What will the academic humanities look like in twenty years?
That’s not intended as a trick question to which one replies, “nothing”—I’m not that pessimistic.
A lot can change in twenty years, and any predictions are risky, but nonetheless I want to think about the future of the academic humanities largely in relation to one thing: the development of large language models (LLMs) and related artificial intelligence tools.
As teachers, we see what the increased use of ChatGPT and its ilk (not to mention other aspects of technological situation, such as smart phones and the entertainment explosion*) are doing to our students; while that’s worrisome, it’s not my main focus here.**
With adequate prompting, today’s LLMs can produce passing papers for most kinds of writing assignments in most college courses. They can create for students better papers than many of them would be able to produce on their own. If you don’t believe this, you are living in a fantasy world.
Our students know what the LLMS can do, are using the technology a lot, and all signs point to increased use over time, as the technology improves, becomes more familiar to more people, and as younger students who grew up with it enter college.
Its use is not currently limited to students. It’s used widely in various professions and one might expect that to increase as well, as it becomes more technologically sophisticated, as students who used it regularly in school enter the workplace, and as social norms adopt to become more accepting of its use.
What happens after a decade or two of all of this?
- Everyone will have the capability to relatively easily produce large amounts of good (enough) writing.
- This means that producing large amounts of good (enough) writing is not a particularly good evaluative criteria.
Some academics have thought about this as it relates to the kinds of assignments they give their students.
But we should also be thinking about it in terms of how it relates to academic labor: training, hiring, tenure, and promotion.
If we come to accept 2, above, will this change professional norms in a way that reduces the pressure to publish so much? Will this reduced pressure mean fewer submissions to our severely overburdened and highly selective journals? Will quality, originality, influence, or other factors come to play a more significant role than they currently do? (Or is that underestimating what AI will be able to eventually do?) Will we move to a “slow philosophy” model? Will writing become less important in our assessment of academics, and if so, what will emerge in its place? What should we be doing now to influence these future changes?
Discussion—and further speculation—welcome.
* I use the term “entertainment explosion” to refer to the internet-facilitated proliferation of a practically infinite amount of instantly accessible and effective entertainments available to nearly everyone almost all of the time.
** Writing is a learning activity in which what is learned is much more than writing. One must think, evaluate their own thoughts and others, organize one’s ideas, figure out what is worth saying, figure out what to pay attention to, find out how to look for information, etc. These are valuable skills that are learned through doing, through practice. There are reasons to think the increased use of this technology this will be bad for students, as it relieves them of the work needed to practice and develop the aforementioned valuable skills—though perhaps alternative means of developing such skills will become widespread in education settings.
If AIs become able to produce academic philosophy papers that are indistinguishable from human academics, then there will probably be way fewer jobs that involve writing academic philosophy papers.
Agreed: If AIs become able to produce X that is indistinguishable from the X produced by a human Y, then there will probably be way fewer jobs that involve a human Y producing X.
In before someone says academia is immune to market incentives!
Agreed: if some X such that X is an AI becomes able to produce some (or other) Y that is identical, with respect to all Y-relevant properties P, to any Y produced by any Z such that Z is a human W, then the probability that there are less than @ jobs (the number of jobs in the actual world) involving Zs which are Ws is greater than .5.
Except in the case of philosophy, the situation isn’t quite the same. In the general case, “if a machine can make X, then fewer people will make X” is true because we get more Xs with less effort.
But in philosophy that can hardly be true—is there really value in philosophy papers merely viewed as a product? I don’t think so.
That doesn’t mean that the effect will not be the same, but the reason can’t be.
I don’t see how it would ever reduce pressure.
Instead, I imagine we’re headed to a Clarkesworld-style future (the magazine was overrun by garbage and nearly had to shut down).
I think the fact that a lot of philosophical writing could be created with AI is a good reason for professional philosophers not to use AI to create philosophy.
Already more philosophy exists than anyone could read. Currently professional philosophers can manage by picking particular research areas. But AI could easily make that infeasible.
This might be worth it if there was some obvious benefit. But there’s not, as far as I can see. A philosopher who submits AI created work might get a publication out of it, but if that were done large-scale the amount of literature will bloat, which will just cause problems.
Contrast with math here. If the AI is given these axioms and there rules, then we have at least some reason to think what it spits out is good mathematics (of course we know AI goes funny sometimes). See the four color theorem. But philosophical arguments don’t seem to be like mathematical proofs in this way.
I want to save my job as much as the next person, but I’m unconvinced by this “bloat” problem. I mean, let’s just assume, for the sake of argument, that I might gain some understanding/insight from a paper regardless of whether it is written by a human or an AI (that is a big assumption, but the comment doesn’t say the AI work is worse). Imagine, now, I want to research topic X, and there are two possible situations: 1) there are 100 good articles on X (only written by humans) or 2) there are 200 good articles on X (some by humans and some by AIs). Clearly, if we assume that “having read everything ever published on X” is a necessary condition for writing on X, I can see that situation 2 makes it harder for me to write on X. It also, I guess, makes it harder for me to say something new about X (because more of the logical space has been covered). Still, even if I face a problem, it’s not clear to me that this is a problem for the world, or even a problem about AI.
After all, some research areas are already like situation 2. Over 2,500years, we have a lot of commentaries written on Plato’s Meno; if we hold some “read everything ever published on X” principle, it is impossible to write on the Meno. I don’t think, though, that anyone thinks that this situation involves some kind of problematic “bloat”; if anything, it’s rather a good thing – if you want to think about the Meno, you have loads of resources to draw on. It would seem even weirder to imagine someone arguing a few hundred years ago that people should stop writing on the Meno because we would soon be in a situation of “bloat” where it would be impossibly difficult to add anything new to the literature. (Maybe people did argue this way – I could be missing something – but it seems weird).
Yes, but neither of you have mentioned the obvious: an AI that can be used to write good philosophy papers also likely can be used to competently synthesize philosophy papers. “Reading” will be as AI-powered as “writing” in the scenario you envision.
The issue with bloat is that it exacerbates extant problems with peer-review, among other pipeline problems (see the case of Clarkesworld, which very nearly died as a result of AI bloat).
If the solution to that is AI review, so that there’s a nice little circle of AI refereeing AI, then… I quit. I publish to participate in my research community. A word-prediction algorithm is not a member of my community, and if it replaces my community wholesale, then there’s no point. It’s bad enough already just reading all the student-generated AI bullshit.
First the AI came for reviewer #2, and I did nothing…
> If the AI is given these axioms and there rules, then we have at least some reason to think what it spits out is good mathematics
This is based on a naive view of mathematics. Just because something is a valid theorem provable from the axioms doesn’t mean it’s good mathematics.
Just like philosophers, mathematicians think that most of what matters is whether the ideas are interesting, novel, and help us think about other things we care about. If AI is able to come up with work that does that, it will likely be able to do so in both math and philosophy. If not, then it will still be useful for people working in both fields, but won’t replace humans.
I would tend to reformulate the question as:
Where should the pursuit of social utility and the skills valued in the job market align in the future?
Maybe an answer lies in recognizing that there will be an even greater premium on originality and work that is epistemically and socially significant —originality that will be judged relative to what AI technologies will enable for the average candidate.
Of course, as has been reminded numerous times, it might be the case that as certain abilities lose prestige, others will gain in status. Skills now considered ‘soft,’ such as interpersonal abilities or managing interpersonal relationships, may become increasingly valued. But what will those “human” skills in the fields of humanities or philosophy be like? Perhaps “enabling dialogue across barriers and disagreements”? Perhaps something else?
Another possibility is that we begin to resemble more and more like IT specialists (scholars-technicians), where the key abilities selected for involve mastering AI tools specifically designed for the humanities or philosophy.
A combination of all of the above seems the most likely outcome.
I sure hope that “producing large amounts of good (enough) writing” was never our standard to begin with. The writing should also have a point of view, a spark of a new idea, or some kind of new synthesis.
Currently, it does not look like AI is anywhere near replicating these. My presupposition is that there is a kind of creative spark that human philosophers bring to the table which is not, at least not in the medium term, imitable by AIs. I don’t think this is entirely unreasonable.
AIs appear to be very good information retrieval machines, but rather terrible at advancing on the information they possess. Thus, I can imagine AI becoming a new stage of peer review. Retrieving the standard objections from the literature is what it’s good at. Perhaps this could go something like the following.
You upload your PDF that develops your new idea and the AI takes the role of reviewer number 2, raising all the silly objections in the world. In the ideal case, this part of the peer review process takes the form of the human author entering into a dialogue with the AI (rather than trying to anticipate and meet all the objections in the manuscript).
A particular delightful upshot of this future would be that there would be no more “Objections and replies” sections (a person can dream). Perhaps the paper can be published with a link to a saved state of the AI that can rehearse the pre-publication conversation. You tell it your objection and it paraphrases the author’s response. It will also tell you if it’s a new objection. And if it is, perhaps there’s a short-cut to publishing it as a response in the same journal.
So, roughly, this system should reward creativity and insight over “producing good enough writing”. It is fairly utopian, and a lot of things would need to go right for this to be remotely feasible.
One of these things that must go right is that such a new stage of peer review should filter out out the AI-generated submissions. That seems possible; I don’t think it would take much for AI’s to make approximately accurate “novelty” judgments. And if I’m right that AI’s are bad at novelty, that should suffice. So, hopefully, the “AI’s writing for AI reviewers” part of the future can happen entirely away from human eyes.
We need to pay attention to the intersection of LLMs and social media, with its “content creators” and “influencers.” My guess, and it really is just a guess, is that someday the folks we will all be paying attention to are those who have figured out how to (live) feed LLMs the most creative prompts for eliciting genuinely interesting and novel responses from software that was designed to give statistically based summaries of what everyone else is saying. This will be a form of content creation, and if it becomes popular enough in a given demographic, such as an academic demographic, then these influencers will start receiving money from companies to endorse their products to other academics.
If AI reaches the stage where it can produce academic research that is indistinguishable from top-quality human research, I think the implications for the academic job market will rapidly become the least of our worries.
It’s the same for scholarship as it is any other trade: you either use new technology to enrich your process or you quickly become obsolete. Conserve your energy by leveraging AI’s strengths to complete tasks then reorient your practice to overcome its shortcomings. If it can write, let it write. If it can synthesize, let it synthesize. If it can teach, let it teach. So far, it can seemingly convey all the linguistic aspects of self-hood, yet remains disembodied. With any luck, on the 20 year timeline in question, vitality will return as the benchmark for relevant philosophical insight. Not what you write but how you live—this, AI cannot replace.
Haven’t read every comment, but I would note that there’s a difference between good writing and what Justin in the OP calls “good (enough) writing.”
Good writing is a matter of style, not only of substance, and it usually requires, IMHO, having a good ear for the language in which one is writing and also a willingness to revise (for style, not only for substance).
Not all great or important works of philosophy (or important works in other fields) are particularly well written, but the hallmarks of a good prose style — clarity, elegance, and so on — are fairly recognizable. Perhaps the importance of style in academic writing will increase as AI’s get better at substance. (Whether AI’s in 20 years will have become good writers, as opposed to the good enough writers they are now, is maybe an open question.)
Also, AI’s can’t do some things that human researchers and writers can do; an AI can’t walk into and use an archive, for instance.
Before considering what difference good (or good enough) AI will make, in any domain, we need to first think through what the intermediate stage might look like; namely, AI not remotely good enough by the lights of experts, but of sufficient quality to impress any non-expert holding the purse strings.
Current LLMs cannot answer the question, “How can a man, a boat, and a goat cross a river safely?” The right answer is to cross in one trip, with appropriate safety precautions. An LLM will give you a long answer that involves several trips across the river, some of them without the boat. The system is not reasoning. It is imitating the answer to a famous brainteaser, which includes a wolf and a cabbage, and producing nonsense.
In the field I teach (business ethics), current LLMs cannot competently answer even basic questions, such as, “Can you explain how Milton Friedman’s shareholder theory is different from R. Edward Freeman’s (2017) stakeholder theory?”
There is no reason to expect LLMs to get substantially better. The “hallucinations” may change, but they will persist. As a recent article by van Rooij et al. explains, computational complexity theory identifies intrinsic limits on the ability of machine learning to imitate human cognition.
https://irisvanrooijcogsci.com/2023/09/17/debunking-agi-inevitability-claims/
People who think LLMs are the future of writing are living in a fantasy world.
As the discussion in today’s other LLM thread shows, GPT-o1 can correctly answer the above variation of the boat crossing puzzle. That said, it still struggles with other straightforward questions that are variations on classic brainteasers. For example, it takes GPT-o1 over a minute to provide a convoluted answer to the straightforward question, “How can you walk one kilometer north, one kilometer west, one kilometer south, one kilometer east, and end up where you started?”
https://chatgpt.com/share/676328a2-afec-800e-b45d-bc9a11b46e76
These systems are not reasoning.
The question isn’t whether AI can produce papers indistinguishable from those written by a human, but whether it can advance the frontiers of human knowledge. If AI can’t do that, and yet it’s no different to what humans have been doing all these years, what on earth was ever the point of the latter? If, by contrast, AI CAN advance the frontiers of human knowledge, why not just move out of the way, and let it get on with it? (I’m aware, as I write this, that philosophical relativists will deny the coherence of ‘advancing the frontiers of human knowledge’, but there we are).