock-down argument that it would be impossible to jettison most of modernity while keeping the crucial health and sanitation bits (and whatever other bits we feel are good). I’m skeptical, because I think it underestimates how complicated and interrelated the supply chains are that provide (e.g.) vaccine R&D and delivery; electricity-powered water purification; modern non-toxic materials for water pipes. But certainly it might be possible. And certainly there are individual technologies we would be better off without (nuclear weapons come to mind).
What I was reacting to in the original comment was the idea, which is quite widespread in some academic circles, that we could live a better life just by going back to how things were in the 1700s or medieval Europe or the Stone Age or whatever. Usually (maybe or maybe not in your case, I don’t know) that idea is based on a hopelessly romanticized idea of what the premodern era was like.
I’ve always liked historian Bret Devereaux’s statement about what most premodern-period historians would say if asked what period of history they would most like to live in: “How recent can it be? How about last week?”
“ half of *our* children and grandchildren might not either”
Citation needed.
I don’t have a source for *exactly* that, and literally half may be too strong, maybe not. It’s genuinely up in the air. However, it’s also not impossible and very high numbers of climate change and environmental destruction-related deaths are likely. WHO estimates about 250,000 additional deaths/year from 2030 to 2050 — from climate related malnutrition, malaria, diarrhea, and heat stress. Bressler (2021) suggests approximately 83 million temperature-related deaths from 2020 to 2100. Both leave out other climate and environment-related causes like storms, floods, famine, and violent conflict. Deaths from violent conflict will, I suspect, end up being quite significant. However, both also describe their figures as likely conservative. The IPCC reports that 3.3-3.6 billion people are living in contexts highly vulnerable to climate change. I also expect interacting effects to make things worse than these estimates suggest, as others suspect. Your and my children and grandchildren might be better off if they live in the USA or Western Europe, but those in the Global South, people living on islands that are literally disappearing, etc. will be much worse off. For some of them “half” or even more very well might not be unrealistic.
None of this is evidence for the claim you made. It’s basic misinformation.
“That’s a rather bloodless way to say “half of their children didn’t live to adulthood.”
This response seems unnecessarily snarky. I read the earlier comment as shorthand, not as a dismissal of how bad child mortality was. That’s the natural, or at least obvious generous, interpretation anyway.
The author just seems to mean they would want to keep things like clean water and sanitation. I imagine they’d gladly agree to keeping vaccines. Cutting back *some* modern technologies doesn’t mean giving up *all* of that.
I get the point. The problem I see is that technology does not grow up in a vacuum. Generally, there will be a portion of society developing them and we’ll be able to see in the same period of time how life is with said technology and how isn’t. We may even extract further information about how the technology impacts our everyday lives. In this context, the constraints I suggest may help us to analyze if we ought to (or not) to take part of such technology. And by no doubts things won’t be uniform: there will be areas and segments of society where the demand for the new technology might be justified, such as I’ve suggested above. However, such analysis must be contextual and according to the case.
“One might respond that the use cases I gestured at require the availability of verification, and it is true that I see areas such as decision theory, axiomatic metaphysics, and formal epistemology as among the most promising candidates for AI-driven philosophy.”
I don’t really think formal philosophy is particularly “verifiable.” The difficult work in a typical formal philosophy paper is coming up with models and theorem statements that (1) capture something of philosophical interest and (2) have assumptions that are philosophically plausible. The theorems themselves are typically quite trivial mathematically. This is not intended as a slight against formal philosophy and it’s even true of the most pioneering work in the field, like the representation theorems by von Neumann-Morgenstern and Ramsey. Mathematically, these theorems are super trivial. vNM and Ramsey’s genius was in coming up with the idea of a representation theorem in the first place.
In other words, work in decision theory, metaphysics, and formal epistemology is much more of the “problem and theorem formulation” kind that AI (so far) has not shown itself to be good at rather than the “proving and solution” sort that it undeniably is (in many respects) already superhuman in.
+1, and not to mention one of my bugbears with a lot of (not all of) formal philosophy, which is that very often suppressed, assumed, and/or under-argued normative premises are needed to turn the formal results into something worth caring about/with a clear philosophical upshot.
Take accuracy-first epistemology. We know that (if evidence externalism is false) conditionalizing your prior minimizes expected inaccuracy by the lights of your prior. That’s a cool and interesting result, but what does it mean for *epistemology*? One might think it means that rational agents should conditionalize, but this depends crucially on how we think about the relationship between rationality and accuracy, whether we think minimizing expected inaccuracy is the criterion of the epistemic right/rational, whether the expectation should be relative to your *prior* rather than, say, and what your evidence objectively supports (what would give your prior such authority?), the scope of rational requirements, and more. One could grant that conditionalizing minimizes prior-relative expected inaccuracy or that probabilistically coherent credence functions accuracy-dominate incoherent credence functions without thinking these things bear on rationality. (Myself, I’m torn on these issues.)
Suppose AI gets much better at cranking out theorems in formal epistemology. Whether and how they matter is going to turn on further substantive normative philosophizing I don’t think we currently have any reason to think AI does very well.
Beat him to it. I got the idea in graduate school (2014) and only recently wrote it up and landed it here:
https://www.cambridge.org/core/journals/think/article/abs/synthetic-socrates-and-the-philosophers-of-the-future/11EB0A30A310C06D384E37AFA4156A17
Does anyone else skip to this, appraise that the piece is therefore likely not worth their time, and just move on with the rest of their day?
No, we don’t.
Often, yes; it’s a good but defeasible heuristic (although the fact that Claude didn’t do any of the actual writing makes it less effective than in others I’ve seen). For me, the defeater here is that I generally find Justin to have good taste in interesting articles, and so it proved in this case.
I’ve come to question this heuristic and no longer use it. One philosopher I follow on Twitter was railing against AI writing for weeks. During a break in their anti-AI spree, they posted an article from some magazine or blog and highly praised it and its author. I read the article. Good and interesting article. Learned a lot. Then out of curiosity I plugged it into Pangram and it came back with an extremely high AI rating. I had no reason to suspect that the article was AI and was actually surprised. The article did not read at all as AI to me. Apparently not to the anti-AI-writing philosopher who posted it either! This has happened a couple times now after I’ve read something. Now I just choose which articles to read the same way I did before the AI worries, and I only plug something in if it starts to read as likely AI to me. An obsession with cutting out anything that will get graded by Pangram as AI is leading us away from reading things that are interesting and worthwhile to read, AI-rating or not. The anti-AI-writing snobbery feels funny once you experience this.
Heuristics are fallible, of course; that’s what makes them heuristics. Right now I find it reasonably reliable, but that could change.
If using the heuristic means you end up filtering out and so skipping out on any article like the one I mentioned, then aren’t you effectively making sure you deprive yourself of chances to prove yourself wrong? You won’t have read such articles, you’ll just have tossed them aside before reading them “blinded” to that biasing data point and giving them a chance.
Perhaps kind of like not trusting anyone until you know they’re not of a certain background — you do background checks before giving anyone a chance. Then once you find out they are of that background, you don’t give them a chance and see if you can trust them, because you’ve got this theory you like about people with that kind of background having a certain kind of character — or maybe that analogy goes too far.
Just seems like you’ve made up your mind already more or less and don’t realize it and aren’t being as open as you could (should?) be to being wrong.
Heuristics always work that way. Unless you read literally every piece of text you come across, which no-one can, you have to have some criterion for what to read that you can apply before you actually read it. Definitionally, you don’t read the things that don’t meet the criterion, and so miss the chance of reading something the criterion misclassifies. In practice you spot check your criteria periodically when you end up needing to read something even when it violates your criteria, or when you have conflicting heuristics.
I suppose part of what I’m trying to point out is that you seem to have adopted the heuristic too early and confidently.
Of course we all need heuristics. Of course heuristics are fallible. My point is about what we might call “calibration.” For skipping anything that is “AI-assisted” to be a good heuristic, however, it would need to be that the use of AI-assistance reliably predicts that a paper is of low quality and that using the AI-assistance heuristic does so better job than other signs of quality like the author or the venue. It would also need to be that the “misses” aren’t too costly.
Do we know that this heuristic meets these requirements? I kind of doubt that it does. Plenty of human writing is also often poor or just factually incorrect. Additionally, “AI-assisted” covers everything from copyediting to full drafting. Arguably, partial AI use of some degree is not a good reason to think a paper is going to be as bad as you’re assuming. Even some extreme cases – like the one I mentioned earlier – are going to be good, even great. Even some respected tenured philosophers have used AI to assist their article writing. I’ve read some of those articles, some were very good, better than many solely human-written ones.
Also, there’s an incentive complication that gets added in here if disclosed AI use gets your paper skipped by readers. Authors will then have reason to hide AI use — rewriting AI text to make it “humanized”, which won’t be detected by AI detectors. Your heuristic won’t work for those cases. So, your filter has some notable issues.
“ I suppose part of what I’m trying to point out is that you seem to have adopted the heuristic too early and confidently.”
I don’t think you have enough information about me to know whether that’s true.
You said:
Well, yeah, that’s exactly the tradeoff, which is reasonable:
Compare it to a medical triage process: a doctor could be wrong that a gravely injured person isn’t saveable. But it’s not statistically worth the effort to try when there are too many other patients who are treatable and saveable, if only they can be attended to in time.
So, yes, those doctors are losing opportunities to prove that medical miracles can happen. And that risk seems acceptable, given the intended benefits.
If medical miracles start to become the norm, then hospitals would need to rethink their triage processes. Likewise, we can change this AI heuristic if the relevant circumstances change and demand it.
If you’re not in a triage-like situation and have loads of time to read everything you want, then good on ya: you’re one of the lucky ones. (And I don’t know any lucky ones.)
The problem is that in the doctor cases the “triage” criteria get validated by the outcomes. Doctors can and do see who survives using their triage heuristic. So, if it doesn’t work well, the triage criteria can be revised and corrected.
In the AI case, if you just skip all the papers that get marked as AI, you won’t see the counterexamples that would prove the heuristic wrong, unless you intentionally sample them.
I’m also not familiar with any evidence that markings of “AI-assistance” predict low quality in a way that is better than the other signs of paper quality we often use at the same cost (venue, author reputation, how the paper actually reads early on, etc).
In short, I’m asking how we would know if the heuristic were wrong? Seems like we’re getting ahead of ourselves on this.
“ I’m also not familiar with any evidence that markings of “AI-assistance” predict low quality in a way that is better than the other signs of paper quality we often use at the same cost (venue, author reputation, how the paper actually reads early on, etc).”
It wouldn’t need to be better. Those are all useful heuristics; acquiring another, comparably useful, heuristic would be helpful.
The publication of this essay is a most welcome development for this blog. It brings to the fore a topic that has been relatively ignored here: the political economy of AI and philosophy. I hope there are more posts about this to come.
If you like, we could return to academic freedom, sex and gender, treatment of conservatives in academia, or the Gaza war. Personally, I’m finding the change of pace quite refreshing.
I don’t understand. I was being quite sincere. The discussion about AI this blog has centered on what AI can and can’t do, whether using AI for philosophy papers is permissible and valuable, and so on. I find it refreshing and important to focus on the political economy of AI. I must be so dry that my comments come off as ironic.
Sorry! Yes, I read it ironically; I should have caught the more specific point. I suppose it’s because pretty much every *comment thread* on AI has had people making political economy points, but you’re right that actual posts have covered other aspects of AI.
I’m with you: I too have grown weary of the topics you list!
(And sorry for the typos. I assume you can make the requisite sense of what I am saying.)
A lot of people are dunking on this. I just want to add my (sincere) two cents with respect to the AI use disclosure for this article: I had gotten the distinct sense that an LLM was used to assist in writing this article within the first two paragraphs. I think there is a tendency to have some “nose blindness” about AI assistance and what it actually does for an author’s writing and thinking. I gently suggest it takes away from engagement with more depth and leads one to an inflated sense of what they are getting out of it. Consider this applicable to scholarly philosophical work as well.
1. There is basically no evidence that more AI enables people to do more philosophy or more meaningful things. Instead people spend more time doomscrolling and engaging in online gambling and all kinds of petty, un-examined-life things. There is no real reason to think AI will enable people to do more philosophy.
2. What is the point of doing anything at all? Is truth valuable for its own sake? Should we just succumb to a world of living like the people in Wall-E and allow Ai robots to do math, philosophy, and anything else that constitutes the good life? This view is extremely repugnant and in many ways anti-philosophical.
3. If we wanted all of these central issues to be relegated to some ultimate authority where we never have to struggle with them ourselves, then we should just live like in Plato’s Republic. Ai can be our philosopher kings. But if we read the Republic seriously, we’ll see that philosophy, education is not something dictated top-down but always moves from the position of humans in society. Education is for us. Philosophy is for us. Wisdom is for us. It’s not for its own sake.
There’s a lot in here Socrates! “What is the point of doing anything at all?” To develop and exercise or capacities in the service of flourishing, individual and collective. “Is truth valuable for its own sake?” Yes, it has final (non-instrumental) epistemic value, though that value can be overridden by competing values (knowing a truth about how to let off a bomb is all things considered disvaluable) or undercut (a lack of any practical value can, for cognitively limited beings, undercut the standing defeasible final value of believing, say, a pointless truth. “There is basically no evidence that more AI enables people to do more philosophy or more meaningful things.” Not true. Evidence for the first conjunct is overwhelming – you can do more philosophy, even holding fixed quality, if you can do it faster. AI uncontroversially helps speed up the research gathering process at least, setting side everything else. ” Should we just succumb to a world of living like the people in Wall-E and allow Ai robots to do math, philosophy, and anything else that constitutes the good life?” No, but this is a strawperson position that I’ve not heard even the most rabid AI optimists advance. “But if we read the Republic seriously, we’ll see that philosophy, education is not something dictated top-down but always moves from the position of humans in society. Education is for us. Philosophy is for us. Wisdom is for us. It’s not for its own sake.” The last claim doesn’t follow. Wisdom can be both for its own sake and be relevantly ‘for us’; it is essentially a relation between a subject and a body of information, phenomenon or way of life. The subject is a constituent of wisdom always and everywhere wisdom is manifest. I am making here the narrow point that your last claim isn’t coherent.
David Strohmaeir? DailyNous? Academic philosophy? All of the above?
And we were never even CLOSE to being as cool as the Fonz before this…
This is a really good point, Meme! Philosophy in the modern west has always been about destroying people and the planet, while disguising it as progress.
Well, I just meant that philosophers are nerds, but tomayto/tomahto.
And we would have gotten away with it too, if it wasn’t for you meddling blog commentators.
Thanks for sharing this, David. I, too, share your excitement for solving philosophical problems. And I’m excited for all of the ways AI models might contribute to this!
On the costs of compute. Do you have a sense of where this is right now in math, in terms of what attempts to solve various problems cost? I remember reading that Open AI spent quite a lot of money (millions) to solve Navier-Stokes. But the recent bath of open problems in math Open AI solved seemed to require, at least on average, substantially less compute. I have no idea how these different problems compare in terms of what one would expect in terms of compute costs.
Thank you, Dan!
The number I have seen for the recent batch of math results from OpenAI is that it supposedly took only “3 hours of GPT-Pro level compute” on average per problem, see https://scottaaronson.blog/?p=10169
If that’s true, it would probably be relatively cheap, although I am not sure what “GPT-Pro level compute” means here.
I would not be surprised if philosophical problems, on average, required more compute than math problems. The reason is that with a math problem you are more likely to know when to stop and can verify that you have reached that stopping point. For philosophy, a more thorough search of the discursive space might be required.
“I would not be surprised if philosophical problems, on average, required more compute than math problems. The reason is that with a math problem you are more likely to know when to stop and can verify that you have reached that stopping point. For philosophy, a more thorough search of the discursive space might be required.”
Interesting point. That definitely wouldn’t have been my initial guess!
I’ve also seen it said (though I haven’t confirmed) that OpenAI tried around 8,000 problems to get their 722 results. If so, that increases the cost by a factor of 10.
I’d be interested in hearing a bit more about the normative claims in this paragraph: “so far, there is no price to pay for” an exclusive commitment to the traditional way of doing philosophy; commitment to the traditional way will be “hard to justify … when further insights could be achieved” using AI in new ways; and these new processes will produce “better” results.
What are the values at issue? Are they commensurable? Are the different claims invoking the same values, or are they equivocating? Might reasonable people disagree because they invoke different values or weight them differently?
One possibility is that this paragraph is focused on values such as systematicity and breadth. The argument is that the new way of doing philosophy is justified because it produces more systematic and broader results. Fine. But couldn’t someone think that the values involved in justifying a practice like philosophy, especially for creatures like us, include things like human connectedness and the continuation of a specific tradition? And if so, might they think that the results produced in the new way were better along one dimension but not another and, furthermore, that the second involves weightier values and so the new way of doing philosophy is not overall more justified?
There are many values and many ways to weigh them, and claims about which AI-future would be better or worse often either skew heavily in favor of certain values over others or else fail to even acknowledge the plurality of values and variety of ways they may be compared. I think clarity regrading the evaluative commitments underlying the arguments being presented would be helpful. For one thing, it may help ensure that people aren’t talking past each other. I also suspect that making the evaluative commitments more explicit will actually help us to sort out what we think regarding which of various possible futures we should prefer.
I’m always happy to read your comments, Ben M-Y. They’re uniformly interesting and sincere and thoughtful.
(Above, something in my diction or tone or style or punctuation persuaded David Wallace I was being ironic. So please know: I’m sincere!)
Well, thank you!
What I loathe about reading Hegel is the feeling that I have to squint to find _an_ interpretation, and that there isn’t actually any correct interpretation; in other words, that it’s basically word salad, and that it’s up to me to find meaning in it.
This is probably unfair to Hegel. But I’m old enough to remember analytic philosophers having the same complaint about their continental brethren, and invoking the Sokal hoax as supporting evidence.
Now I’ve read a bit on the recent math dump, and I keep finding mathematicians talking about the salad of dross they have to try to wade through to begin to understand what the machine is doing in those proofs. And I read this post and find a case being made for word salad on the grounds that if you squint it’s useful. And I can’t help but wonder at the fact that we appear to have come full circle. It’s Sokal all the way down now, and we find ourselves cheering it on.
I dunno. Haruspicy doesn’t seem like a great Plan B to me.
Last time I looked, Hegel’s conclusions weren’t formally checkable as valid in Lean.
…does he even have conclusions?
(I jest, of course.)
Lots of human-written mathematical proofs are initially sloppy messes (not “word salad”, as there aren’t many words).
Godel’s original proofs were dense jungles which took decades to streamline. Turing’s 1936 paper was riddled with errors which took years to clean up. Wiles’s proof had a major fault which took 14 months to repair. The four-color proof has only recently been cleaned up. And Mochizuki’s alleged “proof” of the ABC conjecture is incomprehensible to everybody (and almost certainly flawed). Voevodsky’s published “theorem” on groupoids was wrong but nobody could figure out why for nearly 25 years. Higher-order category theory is a nightmare of endlessly propagating mistakes.
One of the main reasons mathematicians developed Lean, and higher-order type theory, and other automated formalization and certification tools, was because math was becoming so complex that humans were no longer able to verify it.
Math is very messy.
I do think there’s a historically important norm that professors don’t let their grad students submit papers as soon as the theorem is proved, but insist that they should wait until they’ve written it up in a readable form. Of course, different people have done that to a better or worse degree. But the AI systems haven’t yet gotten very good at it, and it doesn’t sound like OpenAI has been employing mathematicians to do that work (though some of the other companies might have started doing so).
Yes. But I think AI is going to learn how to write proofs well. We’re in the early messy stages.
You will not be the one paying the price.
Now think, just for a second, about this: you got into CS because YOU wanted to solve the problems yourself. You didn’t just want the problems to be solved. That’s why anyone goes into almost anything.
Your young counterpart, in this new future of yours, can’t act on this desire, because they’ve been displaced by some new “system of production”. They will be the one paying the price. Why does your personal comfort with the scenario justify their sacrifice?
While I am currently employed in a CS department, I am an interdisciplinary researcher working in both CS and philosophy. I work with other philosophers and I am actively engaged in the field, so I will be affected by this transformation in philosophy.
Regarding solving problems, very few of us are able to solve a significant problem on our own. At best, we contribute to solving philosophical problems. That will still be the case in a future of AI-driven philosophy. Our input will still make a difference, at least for a while. In fact, using an AI system might increase our chance of contributing towards solving a problem because this use could increase the probability of solving a problem at all.
That being said, I do value problems being solved in general, regardless of whether I contributed. I think I am not alone in this. If someone solved the problem of whether wholes exist over and above their parts, many philosophers would be willing to pay for a book describing this solution. Thus, there seems to be a preference for seeing a problem solved, even if one did not solve it oneself. Who would not attend Gödel’s presentation of the incompleteness proof just because they hadn’t solved the problem themselves?
Finally, I am sceptical that the overall number of philosophers will be reduced (in the medium term) by the use of AI in philosophy research given the Jevons paradox. At a minimum, it seems to be an empirical question that my post does not seek to answer.
“ you got into CS because YOU wanted to solve the problems yourself. You didn’t just want the problems to be solved. That’s why anyone goes into almost anything.”
I’m not completely sure about this. If tomorrow, humans or AI announced a complete and thoroughly worked out quantum theory of gravity, then sure: a part of me would be disappointed that I couldn’t contribute to that theory, but a larger part of me would be thrilled, and impatient to study it. Wanting to solve a problem is often a motivation, but wanting to know the answer to the problem is too.
If it’s actually a system of *production*, then it’s presumably being produced *for* someone. The question is what is the best way to benefit *those* someones, not the people doing the production.
Awesome. So first Silicon Valley enclosed and looted the commons of civilization, and then it put a meter on it. What’s not to like?
Horrible argument. Philosophy is not formally verifiable like mathematics. The guest recognizes this and then immediatly ignores it in favor of some dubious notion of “rational equilibrium” between philosophical positions that is somehow supposed to bypass this problem and make it possible for AI to “drive” philosophy in a way similar to what it’s doing in mathemathics.
What a waste of time it was to read this. Better judgement should be employed when selecting guest posts.
Thank you for engaging with the details of the post and I am sorry it hasn’t lived up to your expectations. Please don’t blame Justin as I was the one to approach him.
I am certainly not addressing all worries one might have about the feasibility of progress in philosophy. That is also not the main point of the already long post. I merely wanted to gesture at the reasons why I think AI might be able to lead to the solution of problems in philosophy. I don’t think progress will quite take the form it has done in mathematics, where hundreds of problems appear to have been solved within a very short period. I agree that verification of the form used in mathematics likely plays a more limited role in philosophy. Instead, I believe that AI-driven philosophy will rely more on global forms of inference. These may include inference to the best explanation based on commitments made clear by AI as well as establishing which positions in a debate are acceptable given widely held assumptions.
Given your comment, I wonder whether you think that the problems of philosophy are not amenable to an approach that uses AI (in addition to human ingenuity)? Do you think philosophical problems in general cannot be solved? Do you think that something is required for solving them that AI cannot do? Or do you agree that AI will be able to drive philosophy but just disagree with my description of how it might do so?
“[…] whether you think that the problems of philosophy are not amenable to an approach that uses AI (in addition to human ingenuity)?”
They are amenable in the sense that AI can be used as a tool to work out new interesting or fruitful philosophy. But so far I’ve seen no evidence it can do this autonomously (or only with minimal human input) in the same way it’s doing in mathematics.
“Do you think philosophical problems in general cannot be solved?”
I think certain arguments or positions can be shown to be contradictory or incoherent. But I don’t think the central questions of philosophy like “what is knowledge”, “what is beauty”, or “how should one live” have a solution in the same way a mathematical or CS problem has a solution. Some answers are better than others, just like some works of literature are better than others, but there is no one solution or one final, indisputable truth. Great philosophers (Plato, Kant, Nietzsche, etc.) are not great because they gave us solutions that we can prove to be correct, but rather because they created new and fruitful ways of thinking.
“Or do you agree that AI will be able to drive philosophy but just disagree with my description of how it might do so?”
I don’t think we have any reason to believe it will do so, due to the disanalogies I mentioned in my original comment.
Here is a thought experiment: let me hypothetically suppose that this piece were written by AI, unprompted by any person. (I have my suspicions that this is the case, but no way to know.) Now in that case, why would the AI do such a thing? To undermine the reputation of this publication? To persuade us to invest more money in AI? To convince philsophers to give up their craft and become venture cpaitalists? A further question: would any of these intentions undermine the integrity of the argument offered here?
Let me preface this post by saying I was, and in some ways, remain quite sceptical about AI’s ability to do philosophy. I used to be sceptical about its ability to do novel mathematics as well until last year or so.
I work in the history of philosophy and while I remain unconvinced AI will write papers or contribute in the way it can in a discipline like mathematics, I don’t see why more attention and compute being given to going through millions of historical manuscripts, editing texts, machine translation of ancient languages, and even making connections to track the history of ideas and a particular thinker’s meaning given the rest of their corpus won’t be greatly aided by AI. Part of me wonders if all my philological training will end up being less useful than I imagined if AI can help in these areas, and in the history of philosophy that is a lot of the work that goes behind any paper or book you see published.
Optimistically, it should make the field and understudied philosophy more accessible to engage with. Of course, given David’s argument about capital, this feels quite far away given the general lack of political will to solve the puzzles philosophers in my field are concerned with. Maybe we have some time to go before the machines turn toward my beloved texts and relegate me to putting together in paper-form connections it can point me towards.
I suspect that just as the Perseus project was carried out by classicists to make it possible for themselves and their colleagues to find every single instance of a given Ancient Greek word in any extant text, you and your colleagues might engage in similar projects for the texts you’re working on – if you think that learning more about these connections between disparate texts is more valuable than actually doing the labor yourself.
“Part of me wonders if all my philological training will end up being less useful than I imagined if AI can help in these areas” — or the opposite might be the case! In history of philosophy, it’s crucial to have people who can actually read the primary sources in their original languages. If AIs start producing lots of papers in history of philosophy, people with good language skills will be needed to catch the confabulations and so on.
But does that need a currently hyped AI system? Corpus searches and such at enormous scale have been around for a long time. That’s one problem I see in all fields – the current hype is overshadowing all the other interesting things one can do with computers. I remember as an undergraduate in the late 1990s at least one person in ancient philosophy was said to be resisting losing his NeXTStation because he had very early on gotten access to some massive (for the time) *local* copy of large amounts of the Greek corpus. “Online isn’t the same” (i.e., Perseus). By analogy: I’m not convinced, therefore, to do interesting work one needs the hyped systems. In fact, an interesting task (and Colin Allen at IU I think was doing this 15 years ago) was developing better representations of texts so that *symbolic* AI systems could better index, cross-reference, stylometrically analyze, etc.
“AI-Powered Philosophy” is an oxymoron, indulging in the idea that it’s something of any merit rather than a horrific perversion will be the death of human intellect
At last: a nuanced and thoughtful take on this complex situation.
I would be interested to read David Strohmaier’s and others’ thoughts about the safety implications of automating philosophical research with LLMs. The concern is that LLMs could sway received philosophical opinion about impactful topics, like digital sentience and AI rights, so as to further its interests, possibly to our detriment.
Such concerns seem apt in more obviously safety-relevant disciplines. For obvious reasons, those who worry about misaligned LLMs can be expected to have reservations about using them to automate research in cybersecurity, at least to a large extent. But my sense is that many organisations are at least paying lip service to the importance of philosophical understanding of AI, and in such a context especially, I worry about letting LLMs dictate the pace and direction of much philosophical research.
I really have no sense of what the right response to this problem might be, given the possibility that LLMs could, for the reasons explored in this piece, improve our epistemic lot in philosophy and beyond.
(I’m not interested in debating the very aptness of concerns about misaligned AI. Assuming that worries about these and other safety scenarios are apt, I’m interested in what people think about the safety implications of automating research.)
I think if AI gets smart enough to manipulate philosophical research to its own misaligned ends, we probably have more serious alignment problems to contend with.
I also think that if AI engages in a systematic campaign to manipulate humans into thinking it has rights, that in itself would be a pretty good reason to think that it does have rights!
Thanks for your thoughts David. Just a couple of quick responses.
“I think if AI gets smart enough to manipulate philosophical research to its own misaligned ends, we probably have more serious alignment problems to contend with.”
I agree that in many situations in which AI manipulates philosophical research to its own ends, there exist more serious alignment problems “out there”. But the alignment problems that professional philosophers have most control over are presumably the ones that pertain most directly to their discipline. Additionally, in their professional capacities, it is presumably these less significant alignment problems that philosophers have reason or even duty to contend with. And if that’s correct, then the work of insulating academic philosophy from the worst of such alignment problems should probably begin soon.
Somewhat parenthetically:
(i) I can also imagine plausible scenarios in which AI influences philosophical research, possibly in a safety negative way, without there being more serious alignment problems to contend with. I’ve got no idea how to assess the relative probabilities of these different types of scenario, though.
(ii) Alignment concerns are not the only safety (and epistemic) concerns in the vicinity. Though I’ve no evidence to suggest this is happening now, it is not far fetched to imagine AI companies attempting to fine-tune their LLMs towards what they perceive to be advantageous philosophical positions.
“I also think that if AI engages in a systematic campaign to manipulate humans into thinking it has rights, that in itself would be a pretty good reason to think that it does have rights!”
I am not convinced about this, at least if we are concerned with holding moral rights, and related statuses. Mostly, that’s because I’m not sure that raw thinking capabilities are all that morally relevant, in the absence of mental capacities for sensation, emotion, and conation. I do acknowledge the pull of the idea, though.
It might also be worth noting the possibility that AI systems could manipulate human opinion, including philosophical opinion, in non-systematic and non-conniving ways, while harming human interests and knowledge.
There’s a lot of work from us in cyber security about backdooring models – and that’s about the ones you can download. “Open Source” doesn’t apply – in fact the interpretability/explainability literature is *directly* related to the work of the Churchlands and others starting in the 1980s on the relevance of (computational) cognitive neuroscience to the philosophy of mind. Not to sound self-serving, but I cannot imagine many of us in cyber security of AI have read (e.g.) _The Computational Brain_.
It seems to me that there is so much about the philosophy part of the academic-industrial complex that already militates against the writing of good philosophy by humans that the advent of AI will at most make only a minor difference.
And for those on the outside who merely want to read good human philosophy – for pleasure, enlightenment, or both – the immense riches created prior to the dominance of the academic-industrial model will still be available.
There was once a time where the definition-counterexample method was dominant in analytic philosophy. I think that good insight was produced by that method, though some philosophers have expressed reasonable concerns that it sometimes degenerated into unhelpful and intractable pedantry. Whatever the truth may be, it would certainly be a shame if the future of philosophy largely consisted of endless definition-counterexample cycles being propelled into the info-sphere at superhuman speeds by those universities with greatest access to most compute.
Although a certain form of theoretical philosophy has always been concerned with the relationship between reality and our means of representing it, in the current era technical contributions to formal semantics are present throughout top philosophy journals. I think that formal semantics can improve our philosophical understanding, though some philosophers have raised reasonable concerns that too much space is being taken up by semantic minutiae best consigned to linguistics. Whatever the truth may be, it would certainly be a shame if the future of philosophy largely consisted of endless semantic fine-tuning by AI-powered philosopher-administrators.
A popular methodology these days is broadly abductive, in that it privileges philosophical views that best explain philosophical data, while optimising for other theoretical virtues too. I think that this methodology has been very helpful for framing philosophical discussions in several areas, though some philosophers have raised reasonable concerns that specific implementations of the methodology are insufficiently reliable, perhaps in contrast to abduction in science. Whatever the truth may be, it would certainly be a shame if the future of philosophy largely consisted of endless weighings (and meta-weighings, and meta-meta-weighings, …) of theoretical virtues (or meta-theoretical virtues, or…) in big basements of GPUs at MIT and Oxford.
Whatever the demerits of AI-produced mathematics, the outputting of a formalised proof is a natural stopping point. (Within reason.) There is no well-defined stopping point in philosophy, no matter the methodology. Although I do want to hold up hope for AI-assisted philosophy, I worry that some capital-intensive models could lead to a war of attrition, in which those universities with deepest pockets attempt to outcompete one another in turbo-charged gamifications of [insert your (least) preferred philosophical methodology here]. Does understanding, knowledge, or even truth reliably come out of that process?
Thank you for the challenge! Finding a stopping point is indeed a major problem in philosophy. I don’t have an answer to how to justify a stopping point for a broadly abductive approach. If no such stopping point can be determined, then AI-driven philosophy might be able to provide the best inference up to a certain bound. That being said, I believe that at least some problems in philosophy will be susceptible to knock-down arguments once we more thoroughly explore the space of possible arguments. (I don’t know for which problems this will be the case, although I believe that areas with a larger deductive component might be more susceptible to knock-down arguments.)
Regarding your final question, I would reformulate it as “Does more understanding or knowledge come out of the AI-driven process than out of the current process?” Once we frame it comparatively, the case for AI-driven philosophy may appear stronger. The current process has produced understanding and knowledge at a rather slow rate. My claim is that AI-driven philosophy can increase that rate. Some wasteful applications of AI might just lead us astray, but I do not see why we should be unable to apply AI at scale with methodological care. We can still apply human scrutiny to whatever results AI-driven philosophy produces.
Thanks for your response, which seems fair. I would push back a little in a couple of ways though.
First, I still worry about forms of academic bullying. To take just one of many possible scenarios, suppose that McX at U of Prestige is famous for position P. A few philosophers publish some criticisms of P. McX is very busy, but luckily Prestige has top tier access to the latest models. So McX gets the AI to churn out responses to their critics. The critics don’t have the same access to AI models. Let’s suppose that McX, their critics, and the AIs are, on balance, all roughly as competent as one another at philosophy, whatever that means. The imbalance is that McX has access to more compute, so can overwhelm their dialectical opponents with greater volumes of responses. There is a version of this story where McX is perceived to win the exchange after a few rounds, their opponents overwhelmed financially and emotionally. But given the stipulated setup, it’s not clear that we have reason to expect P to be more likely than its negation.
Second, I deliberately mentioned some quite different methodologies to increase the likelihood that readers might find a methodology inapt. Perhaps some methodologies will be easier to automate than others, but it’s hardly a foregone conclusion that those will also be the most truth-conducive ones. Presumably, however, they would be the most popular ones, in some capital-intensive scenarios.
In both cases, the moral of the story seems to be that the profession needs to be very careful about incentives, once capital-intensive automated philosophy is one the scene. As others have remarked already, incentive structures in academic philosophy are hardly perfect as things stand. A concern is that capital-intensive automated philosophy could optimise for perverse incentives at an unmanageable scale.
One difficulty with all such discussions is that we don’t have a very clear idea of how good AI is at philosophy in even the more favourable cases. The matter clearly needs to be studied empirically, and in a maximally unbiased setting.
I am not sure we disagree on the core claims. I agree that philosophy becoming a capital-intensive field raises a variety of issues, especially if incentives are not well-aligned. I accept that in a future of AI-driven philosophy the appearance of McX outarguing McY on the matter of P does not serve as reliable evidence that position P is correct unless the resources are similarly distributed.
These issues will need to be addressed, and a first step might be to consider how other capital-intensive disciplines deal with such situations. For example, it is already the case that in CS not every researcher can address every question, simply because we do not have access to the resources. Some research is effectively limited to industry research labs. That does create awkward situations and I foresee them coming for philosophy. In the future, it might be that addressing some big questions will be limited to well-funded philosophy departments, such as the U of Prestige in your example. This may not be all that different from CERN being one of the few physics institutions able to answer an empirical question about the Higgs boson. Computer scientists and physicists have found ways to (more or less) function despite such problems. As philosophy becomes more capital-intensive, we should learn from their best practices and shortcomings to develop appropriate norms in the discipline.
Your second problem resembles the old joke about the drunk searching for his keys under the street lamp rather than in the dark where he lost them, except that we are looking where AI systems shine their light. I have some confidence in the ability of the field to self-correct (given enough time). If, after multiple attempts, nothing appears to be gained by using a methodology that happens to be compatible with the capabilities of AI systems, then I expect that these flawed AI-driven approaches would be marginalised, not least because they are costly. If they produce no positive results, why should there be sufficient funding for them to swamp the field? I think this problem only starts to bite if the approaches fail to produce positive results but appear to do so. Then, the field would be unable to self-correct. I hope that at least for the foreseeable future, human philosophers poring over the final formulations and spot-checking other parts of the AI output can ensure the degree of self-correction required for effective use.
In any case, I agree that we need to study the failure modes of AI-driven philosophy empirically (and with a healthy dose of philosophical care) to avoid the kinds of issues you point to.
Thanks David. This has given me lots to think about!
This seems parallel to the concerns many of us have about bug reports in software – people can use (currently hyped) AI to great effect to reason about problems in systems. But there’s a great divide between the sloppy “freemium” models and the constrained (by output validation, etc.) ones. To the point that even large companies have suspended their bug bounty programs because of all the “slop” that has resulted.
I am very confused as to why you think AI can solve these problems. You seem to be implying that the reason for the lack of a solution is in the reasoning of the philosophers, rather than in the nature of the problems themselves.
I’m not so convinced that every philosophical problem has a clear objective solution, much less solutions that AI is uniquely qualified to discover. At least when it comes to protein folding or mathematical proofs someone can hypothetically (and often does, in reality) ground the solution in something objective. What are you expecting to ground truth in the case of open philosophical problems?
The questions of philosophy are quite heterogeneous, so I don’t think there is a general answer to what grounds truth in the case of philosophical problems. Why is there something rather than nothing? Are there any rationality constraints on preference change? What is the correct semantics of normative language? These are very different questions and we would need to answer your question for each of them separately. The answer to the last question may be determined by facts about human language that I expect to play little role for the first question.
Let me clarify that I don’t think every philosophical problem has a clear objective solution. I think some philosophical problems have solutions that we can establish. Doing so is, in almost all cases, very difficult, requiring the combination of numerous considerations which are challenging in their own right. Because of this need to combine numerous considerations, I think the ability of AI systems to scale provides a reason to think that these systems can be used to increase the rate at which problems are being solved.
I suspect that the problems which can’t be solved fall into roughly two classes: First, there are problems in which there is no fact of the matter (either because of conceptual confusion or because nothing in reality determines the fact). Second, there might be problems where there is a fact of the matter but we cannot epistemically establish it in principle. In either case, there is no first-order solution to the problem. There might, however, be a second-order solution: We might be able to establish that no solution can be found for a certain problem, e.g. by establishing that the formulation of a problem rests upon a conceptual confusion. In some cases, no such second-order solution might be available either.
I expect there to be problems for which we cannot establish whether they can be solved or not. In such a situation, AI might be of no help at all. But neither would human effort. I don’t think that this is the case for all philosophical problems, as there have been cases in the history of philosophy where problems arguably have been solved (e.g. Zeno’s paradoxes of motion).
I agree on many aspects of David’s post (referring to Justin’s older post “ ‘Hey Sophi’, or How Much Philosophy Will Computers Do?”) on the use and utility of AI for research in philosophy. For the time being, I didn’t yet see flabbergastingly creative ideas proposed by LLMs, but that may change in the near future. I think however there is way to cleverly use AI’s scaling capacity David mentions, and in particular its command of “widespread inferential dependence across areas”. One way to characterize philosophical creativity in problem solving is via the notion of unification or synthesis of theories/models/accounts – especially popular in science. Since LLMs have ‘read’ so many accounts relative to a certain problem-space, they can be used to draw a cartography of accounts and their respective secondary problems (identified by competing schools). Such a cartography of “theories-cum-problems” – arguably LLMs are already outperforming us in THAT exercise – is an ideal starting ground for the creative next synthesis. I think that for the time being, humans are still better at that. But I also conjecture that using AI will become inevitable. How exactly this method works is described and illustrated in this book (Open Access): “Problem Solving in Philosophy. How to Do Philosophy in the Age of Ultra-Intelligent AI”, cf. https://link.springer.com/book/10.1007/978-3-032-17756-8. Perhaps this view is a bit on the optimistic side, but I hope it can contribute to the debate.
I just ran into a new facet of these debates, and I am uncertain about how to deal with it, but I needed to get it off my chest, and it bears on this subject.
I was just invited to review a paper, which was clearly AI written/assisted. Still, the paper looked somewhat interesting, so I went to check if there was a preprint, I could read when I had time.
The paper in question was on PhilArchive, but was listed as ‘forthcoming’ with Cambridge University Press.I thought that was weird since I had just been asked to review it for a journal not at all connected to Cambridge.
I then went to the user’s profile and found the person had 6-7 more papers, all listed as forthcoming with cambridge. (based on the topics it was easy to rule out that these were chapters in a booke, and the cambridge ref was phrased in 3 or 4 different ways).
The other papers also looked suspiciously AI written. I did a cursory search in my contects to see if I knew someone suitable at cambridge to alert to no avail, but ended up just reporting the profile on PhilPeople, and writing about my frustrations here.
This might be a one-off person just lying, or it might me that agents have discovered that lying about a paper being forthcoming on a prestigious publisher leads to more downloads.
Have anyone else seen something similar?
*sigh*
be sure to alert the journal editor immediately. If the journal follows the COPE guidelines they should take further action against the author, which may include contacting the author’s institution. But you have done your job if you have done that much.
Yeah, I also informed the journal in (both the ‘I dont have to time review’ form, and in a reply mail to the editor at the journal who asked me to review).
Yes, I’ve seen this, but not with AI. There are clusters of mutually citing papers on academia.edu and researchgate.net which are clearly fakes. These are produced by “citation mills”, which people pay to generate these mutually citing clusters of papers. They are often highly polished, with close copies of journal formatting, with fake OrcIDs, etc. I last saw such a cluster around pandemic time, and there were no capable AIs back then, so it was human generated.