Saving Academics from Themselves

The received wisdom among a certain class of Bay Area/rationalist/AI-pilled thinkers is that universities are a hopelessly broken epistemic technology, especially now that AI is widespread. Some academics believe that they might just as well be replaced by recorded lectures¹ or that the university’s only value comes from networking or costly social signalling.

This essay does not speculate about blue-sky AI-powered epistemic institutions — you can use your imagination. But just remember that no less of a university skeptic than Scott Alexander has said that “the hardest problem for any social technology is getting users.” The cynic might say that the only factors preventing AI-powered neoversities from taking off are protectionism by incumbents and the status-signalling nature of an undergraduate degree. To which I — a political scientist who studies emerging technologies — say: great. You’ve described the process of political-economic contestation that determines whether and how new technologies are adopted. It is almost never smooth.

Public corporations are optimized for the next fiscal quarter, politicians for the next election in two, four, six years… but Harvard has a 50-year plan for its Allston campus and Oxford predates the Aztec Empire.² A recent NBER working paper by David Cutler and Edward Glaeser titled “How Have Universities Survived for Nearly a Millennium?” makes the case that universities will continue their strategy of “survival through reinvention.” The Oxford of today has very little in common with the institution that emerged in the 11th century, but it is still here, and still teaching.

In other words: The university isn’t going anywhere. At the same time, however, academics who ignore the impacts of AI do so at their own peril.

The reality is that academics are selected for being willing to suffer serious financial opportunity costs for the guarantee that, via tenure, nobody can tell them what to do. As a group, we’re going to reflexively resist any imposition of AI by either university administration or tech-corporate evangelists.

Speaking for my field: Unless academics can design and implement major reforms to our research practice in the next few years, there is a real risk that many social science professors will want to ban AI entirely. Soon we’ll be forcing each other to sign statements swearing that we didn’t use AI and AInvestigating not just every em-dash, but every semicolon and three-part list, too. To do so would make academics a sort of dockworkers union of knowledge, preventing the adoption of more efficient technology for doing research in order to preserve their current way of life.

Because, yes, the institutional arrangement of academic research is obviously flawed. Peer review is breaking as submission rates climb. The replication crisis remains a symptom of the underlying dysfunction that comes from pitting a hyper-competitive academic job market against the truth-seeking, communitarian ethos of science. And the incentives for public communication of individual studies are still toxic because a bad study going viral still counts for the tenure packet. AI can be used to address these problems in radical ways. Many specialized quantitative and computational social scientists are writing about how to do this; my own contribution focuses more narrowly on the institution of peer review.

More broadly, the best-case outcome is not that academics boycott AI wholesale, but that we adopt AI in a way that preserves both our autonomy and its expertise, which allows us to better pursue our foundational goal of generating knowledge and transmitting it to future generations. A university of Luddites can’t help anyone navigate a changing world. By thoughtfully adopting new technologies, universities can provide a temporal counterweight to short-run trends. Stewart Brand’s model of pace layers suggests that a healthy ecosystem is composed of layers that change at varying rates, balancing progress with stability. As AI accelerates so much of our society, we will become even more dependent on processes that operate slowly.

As other systems accelerate, the basic temporal loops of university life remain the same: the calendar is divided into semesters, the standard degree takes four years, the PhD takes six years, tenure another six to eight, and then a scientific career can last until the hearse pulls up to the faculty lounge. The research loop must also involve the consumption of research and its integration into some larger network of knowledge, even one that only exists in the heads of academics. As a result, universities have been the stewards of some of the extremely long-term projects.

Consider the datasets that underpin our understanding of democratic development, economic growth, or political opinion: the Varieties of Democracy project, the Penn World Tables, the American National Elections Survey. These exist because academics — usually with university or government funding — decided to start measuring the same thing, in the same way, for decades. Almost no corporation has the incentive structure to sustain that kind of commitment. Government agencies can do this kind of work, but as the current U.S. administration has demonstrated, there exists a temptation to cynically politicize formerly professional data collection efforts. And nothing is easier than to simply defund inconvenient research that citizens barely know exist.

These projects are just the more obvious manifestations of the long-run experiment that is the practice of science itself. When a researcher spends years designing an experiment, collecting data, revising their model of how the world works in light of what they find, the eventual PDF journal article is a relic of that experience, and an essential technology for communicating some of that knowledge to the rest of the academic community — but the experience itself is what produces expertise.

Now compare that to a provocative intervention by researcher Andy Hall, which claimed that Claude more or less one-shotted the design and execution of a published paper in a weekend as part of a proposal for a “100x Research Institution.”

Obviously, academics are rewarded for our output, and in the competitive marketplace, we’d each like to 100x that output and crush our competition. But if everyone 100x’s our output, the whole project becomes absurd. We’ll become deranged, suffering from AI psychosis at the institutional level. Perhaps we will one day achieve full Artificial Social Science (ASS), at which point AI will be doing all of the research production, evidence synthesis, and knowledge application. But we’re definitely not there yet. In the meantime, I’m worried that focusing only on research output means incorporating AI into social science in a half-ASSed way.

Our institutional incentives are already in tension with the stewardship of long-term projects. Academia is not exactly sane at the current scale and speed; indeed, this has been true for decades. In a famous letter to Science magazine in 1963, Bernard K. Forscher describes how science began to mistake the map for the territory:

Academics have long bemoaned the pressure to publish papers ever-more papers earlier and earlier in our careers. This trend is indeed worth bemoaning, and it threatens what is unique about academia.

Properly incorporated, AI can speed up science by removing the grunt work, just as our laptops replaced punchcard computers decades ago. But brickmaxxing is the worst way to incorporate AI into science. Now that we have more time, we don’t need to brickmaxx by churning out research papers. There are better ways to build up scientific institutions.

William Clark, a historian of early modern German academia, offers a useful definition of research: “Works of research usually provide a basis for further research and/or relate to other, related works in a complementary and supplementary manner. They add up to something positive.”

 Research is a three-way conversation between the past (the literature), the present (the world it studies) and the future (subsequent research). The core foci of academia are the many slow, expert conversations that require time to percolate through the minds of those involved. There’s no guarantee of positive consequences, but the justification for the model is the many positive consequences that have in fact emerged from it.

If basic research is an ongoing conversation, we need to have a structure in place to make sure that it doesn’t die out. One way is reading. I’m sorry to report that academics don’t really read very much. Reading isn’t counted towards our career advancement, including the all-important step of tenure. I’ve spoken to many academics who privately admit that they barely read anything because they can’t spare the time from writing papers!

Computational social scientist Duncan Watts noted this in a 2019 speech:

But now, with the help of AI, we can spare the time. If we’re going to bother to produce knowledge in a human-readable format, we need humans to actually be reading it.

We must also spend this time developing a healthier system for incorporating the knowledge that will continue to be produced — now at faster rates than ever.

Historically, the skill set to conduct social science research was sufficiently specialized that only people who were also carefully instructed in research design were able to produce it. Vibe-coding has changed that; vibe-coded social science is already here. And let me affirm that much of this work is far from “aligned.”

Richard Feynman famously said of science that "The first principle is that you must not fool yourself — and you are the easiest person to fool." AI-powered novices performing data science on complex social questions about which they are inevitably biased are doomed to fool themselves and potentially flood the world with bad science.

Any serious attempt to evaluate the weight of evidence for a given topic requires a certain depth of knowledge to be done well; indeed, knowledge synthesis is one of the most challenging open problems in social science methodology. AI enables self-directed empirical inquiry but does not provide the expertise on which academic social science rests. The rigor that goes into the constituent studies is wasted if the studies are combined in an unrigorous fashion.

Don Campbell, one of the founders of policy studies, envisioned the role of the social scientist as a “methodological servant” in a society in which citizens wanted to pursue their own epistemic goals but lacked the technical understanding to accomplish them. This is the direction academics should move toward.

Even the AGI-pilled epistemic revolutionary must account for some period of future time in which human expertise remains necessary. If AI continues to be able to execute more and more of the research process, we still need human experts with the breadth of knowledge to be able to check their work. Even if the long-term destination is fully autonomous AI research of the kind that need not be verified by humans, the capacity to navigate from the medium to the long term successfully, on our terms, will require oversight and feedback from human experts.

Most immediately, then, we need to check the work of vibe-coded social science analysis. 

We need trained social science methodologists weighing in on the data analysis and surveys being conducted by the newly-empowered class of the coding-literate and research-curious. In one innocuous example, an author replicated several entry-level mistakes in survey methodology when trying to estimate the prevalence of GLP-1 usage. To his credit, he realized that his results were preposterously wrong and then moved towards reasonable solutions to his entry-level errors. But this gets harder when the problems are intermediate or advanced, or when the results are merely dangerously rather than preposterously wrong. Survey methodology is a field in which people get PhDs, and one in which social scientists are still pushing the frontier.

There are many such methodological issues where curious amateurs would benefit from expert advice. Social scientists can meet the large pool of epistemologically curious non-academics where they are, to combine our refined expertise with their energy, curiosity and resources.³

The sex-ideology surveys run by the blogger and researcher Aella, for example, are an amazing resource, but there are serious limitations to the conclusions that can be drawn within her survey frame. She discusses this on an episode of Doomscroll, responding to a host praising her for a very influential piece mapping different fetishes according to political alignments: 

Very impressive response, but this shouldn’t be necessary; Aella is, famously, skilled at a more applied kind of rotational analysis. Collaboration with social science methodologists would yield major gains from comparative advantage and help unify distinct conversations being held by rationalists and academics.

Social scientists can improve our collective epistemics by correcting methodological errors emerging from self-directed inquiry, and ideally by setting up more public-facing forms of peer review that can accommodate the questions that our fellow citizens find directly important for their own decision-making or well-being, rather than only the theoretical questions of interest to academics.

Even if AI is capable of original research, human experts still serve an important coordinating function. Most knowledge is more valuable when it is public knowledge, when we have some sense that other people are paying attention to the same thing. We want to be part of the same conversation, not talking to ourselves.

Collaborations between academics and citizen researchers are only one medium through which this coordination can take place. There are others. AI is obviously good at writing, but it is still much worse at talking. By this I mean talking, using vocal cords to vibrate the air between your mouth and another person’s ear. This is a defensive way to put it, but in the short run, this remains an important comparative advantage of human academics.

AI-generated audiovisuals still aren’t nearly as fluent as AI-generated text, and will remain much more compute-hungry. Talking is central to the day-to-day work of academics as teachers, and we need to translate these skills in a way that will enable our fellow citizens to benefit. There is a huge popular demand to listen to people talk.

Academics have failed to participate adequately in the contemporary media environment. I understand their skepticism. The contemporary media environment is messy. Even worse, from the perspective of the wordcels who have self-selected into academia, our mass literary culture has been replaced by a neo-orality.

It’s easy to find fault with the contemporary platform-driven media environment; Lord knows I’ve done so. But in the short term, there’s no way around it: To have influence, you have to post more. And given the tastes of the mass audience, you have to post in the medium of the short-form video.

Academic expertise can be immediately applied if we deign to meet people on the platforms and media where they are. This means we’ll have to give up control of the conversation — which is to say, it entails an actual conversation, which by its nature cannot be controlled the way peer-reviewed PDF or even a 60 Minutes interview can. As I argue in my book, The YouTube Apparatus, the defining feature of social media platforms is that they are social; that is, that public audience feedback co-creates the media object consumed by future audiences. Credibility is constructed only partially by expertise; responsiveness to the audience is also crucial. Much more than broadcast media, social media enables a dialogue between creators and audiences.

The problem facing most YouTube creators is in fact too much responsiveness. Without institutional or financial independence, they become puppets, victims of audience capture. The fact that academics do have that independence, that we are also interfacing with ourselves and the academic literature, allows most of us to dialogue with the public without being captured by them. We can use some of our newly freed up time to listen and respond to the experiences of our fellow citizens as things get weird.

The world is changing quickly and in disorienting fashion. Someone is going to fill the demand for explanations, for sensemaking and clarification; academics have the chance to do so responsibly. The alternative is an intensification of the nihilistic mysticism that now passes for wisdom on TikTok and Twitch. The essential upside of the academic autonomy preserved within universities is that we can decide what to do, and what to value our fellow academics doing. We have the freedom to meet our fellow citizens where they are, knowledge-wise, as well as to spend more time directly observing what's going on in the new world.

Ultimately, the research imperative for universities is to produce new knowledge. Academics agree on this point. But if we want things to stay as they are, things will have to change. The people who built the University of Oxford were more concerned about the God of the Bible than the machine god. They weren’t thinking about how AI should replace the college fellows because doing so would allow them to maximize codex production. The medieval mind could not comprehend codexmaxxing (complimentary). But over the centuries, as the machine god assembled itself and the biblical God lost His grip on Western thought, Oxford and the university system survived through reinvention. Now we have to do it again.

Knowledge generation is the goal, but the old technology of writing linear natural language text is outdated. The PDF is a tidy knowledge brick, but AI allows us to design modern skyscrapers of knowledge, composed not of uniform bricks but rather a diversity of specialized building materials. The long-term solution for academia requires both materials engineering and an architectural revolution, if you’ll allow me to torture the metaphor.

Universities aren’t going to disappear overnight, even if they choose to try to ignore or ban AI. They are powerful institutions, interoperable with some of the core social functions of society and with alumni at the heights of essentially everything that matters outside of the Bay. But unless we get our act together, collectively and soon, I fear that reaction will set in, pitting forward-looking AI evangelists against a shrinking “remainder humanism” of academic intellectual labor.

As much as possible, these ideal future universities should engage directly with the public while preserving academic autonomy, the development of expertise, and the continuation of the research pipeline. And they should incorporate AI that enhances efficiency at every step of the core research loop. The standing of universities cannot last unless we can demonstrably retain an epistemic advantage and provide services that our fellow citizens value. The institutional structure that has lasted over 1,000 years can remain viable — at least over the time horizon I can imagine understanding — but only if it undergoes yet another round of survival through reinvention.