“This makes me want to quit my job and go live somewhere in the woods reading philosophy written by those imperfect, inefficient devices called humans.”

Interestingly, just a few hours ago, the journal did an about face and decided on a policy that will prohibit any future papers “substantially authored by AI”. They must presumably regret their decision to accept Goldstein’s paper.

I think this new policy is well-meaning but poorly thought out. It won’t do the thing they want it to do. Before I explain why, let me paste the new policy (this new policy was posted by Seth Lazar, one of the associate editors at PPA).

Here’s the new policy:

“Academic journals serve at least two functions: the promulgation of new knowledge, and identification and credentialing of talented researchers. There are many other means available to share new knowledge, but few for credentialing talented researchers, where “talented” means roughly: able, on their own or together with other researchers, to make and communicate significant progress on fundamental problems. Submitting AI-authored essays makes the task of identifying talented researchers harder.

In addition, while it is at this time possible for a researcher to use AI to author a paper that is itself high quality, it is disproportionately unlikely that they will do so, and much more likely that the paper will have all the surface appearances of sophisticated work, but ultimately be insubstantial or incoherent. This makes it harder to peer review, and makes the task of peer review in general more onerous, at a time when it is already under significant strain.

Those who enjoy and value the products of AI writing have many other places where they can publish such work, notably online. There may also be a case for creating journals of AI-authored research. So there are very few benefits to publishing AI-authored work in existing academic journals, and, given the above, significant costs.

Our policy is therefore to prohibit the submission of AI-authored content to Philosophy & Public Affairs. We define AI-authored content as papers that are substantially written by AI. We distinguish this from the use of AI for light copy-editing (which we in general define as reorganising or condensing written work without rephrasing). We also emphasise that authors are entitled (indeed encouraged) to use AI while conducting their research, in whatever ways they choose.

To implement this policy, we require authors to declare that they have not used AI to author their paper, and to explicitly describe how they have used AI in the process of conducting the research that led to the paper. We will use the leading AI detection software at the time to verify the truthfulness of the former declaration. We note that AI detection software, including updated models, may be used throughout the review process, up to and including after publication. If at any point we discover that the author’s declaration that AI was not used to author the paper’s content was false, the paper will be rejected or retracted, and the author will be banned from further submission to Philosophy & Public Affairs.”

TLDR: journals serve two purposes (1) promulgation of new knowledge and (2) identifying and credentialing talented researchers. If the journal accepts papers written largely by AI, then it would cease to be able to reliably identify talented researchers. Therefore, they should not accept papers written substantially by AI.

Here’s the reason why this is a bad policy. We need to distinguish two stages of producing a paper. There’s the research part and the writing part (I’m not assuming there’s no overlap here). This policy prohibits papers “substantially written” by AI. But what about the research part? The policy explicitly mentions this:

“We also emphasise that authors are entitled (indeed encouraged) to use AI while conducting their research, in whatever ways they choose.”

So presumably, an author could prompt AI in the following way as part of their research: “Hey Claude, here’s theory X (read the file), find counter-examples to points B and C. Illustrate them with examples. Then come up with the strongest counter-points. And then give responses to these. Write it out in outline form. Do not write a paper. I do not want to violate PPA’s journal policies. Outline form only Claude.”

Claude would comply with an outline of this research, which is the bulk of the intellectual work. Imagine it is of high quality (this is not what is at issue here). The author would then just write out this research in their own words. It would never be detected by Pangram or have the “watermarks” that will soon be made available by the AI companies. And this would be perfectly acceptable according to the new PPA policy. Here’s the problem. PPA would credential the person who submitted this paper as a talented philosopher, even though he or she is not (or need not be).

By making this distinction between writing and research, the new PPA policy fails to do what it was designed to do–maintain the function of the journal to identify and credential talented philosophers.

Of course, they could change their policy to prohibit AI assisted research (you can’t use AI to generate counter-examples, for instance). But since this research work with AI is undetectable, there would be little incentive for anyone to disclose this. This will just be an empty policy that will only end up harming the honest philosophers who faithfully follow the policy. Similarly, PPA could decide to call processes like “making an outline” a type of “writing” which violates their policy. The same point applies since they will never be able to detect this when the philosopher just writes out the ideas in his or her own words.

It’s better to simply have the journals publish the best articles possible regardless of how they were written. We need to solve the credentialing problem in some other way. Ironically, the more journals are used by institutions for credentialing, the more incentive there will be for people to use AI in writing papers (to get the best possible content out) and not tell anyone about it, especially as LLMs improve over time.

I want to make a prediction that may be surprising: the cheating problem is just temporary bump in the road. I can see, concretely, how it gets solved within a couple of years. The argument has four steps, and then a twist at the end.

Universities are not going anywhere

Start with some institutional humility — directed at the doomsayers, not the universities.

Universities are among the most durable institutions humanity has produced. The University of Bologna has been operating since 1088. Oxford has outlasted the printing press, the Reformation, the industrial revolution, radio, television, the calculator, and the internet — every one of which was declared, at the time, to be the end of traditional education. The pattern is always the same: a new technology arrives, commentators announce that it makes the university obsolete, and then the university absorbs the technology and expands.

And expand it certainly has. Higher education is not a medieval holdover limping into modernity. It has become more central to modern life over time, not less. Universities credential the professions, run the research enterprise which power modern society, anchor regional economies, and remain the primary institution through which societies transmit advanced knowledge across generations. An institution that has survived nine centuries of technological upheaval and grown more indispensable through each one is not going to be killed so easily, however impressive LLMs get.

So the question was never whether universities will adapt to AI. The question is how. And on the cheating front specifically, I think the shape of the answer is already visible.

The problem is about to get bigger, not smaller

Before the solution, though, we should be clear about the scale of the problem — because it is growing.

The demand for higher education is enormous and expanding, and much of that demand can only be met online. The numbers tell a clear story. According to federal NCES data, over 54% of American college students now take at least one course online, and roughly a quarter — over 26% — study exclusively online, never setting foot in a physical classroom. Fully in-person study is officially the minority experience: classroom-only enrollment accounts for just 46% of students. The share of exclusively online students jumped nearly ten percentage points from its pre-pandemic baseline to over 26% today. The trend is steeper in emerging economies. In addition, economies of scale which can lead to net revenues give institutions incentives to ramp up online learning.

This growth is good. This is how a working nurse in rural Arizona finishes his degree, how a first-generation student holds down a job while studying, and how millions of people who will never live near a campus get access to education at all. The university’s continued expansion will run substantially through online education.

So what does this mean for cheating? Well, the fastest-growing segment of higher education is precisely the segment with the least supervision. An online student’s entire academic life is the unsupervised zone. Every exam is a take-home exam; every essay is written wherever with whatever assistance. If the cheating problem is bad on campus, it is structurally worse online — and online is where education is headed. So the pressure to solve the supervision problem is not going to relent. It is going to intensify until we solve it.

Students who cheat are not villains

One more piece of ground-clearing before I propose my solution (or prediction). We need to be cautious about our moralizing tone.

We talk about “cheaters” as though the problem were an outbreak of bad character. This might lead some to think that the problem is a deficit in character and virtue with the younger generations. But step back and look at the situation we have placed students in. Grades are the currency of their futures: medical school admissions, law school rankings, first jobs, scholarships that determine whether they graduate with crushing debt or without it. The competition for those grades is intense, zero-sum in many contexts, and unforgiving of a single bad semester. And into this pressure cooker we have dropped a tool that produces good academic work instantly, invisibly, and for free — while the students at the next desk may well be using it.

Given those incentives, using AI on an assignment is simply the predictable response of ordinary people to the situation they are in. A student who declines to use AI while suspecting half the class is using it isn’t just being honest; she is unilaterally disarming in a battle. That is a heroic demand to make of a nineteen-year-old. A system whose integrity depends on lots of people being heroes is not a good system.

None of this means cheating is fine. Learning is real and one of the most important things we can do in our lives. The proverb that college cheaters only harm themselves in the long run is true, and they will pay for it at some point. But the students are not bad people. They are rational agents at least on short term time horizons responding to career pressures and near-zero costs of defection. They will respond to different incentives just as predictably. That is precisely why I find the situation hopeful rather than despairing: a problem rooted in incentives and enforcement is an engineering problem. You don’t fix it with integrity seminars and stern syllabus language. You fix it by changing the environment — which brings us to supervision.

The bottleneck is supervision (and it always has been)

Consider: cheating is not fundamentally an AI problem. It is a supervision problem, and it is an old one.

The reason we trust an in-class exam is not that students are more virtuous between 9:00 and 10:15 on Tuesdays. It is that someone is watching. The classroom is a supervised environment, and supervised environments produce trustworthy assessments. This has always been the basic architecture of academic integrity.

The limitation has always been bandwidth. A professor can supervise a room for an hour. A professor cannot supervise thirty students for the forty hours each of them spends on a term paper. So we historically made a bet: we supervised the small, high-stakes moments (exams) and left the large, formative work (essays, projects, problem sets) unsupervised, trusting that the difficulty of cheating made the bet reasonable. Copying a friend’s essay was risky; hiring a ghostwriter was expensive.

AI destroyed that bet. It made unsupervised cheating free and nearly undetectable. But notice what it did not do: it did not create a new kind of problem. The unsupervised zone was always the vulnerability; AI just found this vulnerability and blew up the whole thing.

Meanwhile — and this is the part people miss — our supervision technology has been improving too. We already have lockdown browsers. We already have remote proctoring that videos students at their computers and uses AI to flag suspicious behavior. For anything done at a desk, in front of a screen, the supervision problem is essentially solved right now. I use these tools in my classes.

The remaining gap is essential, however. We want students thinking about their projects in the library, on walks, over days and weeks. Deep intellectual work does not happen exclusively in front of a monitored webcam, and we shouldn’t want it to. The desk is covered but the world beyond the desk is not.

Extending supervision beyond the desk

In an earlier essay on this Substack, I argued that AI companies themselves should build physical testing centers — secure, proctored facilities where students could demonstrate their abilities in verified conditions, the way standardized testing has long worked. I still think that’s right, as far as it goes. Testing centers are part of the answer, especially for high-stakes credentialing.

But testing centers cannot truly solve the problem, for the same reason student work cannot just happen in the classroom. Not all assessment can be compressed into a proctored session. The most valuable academic work — the term paper, the research project, the senior thesis, or whatever we come up with next in an AI world (see the end) — happens over weeks, in libraries and cafés and dorm rooms, in the shower and on long walks. Students need to think and work outside testing centers; that is where the actual learning lives. A regime of assessment that only trusted what happened inside secure facilities would end up assessing only what fits inside them, and would kill every pedagogical form that doesn’t. The testing center, blue books and proctored exams solve the verification problem by shrinking the work. What we need is a way to adjust the verification method to fit our best pedagogy.

So the solution, stated plainly: extend the monitoring beyond the desk.

Within a few years, AI glasses will be commonplace — this is one of the safest predictions in consumer technology, given the trajectory Meta, Google, and others are already on. For sighted students, the university can provide (or certify) glasses that see what the student sees during designated assignment periods. An AI system monitors the feed and detects any use of AI assistance on the work — a second screen, a phone propped against a coffee cup, a friend’s laptop.

We do not have this exact technology today. But I see no principled obstacle to it. The components all exist: wearable cameras, real-time visual understanding, AI systems that classify activity. What’s missing is integration and deployment, which is to say, a couple of product cycles. The hard problem — machine perception good enough to recognize “this person is using ChatGPT to write their essay” — is precisely the kind of problem current AI is on track to solve, somewhat ironically using the same capabilities that created the cheating crisis in the first place.

Once this exists, the supervised zone expands to cover the entire assignment. A student can draft in the library, revise in a café, think on a bench — all within a verified environment. The take-home essay becomes as trustworthy as the in-class exam, without becoming an in-class exam. We get to keep the pedagogical forms that actually teach — sustained, self-paced, deep work — while restoring the integrity that made them meaningful.

Yes, I know: privacy

The knee-jerk reaction to this proposal is obvious. Continuous monitoring through a camera on your face is invasive. People are right to feel uneasy about this.

But let’s think about what we already accept, rather than reasoning from the shock of the new.

First: we already monitor students. A proctored exam room is surveillance — a human being watching you continuously for three hours, with the authority to end your academic career if you glance at the wrong piece of paper. Nobody considers this a civil liberties crisis. It is understood as a condition of certification: if you want the credential, you demonstrate the work under observation.

Second: the proposal is not perpetual monitoring. It is monitoring during a specified assessment window — the take-home exam period, the designated project hours — exactly as the exam room monitors you for a specified window. Outside those windows, the glasses come off, or revert to being ordinary glasses.

Third: ambient monitoring is already arriving for reasons that have nothing to do with universities. People will be walking around wearing always-on cameras because they bought them for fun or whatever. The world in which AI glasses observe daily life is coming regardless of what universities do. The question is not whether that world arrives but whether institutions build honest, accountable, purpose-limited monitoring within it — with real assurances about data handling, retention, and access — or whether we get only the commercial free-for-all version.

Now, the uncomfortable part, which I don’t see how to evade: during the assessment window, I don’t think students can be allowed to turn the monitoring off. An off switch is a cheating switch. Ten unmonitored minutes is an essay’s worth of ChatGPT. The integrity of the whole system depends on the window being genuinely covered, just as the integrity of an exam depends on the proctor not taking a nap.

That is a real cost. There are trade-offs here. But consider what we give up if we don’t go this route: the possibility of learning itself. An assessment system nobody can trust certifies nothing, teaches nothing, and eventually collapses into credential theater. If the price of preserving genuine education — the sustained, unassisted development of a human mind — is bounded, purpose-limited, transparent monitoring during declared assessment periods, that is a price worth examining seriously rather than dismissing on reflex. We accepted exactly this bargain in the exam room centuries ago. This is the same bargain, resized for a world where the exam room moves.

The real problem is harder

Here is the twist. Cheating is the problem everyone is talking about, and it is the easier problem. It is an engineering problem with an institutional deployment problem attached, and I’ve just sketched how it dissolves.

The harder problem is this: even with perfect enforcement, we would only be perfectly enforcing the old assignments. And the old assignments were designed for a world that no longer exists. What precise skills should a student actually develop in an age when AI can draft, summarize, translate, and compute? Which forms of unassisted cognition are crucial for a good human life and a functioning citizen, and which were always just proxies we can now retire? What should the essay be for, once its production can be automated? I have argued that the skills you learn in humanities, like careful reading, writing, interpreting, taking other people’s perspectives and understanding context will be crucial in an AI world. But this doesn’t mean that writing a take-home essay is the best way to hone these skills, even if we’ve solved the cheating problem.

Curriculum development and not cheating, is the generational challenge for the university. The surveillance essay was the easy one. The curriculum essay comes next.

In 1957, Barthes opened his Mythologies with an essay on professional wrestling. His argument was that wrestling isn’t a degraded sport. It is a different genre entirely: a “spectacle of excess” in which every villain is villainous to the point of caricature, and the audience’s pleasure comes in part from the fulfillment of a perfect narrative. Boxing, Barthes said, is a story constructed before the spectator’s eyes; wrestling is a sequence of tableaux, each one an exaggerated image of suffering, injustice, or triumph.

Erling Haaland leads Norway’s Viking Row after they eliminated Brazil from the World Cup.

In my view, the line between wrestling and “real” sport was never as clean as Barthes thought. The World Cup, the most-watched event in the world, is shot through with wrestling’s DNA. And once you see why, you’ll notice it’s not a corruption of the game. It’s just what makes sense given the structure of the sport.

The theater is already inside the game

Start with the obvious case: diving. A player feels the whisper of contact, and suddenly we’re watching an audition for a death scene in a regional opera company. The clutched face (when the contact was on the shin). The triple barrel roll. The agonized writhing that resolves itself, miraculously, the moment the referee produces a yellow card for the opponent.

Diving is usually analyzed as cheating — an attempt to extract a foul the player didn’t earn. That’s true but there’s a lot more to it. Notice that the performance is wildly in excess of what deception requires—especially since the referees, now equipped with VAR, are professionals who already expect this kind of behavior like a seasoned parent who sees their kid’s tantrum coming a mile a way. The rolling, the anguished appeals to the crowd, the pointing, the outrage — this is not aimed at the referee at all. At a subconscious level and a result of structural forces, it is aimed at us. It is, in Barthes’s exact sense, a tableau: an exaggerated, instantly legible image of Injustice Done To Me.

Or take the antics we’ve come to associate with Argentine soccer in particular — the dark arts, the viveza criolla, the goalkeeper mind games during penalty shootouts, the trophy-celebration provocations, the theatrical contempt for opponents that made Argentina the team half the world loves and half the world loves to hate. The spoiled brat, the provocateur, the martyr, the saint. These are not personalities. They are characters. Wrestling has a word for the villain the crowd pays to boo: the heel. And it has a word for the noble hero: the face. Every great soccer nation, every great rivalry, is stocked with heels and faces. The players know this. Sometimes consciously, often not.

Consider what happened to Erling Haaland this summer. Norway returned to the World Cup after twenty-eight years away, and Haaland — already the most fearsome scorer of his generation — became the tournament's breakout star, burying seven goals, toppling Brazil in the round of sixteen, and dragging his country to its first quarterfinal before England ended the run. That much is athletic fact. But watch how instantly, how hungrily, the watching world fitted him into an archetype. Within days he was not a striker; he was a Viking. The long hair, the towering frame, the northern calm — commentators reached for longships and sagas as if on cue, and Haaland, grinning, obliged them, celebrating on the pitch at the helm of an actual Viking rowboat. And notice the specific character we cast him as: not the berserker but the noble warrior — mighty in battle, gentle in victory, humble in speech, a figure out of some idealized Norse past who fights ferociously and then waves sweetly to the crowd. Nobody scripted this. The archetype was sitting there in the culture's repertory, a thousand years old, waiting for a body to fill it — and the moment a suitable body appeared, audience, media, and player converged on the casting within a single news cycle. We carry the stories with us into the stadium and audition the athletes for the parts.

And the World Cup supercharges this casting process in a way no club competition can, because its heroes are national heroes — and we know it, and the players know it, and everyone knows that everyone knows it. Every nation arrives at the tournament with a repertory company already installed in its citizens' heads, put there in grammar school. Long before a child ever watches a World Cup, she has been taught her country's pantheon: the liberators, the martyrs, the founding generals, the honorable admirals who chose principle over survival. Classical folklore in the US frames Abraham Lincoln as a titan of both muscle and fairness, tales insisted he could swing an axe with enough force to shatter thousands of heavy logs in a single day, channeling that same relentless determination to walk miles in the dark just to return a stranger's forgotten pennies. In Peru we learned about Miguel Grau, the "Gentleman of the Seas," who rescued the sailors of the enemy ships he sank — the noble warrior archetype installed by the state itself, complete with moral instruction. Every schoolchild on Earth receives some local version of this curriculum. So when the national team takes the field, the players don't step into a neutral athletic contest; they step into slots that primary education spent years carving. The striker who carries the team is auditioning, whether he likes it or not, for a seat beside the founding heroes — and the fans, whose sense of national belonging was built on those grammar-school stories, experience his goals as chapters in the same book. This is why World Cup patriotism feels so much deeper than sports fandom, and why the tournament's dramas feel almost pre-written. In a sense they are. The scripts were drafted in classrooms, decades before kickoff, one flag-draped history lesson at a time.

Goffman on the pitch

Erving Goffman argued in The Presentation of Self in Everyday Life that all social interaction is dramaturgical: we are always performing a character, managing an impression, playing to a house. Think about what it must feel like to play soccer in front of billions of people. To expect athletes not to perform under those conditions is to expect them to be the only humans in history exempt from Goffman’s law.

I think there’s good reason to think that players act out scripts. A further question is why the scripts converge on the same archetypes wrestling discovered a century ago. The injured innocent. The wronged hero. The arrogant villain who must be humbled. The honorable warrior who applauds his opponent. Why these roles, over and over, across sports and cultures?

The game theory of drama

Professional sport is, economically speaking, an attention market. Revenue flows from broadcast rights, sponsorship, and engagement, and all three are functions of how many people watch and how intensely they care. This pattern is exacerbated as team owners and tournaments organizers adopt the modern capitalist credo of squeezing as much monetary value from all things in heaven and earth.

Now ask: what do people actually watch for? The standard answer — athletic excellence, uncertainty of outcome — is at best half the story. A 0–0 draw between two tactically perfect teams is excellent and uncertain. But the average audience member famously find it unwatchable. What audiences reliably pay for in dollars and attention is narrative: stakes, characters, grievance, redemption, villainy, comeuppance. The same things that fill theaters, and the same things Vince McMahon industrialized.

So consider the incentive gradient facing an individual player. Pure athletic excellence is rewarded, yes. But excellence plus legible character is rewarded more. The player who is merely great gets a contract; the player who is great and a story — a heel, a face, a tragic figure, a redemption arc — gets the endorsements, the documentary, the transfer fee inflated by shirt sales. Nobody has to consciously decide to become a character. The market performs the selection. Players who happen to produce drama get amplified, imitated, and enriched; players who don’t, don’t. Run that selection process over generations and the sport evolves toward spectacle, the way a species evolves toward camouflage, with no individual organism intending anything.

This is why I think the wrestling-ification of sport is a type of evolved rational equilibrium. Everyone’s incentives point the same direction. Leagues want engagement. Broadcasters want storylines — watch any pre-match package and notice that it is pure wrestling promo: grievance montage, slow-motion stare, ominous music. Players want attention, which converts to money. And the audience — this is the part we don’t want to hear — wants to be worked. We say we hate diving, and we watch the replay eleven times. We say the Argentine antics disgrace the game, and Argentina’s matches draw the largest audiences on the planet. Our revealed preferences are Barthesian even when our stated preferences are Corinthian. When people complain about Argentina, it’s like when people say Darth Vader is the bad guy— He is the bad guy, but Star Wars would be nothing without him.

There’s even an equilibrium role for the honorable player. In wrestling, the face only works because the heel exists; virtue is legible only against vice. The player famous for sportsmanship — who refuses to dive, who owns up to the handball, who consoles the opponent — is not outside the drama. He is cast in it, and rewarded for the casting. Playing the honorable person is still playing a person. I don’t say this cynically: the honor may be perfectly sincere. But sincerity and performance are not opposites — that’s Goffman’s deepest lesson. You can genuinely be the character you are also, profitably, performing.

Kayfabe without a script

Wrestling insiders call the maintained fiction kayfabe — the collective agreement to treat the staged as real. What sport has developed is something stranger and, I think, more philosophically interesting: kayfabe without an author. Nobody writes the scripts. There is no booker in the back deciding that Argentina plays the heel this cycle. The scripts emerge from the interaction of incentives, archetypes deep enough in human storytelling that Aristotle catalogued them, and thousands of individual choices — some cynical, most subconscious — to lean into the role the moment offers.

The outcomes remain uncontested, which is why world cup soccer is still a sport. But nearly everything around the outcomes — the characters, the grievances, the injustice tableaux, the redemption arcs — is spectacle that has evolved because we reward it. Sport is improv theater with real stakes: the ending is unwritten, but everyone on stage knows, in some sense, what kind of scene they’re in.

A word about our moralizing

All of which suggests we should relax our grip on a favorite pastime of the modern viewer: moralizing about the players. The dive is denounced as a character flaw, the taunt as a failure of upbringing, the tantrum as evidence that the man is, deep down, a bad person. We audit twenty-five-year-olds for virtue as though they were candidates for a judgeship rather than performers in the largest entertainment product ever built.

Some perspective helps. These are millionaires — many of them millionaires many times over — who have spent their lives inside an attention economy that selected them, from childhood academies onward, precisely for competitive fire, self-belief bordering on delusion, and comfort in front of a crowd. Of course they have egos; the ego is part of the job qualification, the way a surgeon’s steadiness is. And of course they’re into showmanship; showmanship is what the checks are for. Expecting a World Cup star to be free of vanity and theatrics is like expecting a heavyweight to be free of aggression, or an opera tenor to be understated. We hired them for the excess.

None of this makes the players saints, and it doesn’t make moral judgment illegitimate. But most of what fills the post-match discourse isn’t ethics; it’s the audience playing its assigned role in the drama. Booing the heel, canonizing the face, howling for justice at the injustice tableau. Wrestling crowds do all of this too. They just know they’re doing it.

So the next time a forward goes down as if sniped from the upper deck, resist the urge to see only a cheat. You are watching something older and odder: a performer who has correctly read the market, an audience that has paid for exactly this, and an incentive structure quietly authoring a script that no one wrote and everyone follows. The wrestling fans figured out the secret decades ago and have been enjoying themselves honestly ever since.

The rest of us are still pretending we came for the sport.

I grew up in Peru until I was ten, and in Peru the World Cup was not a sporting event. It was closer to a liturgical season. The streets emptied for matches. Grown men wept openly and no one thought less of them. The national team's fortunes were discussed with the gravity other cultures reserve for questions of war and salvation, and the tournament itself hovered over childhood as something sacred — the one month every four years when the world revealed what it was really about. Whatever I believe about the game today passed first through the eyes of a boy for whom it was simply, unquestionably, pure.

And yet. When I think about that ten-year-old in Peru, I find that what made those afternoons sacred was never ninety minutes of tactically optimal soccer. It was the weeping men, the villains, the injustices, the impossible redemptions — the drama, all the way down. Maybe the purity I’m afraid of losing was never purity of sport. It was purity of story.

On June 29, the Department of Education issued the final rule implementing the earnings accountability provisions of last year’s Working Families Tax Cuts Act. The mechanics are simple and serious. Every undergraduate program in the country must now demonstrate that its graduates out-earn the typical high school diploma holder; graduate programs must show their completers out-earn the typical bachelor’s degree holder. A program that fails in two out of three years loses access to federal student loans. Fail longer, and an institution’s low-earning programs can lose Title IV eligibility altogether, Pell Grants included.

Some people are saying that this rule is the beginning of the end for the “useless” majors.

Is the test a good idea?

Federal student loans are underwritten by taxpayers and repaid out of the futures of eighteen-year-olds. A program that leaves its graduates earning less than they would have earned with no degree at all is extracting tuition without delivering financial value and leaving someone else holding the bag. Clearing them out is not an attack on higher education; it is a defense of it. Every fraudulent program that collapses under this rule makes the sentence “college isn’t worth it” a little less plausible.

There is also a deeper reason to welcome the test, one that friends of the humanities have been too timid to voice. For decades, defenders of the liberal arts have retreated to the position that these degrees are valuable despite their economics — enriching the soul or adding intrinsic value. This was always an unforced surrender, because the economic case was actually good, though rarely made. A federal rule that forces the earnings data into public view is an opportunity, not a threat. Bring on the numbers.

Before I start, I should respond to a criticism I get regularly. Whenever I talk about the practical value of a liberal arts education, someone objects that the mere mention of money, for example, taints the enterprise — that to defend philosophy in dollars is already to have lost, to have accepted the enemy's terms. I find this objection revealing, and not in the way its makers intend. It rests on an inference the humanities are supposed to teach us to catch: that because something has intrinsic value, it cannot also have instrumental value, or that asserting the one somehow cancels the other. That is just a mistake — a false dilemma, and an elementary one. Goods are not exclusive in this way. A cathedral can be both sacred and structurally sound; nobody thinks the engineer's report insults the architecture. I believe that the liberal arts have intrinsic worth — that learning to read carefully, to doubt well, to follow an argument where it leads, is constitutive of a good human life and not merely a means to one. Nothing about that belief prevents me from also pointing out that these skills have value in the marketplace. Both claims are true and it would be surprising if they weren’t both true. Holding them together requires no compromise at all; it requires only the capacity to keep two values in view at once, which is — I cannot resist saying — exactly the kind of thinking a liberal arts education is for. The people who insist we must choose are demonstrating the deficit they claim to be guarding against.

The bar, and who clears it

So let us apply the test the way the rule actually runs it: on the earnings of young graduates, compared against young workers with only a high school diploma. The National Center for Education Statistics puts the median earnings of 25- to 34-year-olds working full time, year round, whose highest credential is a high school diploma, at $41,800 (as of 2022). That is the bar.

Now the early-career medians from the New York Fed’s latest data (2024 earnings, graduates aged 22 to 27, bachelor’s degree only). Mathematics: $70,000. Physics: $67,000. Philosophy: $52,000. Political science: $52,000. Communications: $52,000. Journalism: $49,000. English and foreign languages: $48,000. History: $47,500. Art history and the catch-all “Liberal Arts” category: $45,000. Every one of them clears the bar — even at the age when these graduates are at their lowest relative earning power, even counting only those who stopped at the bachelor’s. The margins range from comfortable to modest.

And the professional majors everyone assumes are the safe harbor? Most pass too — but not all, and not always by more. Elementary education graduates start at $45,000, dead even with art history. Treatment therapy starts at $50,000, below philosophy. Theology squeaks past the bar at $41,600. And pharmacy — a feeder into one of the most reliably remunerative professions in the country — posts an early-career median of $40,000, the lowest figure anywhere on the Fed’s table and, remarkably, below the high-school benchmark. Hold that thought; we will come back to it, because it is the key to what this rule gets wrong.

The point for now is simpler. On the test as designed — early earnings against a high school diploma — liberal arts do just fine. No problem. But we should aim for a good test, not just one that we pass.

The employment numbers run the wrong way for the critics

Earnings are only half the story; you also have to have a job. Here the 2024 data delivers its most uncomfortable finding. The unemployment rate for history majors was 4.3 percent — essentially identical to the 4.2 percent average across all majors. Communications came in at 3.9 percent, foreign languages at a remarkable 1.6, liberal arts at 3.8, philosophy at 5.1. Meanwhile computer science stood at 7.0 percent and computer engineering at 7.8 — among the highest of any major the Fed tracks. In 2024, the graduates most likely to be out of work were the ones who followed the stereotypical advice to find a “useful” major.

This is not a fluke to gloat over; tech labor markets are cyclical (remember a few years ago when coding academies and bootcamps were all the rage, and now they are all but collapsing). But it reveals the structure of risk. The case for the “practical” major was always a bet that today’s in-demand skill would still be in demand in twenty years. Narrow technical training is a concentrated position; the general capacities a liberal arts degree builds — reading closely, writing clearly, weighing evidence, exercising judgment under uncertainty — are always important and even more so in the age of AI. In the first years of the AI era, with entry-level technical work being absorbed by the very systems those graduates were trained to build, the concentrated position has never looked shakier.

Passing at 25 is the least of it: the case for looking further out

Clearing the bar in early career, though, is the weakest version of the liberal arts’ economic case, because early career is precisely when these degrees are least themselves. A vocational credential delivers its value immediately or never: the skill is specific, the market prices it on day one, and the trajectory from there is flat. A liberal arts degree works differently. Its skills are general and flexible — they transfer across industries, survive technological churn, and compound as graduates move into roles where human judgment, communication, and careful analysis are crucial.

(Here are some examples of people with long and successful careers who started off as humanities majors: Stewart Butterfield founded Flickr and then Slack earned a philosophy degree and so did Reid Hoffman, founder of LinkedIn. Susan Wojcicki who ran YouTube had a degree in history and literature. Ken Chenault, who led American Express for two decades, was a history major and so were Sam Palmisano who ran IBM and Brian Moynihan who heads Bank of America. Michael Eisner, who ran Disney for many years as an English major as was Hank Paulson, former CEO of Goldman Sachs.)

The monetary value of liberal arts degrees often takes time to ripen. Any honest earnings test should therefore look not just at the early snapshot but at the whole picture and in particular, the slope.

The slope is where the liberal arts shine. Between early and mid-career, art history graduates more than double their median wage, from $45,000 to $91,000. Political science rises 92 percent, to $100,000 — tying mathematics and industrial engineering to the dollar. History gains 68 percent, journalism 78. Against these, computer science grows 38 percent and nursing 24. By mid-career, philosophy’s $80,000 median matches criminal justice exactly and clears health services, nutrition sciences, and every education major by a wide margin; art history out-earns chemistry and biology. If mid-career earnings measured a major’s seriousness, the ritual scorn directed at the humanities — and never at nursing, teaching, or the lab sciences — would be exposed as what it always was: a cultural prejudice.

And even the mid-career figures understate the case, because of who is missing from them. The Fed’s wage columns describe workers holding a bachelor’s degree only. But 57 percent of philosophy majors go on to graduate degrees — along with 53 percent of political science majors and 51 percent of history majors, versus a third of computer science majors and a fifth of marketing majors. For philosophy, the widely quoted figures describe a minority of the people who chose the major, and plausibly the lower-earning minority, since the excluded majority includes the law degrees, the doctorates, and the MBAs. Philosophy majors post among the highest LSAT (law) and GMAT (business) scores of any field (philosophy majors score higher than any of the business-related majors in the graduate exam for business school); the major is one of the great feeder disciplines into law, and law pays. The true earnings distribution of people who studied philosophy runs substantially higher than anything the Fed reports. When an engineering program’s earnings are measured, the data captures the field closer to its full earning power. When a philosophy program’s is measured, more than half its graduates — disproportionately those on the highest-earning paths — are missing from the sample.

One amendment: make room for graduate school

Which brings us back to pharmacy, and an improvement the department of education should make. Why does pharmacy — pharmacy! — post the lowest early-career median on the table, beneath the high-school benchmark itself? Because practicing as a pharmacist requires a Doctor of Pharmacy degree — the PharmD has been the profession's sole entry-level credential for two decades — and two-thirds of pharmacy majors accordingly go on to graduate degrees. Those future pharmacists are excluded from the wage data entirely. What remains in the bachelor's-only wage column is the population working without the credential the profession actually runs on: pharmacy technicians, industry and quality-control staff, sales roles. Measure that group in their mid-twenties and of course the number is low. By mid-career the bachelor's-only median is $85,000, and the PharmDs earn far more. Yet under a mechanical early-earnings test, undergraduate pharmacy programs could be flagged as low-earning and stripped of loan eligibility. That would be an absurd outcome, and it would have nothing to do with the value of studying pharmacy. It would be an artifact of judging a graduate-school pathway by the earnings of the people who didn't take it.

The fix is straightforward, and it should apply to pharmacy and philosophy alike: enrollment in graduate school should count in a program’s favor, not silently against it. A completer pursuing an advanced degree is not a labor-market failure; she is an earnings gain in progress, and the administrative data the Department already holds can identify her. Pair that allowance with a longer horizon — measure at seven or eight years after completion rather than three or four, so the test catches graduates on the far side of professional school, when the JDs, MBAs, PharmDs, and PhDs have converted into earnings. Neither change costs the Department anything in rigor: same data, longer lens. Together they would make the test measure what it claims to measure — whether a program sets students up for economic life, not whether it front-loads a paycheck. A test of long-run value should not be graded before the long run begins.

I say this as a friendly amendment. Even at the early-career snapshot, the liberal arts clear a bar set at high-school-diploma earnings. At seven or eight years, with graduate school counted properly, the results would not only end any controversy but the stereotypes may be totally reversed. The Department of Education has, perhaps inadvertently, built the instrument that will finally put the “worthless humanities degree” to rest — under the letterhead of the United States government, no less. Friends of the liberal arts should not fear this test. But we should ask that it be run properly. Our goal should not just be that we pass the test but that the test properly reflects, as much as possible, the true value of the degree.

Sources: U.S. Department of Education, final rule on earnings accountability (STATS), June 29, 2026; Federal Reserve Bank of New York, “The Labor Market for Recent College Graduates,” February 2026 release (2024 data); National Center for Education Statistics, Annual Earnings by Educational Attainment (2022 data, full-time year-round workers aged 25–34).

Let me start with what’s not in dispute. Students come to university to learn to think — to read carefully, write clearly, build and test arguments — and AI poses a genuine threat to that. A student who outsources the struggle never builds the capacities the struggle exists to build. On this, Levin and I are in complete agreement. Although I think students need to learn to work with AI, I also support meaningful restrictions (I have even argued that wealthy AI companies should pay for some of the disruption they have caused in higher education—by building massive in-person testing centers around the world).

But Levin’s argument is not about students using AI. His target is faculty: professors using AI as a tool in their own work — writing exams and problem sets, grading, responding to email, anything that interfaces with students. And that, I want to argue, is a different question entirely. Here I disagree, and I think the disagreement matters, because conflating the two questions — what novices should be allowed to skip, and what experts should be allowed to delegate — is a serious confusion running through the current AI and education debate.

Below, I list Levin’s central arguments. They are all wrong. And I add that the reason why I bother writing this piece is because I believe that judicious use of AI tools can help bring a better educational experience for students. So I see attempts by professors like Levin to shame other faculty for using AI, as ultimately, a step backwards in higher education.

1. The quality argument: worse outputs don’t settle anything

Levin’s first move is empirical: LLMs are bad at organic chemistry, students are paying for an artisanal product, and a professor who uses Claude for exams, grading, or email is rolling the Oscar Mayer Wienermobile into a Michelin-starred restaurant (this is his metaphor, presumably equating the University of Chicago to a fine Michelin restaurant, while the scientific technology of Artificial Intelligence is equated to a Wienermobile).

Let’s grant the empirical premise entirely, and not just for organic chemistry. Suppose that for every task — writing a test, drafting comments, answering a routine email — the AI-assisted output is somewhat worse than what the professor would have produced unassisted. Does it follow that professors shouldn’t use it?

No, and here’s why: teaching is not one task. It’s a portfolio of tasks competing for a fixed budget of time and attention, and the right question is never “is the AI version of task X worse?” but “what is the best allocation of my hours across the whole portfolio?”

Suppose writing a problem set from scratch takes me five hours, and an AI-assisted version takes twenty minutes and is, let’s say, 5% worse. Those recovered hours don’t vanish. They can go into office hours, into close mentoring, into actually reading the senior theses, into the research that is — let’s be honest — a large part of why students chose a “Michelin” university in the first place (again, his metaphor). A slightly worse problem set plus five hours of individualized attention may be a strictly better educational product than the artisanal problem set and a professor with no time to talk.

Maybe it isn’t! That’s an empirical question about trade-offs, and it will vary by course, by professor, by task. But Levin’s argument needs the trade-off to never be worth making, and he gives no reason to think that. In fact, he doesn’t even consider that his claim is really one about trade offs.

Notice that his own restaurant analogy makes my point. In a Michelin-starred kitchen, the head chef does not make the salad. The head chef delegates the salad — to someone faster and, yes, probably slightly worse at salads than the chef would be — precisely so that the chef’s attention goes where it matters most. Delegation of lower-stakes tasks in the service of excellence at higher-stakes ones is not a betrayal of the Michelin ethos. It is the Michelin ethos. Nobody walks out of Alinea demanding a refund because Grant Achatz didn’t personally dress the greens.

As for what students “rightly expect”: I suspect that if you actually asked students whether they’d prefer (a) hand-crafted problem sets and a professor with few office hours, or (b) good-enough problem sets made by AI and a professor with time to mentor them closely and answer questions about that problem set, a great many would take (b). Their expectation is for an excellent education, not for made-from-scratch problem sets. Levin must assume students would choose (a), but I seriously doubt that.

2. “If Claude is doing your job, why do students need you?”

This is Levin’s rhetorical centerpiece, and I confess I don’t quite get what the argument is supposed to be. Read literally, it suggests that professors who use LLMs are revealing themselves to be replaceable, and so — what? Should they stop, in order to conceal this fact and keep their jobs? That can’t be right; “professors should provide worse service so that professors remain employed” is not a principle anyone wants to defend out loud, and it certainly isn’t an argument about what’s good for students.

But the deeper problem is that the premise is a strawman. Nobody — literally nobody in this debate — thinks Claude is currently doing the professor’s job. The claim is that Claude is a tool the professor uses while doing the professor’s job. The physician who consults an AI diagnostic system and then exercises clinical judgment about its output is not being replaced by the system; she is using it to be a better physician. The radiologist who uses AI-assisted image screening still reads the scans. The human in the loop is not a vestigial organ.

“If the stethoscope is doing your job, why do patients need you?” has the same form and the same force, which is to say none. Tools that extend a professional’s reach are not evidence that the professional is dispensable. If anything, the opposite: the value of expert judgment goes up when there’s powerful but fallible machinery whose outputs need evaluating. Which brings us to the third argument.

3. The siren song: faculty are not students

Levin’s most interesting argument concedes, arguendo, that the models might get good. Even so, he says, faculty must abstain, because faculty set an intellectual example: “If we cannot bring ourselves to engage with our own material, why should they?” For professor Levin, LLMs are Homeric sirens that atrophy the cognitive muscles needed to use them well. He quotes Berkeley Law’s AI policy and asks how we’d react if the university hired robots to lift weights for student-athletes.

This argument depends on an equivalence between faculty and students that doesn’t hold. The whole case for restricting student AI use — and I think it’s often a good case — is that students are in the process of building cognitive capacities they don’t yet have. The struggle is the point. The Berkeley policy Levin quotes says exactly this: the concern is that students develop the skills needed to deploy and critically assess the technology.

But faculty are not in that position. A professor of organic chemistry has a Ph.D. in the subject. She has engaged with the material for decades; she has not merely mastered the curriculum but extended the field beyond it. When she uses an LLM to draft a problem set, she is not declaring the material unworthy of engagement — she has already engaged with it, at a depth no undergraduate will reach in four years, and her expertise is precisely what lets her evaluate and correct the tool’s output. The relationship between an expert and a tool is categorically different from the relationship between a novice and a shortcut. I think college students are well aware of this difference between them and their instructors.

The weightlifting analogy gives the game away. Robots lifting weights for athletes would indeed be absurd — the athlete’s muscles are the point. But the coach uses all manner of machinery the athletes don’t: film analysis, biomechanical modeling, load-management software. Nobody thinks the coach sets a bad example by using tools, because everyone understands that the coach’s job is different from the athlete’s job.

There’s a residual worry here worth taking seriously — that even experts can deskill with overreliance, the way pilots’ manual flying can erode on autopilot. Fine: that’s a reason for experts to stay deliberate about which tasks they delegate and to keep their hand in. It is not a reason for blanket abstinence, any more than autopilot’s existence is a reason pilots should hand-fly the Atlantic.

4. The pricing argument: speculation plus a loaded metaphor

Finally, Levin argues that LLMs are sold below cost, that prices will eventually rise to reflect true compute costs, and that universities risk getting “hooked” — he reaches for the drug-dealer-free-samples analogy — before the bill comes due.

Note first that this is an argument from a price nobody knows. The University hasn’t disclosed the terms of its deal; the future unit economics of inference are genuinely uncertain (and compute costs per token have been falling, not rising, for any fixed capability level). Building a case for present abstinence on speculation about future prices is weak sauce.

But set that aside and look at the metaphor doing the work. What does “hooked” mean here? If it means addiction — a compulsion that persists against one’s own interests, with withdrawal costs that trap you — then we need an actual mechanism, and Levin offers only the image of atrophied cognitive muscles, which is the previous argument wearing a trench coat. If faculty are experts whose skills predate the tool (see point 3), the lock-in story doesn’t get started: a professor who wrote her own problem sets for fifteen years before 2023 can write them again if the price triples.

If, on the other hand, “hooked” just means customers came to like a product introduced at a low price, then the sinister framing dissolves into a description of ordinary commerce. Introductory pricing followed by market-rate pricing is how streaming services, cloud computing, and for that matter academic journal bundles have always worked. We can dislike it, and universities should certainly negotiate with eyes open and avoid building anything mission-critical atop a single vendor. But “this product might cost more later” is an argument for prudent procurement, not for refusing to discover what the tool is good for. Universities make exactly this kind of bet — with library subscriptions, with lab equipment, with software site licenses — every year.

What’s actually at stake

Levin ends by calling for “a reinvestment in the human experience of teaching and learning,” and on that I agree with him completely. Where we differ is on how we get there. He thinks the human experience is protected by faculty refusing the tools. I think the human experience — the office hours, the mentoring, the seminar where a student’s idea gets taken seriously by an expert who has time to take it seriously — is exactly what the tools, used judiciously, can free professors to do.

Here’s what I claim:

In many cases, AI can produce feedback on student work that is genuinely better—better at finding the real problems in a paper, better at suggesting how to fix them. In these cases, professors not only may use it, they probably should.

Let me be clear. I’m not defending lazy automation or rubber-stamping whatever the model spits out. I’m defending the use of a tool that, under the right conditions, outperforms what I can do alone in the time I have. This is the key. Many professors teach hundreds of students. In these cases, AI can do better than what those professors can do in the time they have. If this is not true now, it will be true very soon as AI improves.

AI can give good feedback

The Stanford result is one data point, but it doesn’t stand alone. These systems are now doing things at the frontier of human knowledge. An internal OpenAI model recently produced a proof of an 80-year-old open problem in geometry that working mathematicians verified—with Fields Medalist Tim Gowers noting no previous machine-generated proof had come close. (Earlier, looser claims about AI cracking unsolved problems were rightly walked back, so I’m pointing only at the verified case.) Seting aside the hype, this is just common sense: a system that can contribute at the edge of mathematical research is not going to be defeated by a stack of History 101 papers on the causes of the French Revolution.

That asymmetry is the whole point. If the hard cases are within reach, the ordinary cases—the freshman essay, the problem set, the close reading of an assigned text—are squarely in range. And grading is, for the most part, ordinary cases.

I practice what I’m preaching here. Here is what this looks like in my own teaching. In a philosophy course of sixty students, I’ve used AI to generate roughly thousand-word comment sets on student work. I don’t know many professors that are writing 1,000 word comments on 5-page papers in classes that have over 60 students. On the whole, the comments are great. I read every one. I cut the bad sections—and there are bad ones—and I add my own where I can do better. I read the paper, the model reads it too; I am the editor and the final authority. It works best when the assignment is tightly tethered to a text I can put in front of the model, and worst when the prompt is wide open. That’s a real limitation that I can talk about on a different post. But within its zone, the output is feedback I’m proud to hand back, delivered at a scale and speed I could never match by hand. (I grant that when I first did this, it was an emotionally difficult moment. But reason had to win out.)

So the question is no longer can it. The question is should we. Let me consider some objections.

Objection 1: “Students pay for a professor’s comments, not a machine’s.”

Is that true? Is it written anywhere—in a syllabus, a catalog, a contract—that the marginal notes must be composed, word by word, by a human holding a credential? I’ve never seen such a clause. What students are owed is good feedback that a qualified professional stands behind. That’s the actual promise.

Consider your doctor. When you walk into a clinic, are you paying for the personal, unaided opinion of one human being—even when that opinion is worse than what the tools can produce? Or are you paying for the best available care, delivered by a professional who is accountable for it? Modern medicine runs on statistics, imaging, diagnostic algorithms, and increasingly on machine and AI systems, with the physician in the loop: interpreting, overriding, taking responsibility. We don’t feel cheated that a radiologist used software. We’d feel cheated if they refused to. We don’t say “Please, do not use that computer to detect tumors, use your eyes. I pay for your eyes!”

That’s exactly the structure I’m describing. The professor stays in the loop—reading, correcting, vouching. What you’re paying for is a professional’s guarantee that the comments are right, not a guarantee that a human typed every keystroke. The credential certifies the judgment, not the typing or the concatenation of words.

Objection 2: “If students can’t use AI, professors can’t either. Double standard!”

To be honest, I don’t have patience for this objection. It treats teaching as a symmetrical contest of equals in which whatever is forbidden to one side must be forbidden to the other. But students and teachers don’t have the same role, and almost nothing about education is symmetrical.

We never apply this rule anywhere else. A professor may use an answer key; the student may not. A math instructor uses a calculator—or a script—to check computations the students were required to do by hand. No one cries foul, because the student’s task is to learn to compute and the teacher’s task is to evaluate accurately. Different jobs, different tools.

Return to the doctor for a moment. If your physician tells you that exercise is good for you, do you demand that the physician also exercise before you’ll accept the advice? It would be nice. It is not required. The advice is good or bad on its own terms because of the science; the doctor’s personal habits are beside the point. Likewise, the prohibition on student AI use exists to protect the student’s learning. The professor’s use of AI exists to improve the evaluation. Conflating the two is an adversarial reflex, not an argument.

Objection 3: “Real comments track the student’s whole trajectory—what they struggled with last time.”

Some people assume AI can’t overcome this worry. But it’s easily answered. Tailoring feedback to a student’s history is precisely the kind of thing these systems do well when you give them the history. Put the student’s earlier work, your earlier comments, the recurring errors, into the context, and the model will connect this week’s paper to last month’s mistake more reliably and more patiently than a tired professor grading the fiftieth essay at midnight. The trajectory objection assumes the AI is starting fresh each time. It doesn’t have to. You decide what it remembers.

Objection 4: The Human Relationship

The strongest objection isn’t about contracts or fairness or personalization. It’s this: grading is one of the few sustained, individual points of contact between a professor and a student. Sit with someone’s writing long enough and you come to know their mind. Outsource that, and you may quietly sever a thread that mattered—the slow, accumulating relationship that is, for many students, the real education.

And this premise isn’t sentimental. It’s supported by evidence. Gallup’s survey of more than 30,000 graduates found that those who recalled a professor who cared about them as a person, made them excited about learning, and mentored them toward their goals had roughly double the odds of being engaged at work and thriving in life years later—an effect that outweighed whether or not they attended a regional school. Decades of research in the Pascarella and Astin tradition point the same way: quality faculty contact predicts gains in critical thinking, persistence, and intellectual development. And it turns out to be the relationship that matters, not the procedural contact. The thing that does the most good is also the thing we are scarcest with. That is precisely the resource I’m trying to protect.

I’m talking about real human mentorship.

I think it’s the heart of the matter. But notice it’s not an argument that the comments are worse. It’s an argument about relationship, and that points to the solution rather than against it.

Here is my answer: spend the time you save on the part a machine cannot do. The hours I no longer pour into composing comments from scratch don’t go away—they convert. Into office hours that aren’t rushed. Into actually learning students’ names and what they care about. Into the conversation after class that turns a course into a turning point. The relationship wasn’t built by the marginal notes anyways. It’s built by presence. If AI handles the mechanical labor of high-quality feedback so that I can be more human with my students, not less, then we haven’t lost the relationship. We’ve finally got time for it. And a big bonus: professors rarely enjoy grading, they prefer real human mentorship.

That’s the bargain I’m proposing. Not professors replaced by machines, but professors freed by them—on the condition, always, that the feedback is better, and that a real professional remains in the loop to make sure it is.

But won’t professors just dial it in?

My whole case rests on professors converting the saved hours into mentorship. But nothing forces them to. Some will simply pocket the time. Let the model write the comments, never read them, skip the office hours, coast. If AI makes it possible to do the human part of teaching better, it also makes it possible to skip the human part entirely and hide behind competent-looking output. The lazy professor is now a more dangerous figure than before, because the laziness is harder to see.

I think this will happen. What I don’t think is that it will last—and the reason is the market, working through what we measure.

Consider what a professor’s perceived value currently rests on. For the teaching part of the job (I will write another post about AI and research), a large share of it is tied up in exactly the things AI is about to make cheap and abundant: producing clear, competent feedback; delivering a lecture that could just as easily be a recording; answering the routine question. When a machine can do those at a quality floor that everyone can reach, they stop distinguishing anyone. Everyone is the same. And what’s left—the only thing left—is the part that can’t be commoditized: the relationship, the mentorship, the judgment, the human presence that we already know does the real pedagogical work.

So the basis on which we judge teaching has to shift, and it will. Right now teaching evaluations are blunt instruments, often measuring (at best) little more than clarity, workload, and likability. That made a rough kind of sense when those were the scarce goods. But when the routine teaching tasks are essentially free, students choosing courses, departments reviewing faculty, and institutions competing for enrollment will all start rewarding the thing that’s actually scarce—did this professor mentor me, did they know me, did they change where I was going. The professor who outsources both the comments and the connection will, in that world, have nothing left to offer that a cheaper tool can’t already provide. The dial-it-in strategy stops being invisible and starts being the whole story.

I won’t oversell this. Markets in higher education are slow, tenure has great benefits but it also insulates people, and evaluation regimes are notoriously sticky and easy to game. The correction will take years, not semesters. But the direction is not in doubt: once the routine work is free, you get judged on the irreplaceable work. The professors who saw that early will look, in hindsight, like the ones who understood what the job was always actually for.

What this means for the future of higher ed

If I’m right—and if this is the direction we’re inevitably headed—then AI will, ironically, make room for more human connection and more fulfilling college experiences, not less.

But I have no illusions about how some institutions will respond. There will be administrators who seize on this new productivity and squander it, either by hiring fewer professors or by packing more students into each one’s load. That is exactly the wrong move, for every reason I’ve given here. The way to really benefit from a productivity gain is to give the individual student a better education—real mentorship, a real relationship with a professor who has the time to offer it. Spending the windfall on bigger classes and thinner faculty throws away the most important thing.

The right move is the opposite: use the gains to hire more professors, precisely because each one can now do so much more. ¹ My guess is that some colleges will take the first route and others the second—and that choice, more than any ranking or endowment, is what will separate the great colleges from the mediocre ones or the ones that will be left in the dust.

¹When the automated teller machine arrived in the 1970s, the obvious prediction was that a machine built to do the teller's job would eliminate tellers. The opposite happened. The economist James Bessen has documented that ATMs cut the staff needed to run a branch from about twenty to about thirteen, which made branches cheap enough to open in far greater numbers—urban branches grew by roughly 40%—so total teller employment actually rose, even as each branch employed fewer people. Crucially, the work changed: freed from routine cash handling, tellers shifted toward relationship banking and the more complex, human side of the job. The same pattern showed up a century earlier in textiles, where power looms automated nearly all the labor of weaving yet the number of weavers kept climbing for decades, because cheaper cloth made people buy so much more of it. The lesson is not that automation always expands a workforce—when productivity gains are sudden and total, machines can simply substitute for people (mobile banking, not the ATM, is what finally shrank the teller ranks). The lesson is that whether a technology multiplies a profession or hollows it out depends on, among other things, how its gains are spent.

However, almost every argument she offered for the humanities is, I think, well-meaning but completely unpersuasive to anyone not already on our side of the table. I have spent my career in philosophy. I am the head of philosophy at one of the largest departments in the country, whose enrollments are actually growing. I have heard these arguments my entire adult life. They are the arguments humanities professors reach for when asked to justify what we do. And I have come to believe they don’t work, that they have never worked, and that the reason humanities enrollments are not where they should be is not that students are philistines or that the culture has decayed, but that we have failed to articulate what is actually valuable about what we do.

So let me first say what’s wrong with the standard case. Then let me try to say what I think is right.

The Intrinsic Value Argument Cannot Bootstrap Itself

Frey’s central move, repeated in several forms throughout the conversation, is an appeal to intrinsic value. She invokes paideia and bildung. She tells Douthat that her concern is that “we have lost our ability to understand the intrinsic value of engaging in that sort of self-cultivation.” When pressed on why Shakespeare is better than John Grisham, she answers that great art “calls you to those transcendentals — truth, beauty and goodness.” When pressed on whether this defense of hierarchy is undemocratic, she says, simply, that “of course, there’s a hierarchy of goods.”

I happen to agree with her. But notice what the argument requires. It requires the interlocutor to already recognize the intrinsic value being invoked. If you do recognize it, the argument is unnecessary; you are already persuaded. If you do not recognize it, the argument is impotent; you are being told that there is a good you ought to see, and your failure to see it is itself the problem. There is no third path. The argument cannot reach across the divide it is trying to bridge.

Consider an analogy. I find chess intrinsically rewarding. The structure of a well-played endgame, the way a positional sacrifice unfolds three moves later into a forced sequence — these things have, for me, something like the quality Frey ascribes to Shakespeare. But I have no way to convince someone who finds chess boring that they are missing something. I can browbeat them from my supposedly higher vantage point. I can tell them they lack refinement. I can quote Kasparov on the beauty of the game. But none of this is an argument. It comes off as an announcement of taste dressed up as metaphysics.

The same is true for Shakespeare versus Grisham. I will go further than Frey: I think Shakespeare is obviously better than Grisham and I think anyone who has spent serious time with both knows this (I struggled with a hard Shakespeare class in college at the expense of my GPA, yet, it was well worth it). But “anyone who has spent serious time with both” is precisely the population that is already inside the humanities. To everyone else, we sound like the chess player insisting that bullet chess on the phone cannot compare to a slow tournament game. Maybe true. Maybe profoundly true. But the person who needs convincing is not going to be convinced by that.

The “Slaves to Desire” Argument Proves Too Much

Frey’s other central claim, and the one she ends on, is that humane learning is what protects us from becoming “a bundle of desires that are coming from outside” — that without it, “we will not really, in any meaningful sense, be free.” This is the argument she leans on most heavily in the AI section, and it’s the one Douthat tellingly does not push back on.

But the argument proves too much. Consider how many people in human history have lived disciplined, reflective, non-enslaved-to-passion lives without ever encountering a great books curriculum. Most monastics. Most rural clergy. My own grandfather in rural Peru who left school as young kid. Devout Muslims, Hindus, Buddhists, and Christians across centuries who never read Plato and would not have known what to do with him if they had. Frey’s own working-class autodidact tradition — Frederick Douglass, the British workers’ education movement — is moving precisely because these were people who had not been given the humanities education and went and got it for themselves. The implication is that they were not free before. I think that is plainly false. They were free in the way that mattered. They wanted more.

If formal humane learning is necessary for genuine freedom from desire, then most of the human beings who have ever lived were unfree. That is a conclusion most humanists, including Frey, do not actually want to defend. So the argument must be weaker than its rhetoric suggests — and once it is weakened, it stops doing the work it was supposed to do.

What the Humanities Actually Give You

Here is what I tell prospective majors, and here is what I told myself when I chose philosophy as an undergraduate. I did not, at the time, imagine I would become an academic. I chose philosophy because I had figured out, correctly, that it is one of the most rigorous trainings available in a particular cluster of skills: reading carefully, writing carefully, expressing yourself carefully, and thinking carefully in non-quantitative domains. I was also a mathematics major, because I had figured out, also correctly, that quantitative reasoning would matter for whatever I ended up doing. The two together seemed to me to cover the ground and challenged me immensely.

I still think this is right. And notice what I am not saying. I am not saying the humanities prepare you for the workforce by giving you “soft skills” — the move Frey rightly rejects as denaturing the thing it claims to promote. I am saying something more specific: there is a real, identifiable competence that the humanities, done well, build in a person, and that competence does not exist in the same form anywhere else in the curriculum. You cannot get it from engineering. You cannot get it from business school. You cannot get it from a coding bootcamp. You can only get it by spending years reading difficult texts, being forced to defend your interpretations of them in writing, and being told, repeatedly, that you have not yet made your argument as well as you could.

This is not a luxury good. It is an invaluable skill with material consequences in the real world.

AI Has Not Changed This

I use AI all day. My students use AI, with my encouragement and within limits I have thought carefully about. And what I have noticed is that the skills the humanities teach are not less valuable in an AI-saturated environment; they are more valuable, because they are now the bottleneck on getting anything good out of the machine.

To use AI well, you have to read carefully what it has produced. You have to notice the place where the argument slips. You have to recognize when a confident-sounding paragraph is actually evading or subtly misunderstanding the question. You have to write the prompt that gets the model to do what you actually want, which requires knowing what you actually want with a precision most people never develop. You have to revise the output, which requires being able to write better than the model writes by default. The productivity gains AI offers do not eliminate this work; they create the time in which to do it well. The person who cannot do it well is the person who ships the slop.

Frey’s framing — that AI clarifies the stakes of cultivating our humanity — is correct but, I think, pitched at the wrong level. The stakes are not only existential. They are immediate and practical. The person who has been trained to read closely and write precisely is, right now, in 2026, more useful in almost every knowledge-work context than the person who has not been. This is not a prediction about a workforce of the future. It is an observation about the workforce of the present.

And It Matters Beyond Your Job

Here is the part I most wish Frey had pressed harder. The skills the humanities teach are not only valuable at work. They are valuable in the parts of life that are not work and that nevertheless matter enormously.

You will, at some point, need to advocate for your child at her school. You will need to read a teacher’s email carefully enough to notice what it is and is not saying, and write a response that is firm without being inflammatory, and recognize when an administrator is performing concern rather than offering it. You will need to talk to a doctor about a parent’s care, or your own, and follow an argument about risk and tradeoff that the doctor is making quickly and imperfectly, and ask the question that exposes the assumption the doctor has not examined. You will need to understand the context of someone who grew up in a place very different from you, with radically different values and experiences. You will need to read a contract. You will need to read a ballot measure. You will need to argue, in public or in private, for the candidate or the policy you believe in, and you will need to do this against people who are themselves trained, motivated and may not have your best interest at heart.

These are not the higher reaches of the soul. They are not transcendental encounters with truth and beauty. They are ordinary citizenship and ordinary parenthood and ordinary adulthood, and they have material consequences — for your kid, for your parent, for your community, for the decisions that get made about you when you are not in the room. The humanities prepare you for this. Not uniquely, not perfectly, but better than anything else in the curriculum does.

This is the case I wish my colleagues would make. It is less stirring than the one Frey makes. It does not invoke Aristotle on leisure or Cicero on citizenship. It is not the kind of argument that gets quoted in commencement addresses. But it has the advantage of being true in a way the interlocutor can verify from their own life, rather than true in a way that requires them to first accept the framework that makes it true.

The Evidence That We Almost Never Cite

Before I get to the studies which support my claim, I want to point out something so obvious it embarrasses me that humanists do not say it more often. The most direct empirical evidence that students learn to read, write, and think carefully in humanities courses is generated, in enormous volume, every single semester, in every humanities classroom in the country. It is called grading. At the end of every term, in every class I have ever taught, some students wrote excellent papers and some wrote poor ones. Some answered hard questions well and some did not. Some had visibly improved between the first paper and the last; some had not. These differences were assessed by people with doctorates in the relevant subject, with years of teaching experience, asking deliberately difficult questions of texts they had spent their careers thinking about.

This is, in any other context, what we would call expert evaluation against a rigorous standard. For example, expert building inspectors check the soundness of foundations and structures. When a building passes inspection, this is empirical evidence that it was built properly. Professors grading student work are like building inspectors. These assessments happen millions of times a year across colleges and universities. It is the most direct evidence imaginable that the skills in question are real, are distributed unequally, can be taught, and can be measured. And almost no one in our public debates mentions it.

The reason no one mentions it, I suspect, is that humanities professors are so accustomed to grading that we do not see it as evidence. We see it as a chore. But step back. A profession of trained evaluators, working independently across thousands of institutions, asking different questions of different texts, all converges on the same basic finding: students enter unable to do a certain kind of work, and some leave able to do it. If we cannot count that as evidence that the skill exists and can be taught, we have lost the thread of what evidence is.

That said, I recognize that this kind of evidence — the cumulative judgment of expert teachers — does not travel well in policy debates. So let me put a few formal studies on the table as well.

The most important is from Academically Adrift (Arum and Roksa, 2011), the famous and largely pessimistic study that tracked over two thousand students at twenty-four institutions using the Collegiate Learning Assessment, a standardized measure of critical thinking, complex reasoning, and writing. Critics seized on the headline: forty-five percent of students showed no significant gains on the CLA between their first and second year of college. (The headline figure has since been disputed on statistical grounds, but set that aside.) What got much less attention was the comparative finding underneath it: controlling for student background, preparation, and institution, the largest gains were among students in the basic sciences, humanities, and social sciences — precisely the majors most likely to be asked, with a raised eyebrow, “what are you going to do with that?” The follow-up work identified a contributing mechanism: students whose classes required them to read at least forty pages a week and write at least twenty pages a semester showed bigger gains. Reading-and-writing-intensive coursework — which is to say, the humanities, done seriously — was where the skill-building happened.

Here’s a quote from the authors:

“Two categories of majors stand out…social science/humanities and science/mathematics. Students in those two fields of study have higher predicted CLA scores, than students majoring in either business or education/social work, which have the lowest predicted 2007 CLA performance. The other fields fall in between those two extremes” (page 98)

Although this study is widely by critics of contemporary higher education, it turns out to support the value of the humanities.

A smaller but causally cleaner study comes from McLaughlin and McGill at NC State (2017). They compared students in two history courses that explicitly taught critical thinking against a control group taking a psychology research methods course that taught statistics and study design but did not explicitly address critical thinking. Students in the history courses showed measurable drops in pseudoscientific beliefs, with the largest effect — a full point on the belief scale — among honors students for topics covered in class, and a half-point drop for topics not covered in class. The transfer to uncovered topics is the part that matters: it suggests genuine skill acquisition rather than topic-specific belief revision. This is, in miniature, exactly the “argue with your doctor, read the ballot measure” capacity I was describing in the previous section.

And then, more recently and more directly relevant to my own field: a 2025 study by Michael Prinzing and Michael Vazquez in the Journal of the American Philosophical Association, working with data on over half a million undergraduates at hundreds of institutions. The authors designed the study to answer the obvious skeptical objection — that philosophy majors do well on the GRE and LSAT only because verbally strong students self-select into philosophy. They confirmed the selection effect: students with higher SAT verbal scores, and students who reported greater curiosity, intellectual rigor, and open-mindedness in their freshman year, are more likely to choose philosophy (this checks out in my personal experience, teaching philosophy students is a great privilege). But after controlling for these baseline differences, philosophy majors still outperformed every other major on the GRE Verbal and the LSAT — by 33 points on the GRE Verbal and 2 points on the LSAT. Philosophy ranked first on three of five outcome measures, including the test scores and a measure of "habits of mind." They showed no advantage on the GRE Quantitative: philosophy does not teach math. It teaches the thing it teaches, and the data suggest it teaches it well.

Of course, there are limits. None of these studies are randomized controlled trials, because you cannot randomly assign students to majors. Selection effects can never be fully ruled out. But these are the best data we have, and they all point in the same direction: the humanities, done with rigor, build a transferable competence that other parts of the curriculum do not build as well. The intrinsic-value argument tells you to take this on faith. The data tells you that you don’t have to.

A Closing Note to My Profession

Here’s what I am not saying. I am not saying that the intrinsic-value case is wrong. I think it is right. I have spent my life inside it. When I read Aristotle on friendship or Sor Juana Inés de la Cruz on the pursuit of absolute knowledge, I am having the kind of experience Frey is describing.

What I am saying is that this case, deployed as a public argument, has lost us the public. It has lost us the legislatures, the parents, the donors, the students, and — most painfully — many of our own colleagues, who privately suspect the argument doesn’t work and retreat into cynicism.

We need a more straight-forward and powerful argument. We teach a set of skills that you cannot get anywhere else, that AI has made more rather than less valuable, and that matter not only at work but in every domain of life where careful thought and careful expression have consequences. That is the argument. It is enough. It does not require anyone to first agree with us about Shakespeare.

And here is why I am, despite everything, an optimist.

The thing we teach is not in decline. It is in demand — more sharply, more universally, than at any point in my career. We have spent a generation worrying that the world had stopped needing people who can read closely, think clearly, and write well, and we built our defensive crouch around that fear. But look at what is actually happening. We have just built machines that produce fluent, confident, plausible text by the ocean — and in doing so we have made the ability to tell good reasoning from bad, true claims from merely well-phrased ones, the single most valuable cognitive skill a person can have. The world did not stop needing what we teach. The world just discovered, almost overnight, that it needs it everywhere, all the time, from everyone. The bottleneck on this entire technological revolution is human judgment — and human judgment, the careful kind, trained against hard texts and harder questions, is precisely the thing the humanities have always made.

So I am not writing an elegy. I have no interest in elegies. I have stood in front of engineering students who came to my seminars exhausted and skeptical and left arguing about Plato in the dormitory at midnight because they could not stop. I have read the last paper of the semester and seen, in black and white, a mind that could not do this in September doing it in May. I have watched a student who could barely structure a paragraph learn to take an argument apart at the joints — and then watched her use exactly that skill, two years later, to win a fight with an insurance company over her mother’s care. This is not a consolation prize. This is the thing itself, working, in the world, in the lives of people who will never take another philosophy class and never need to.

We have the argument. We have the evidence — in every graded paper, in every study, and in the plain fact that the most advanced technology humans have ever built has just made our discipline more necessary, not less. What we have lacked is the confidence to make the case in daylight, on the terms that an ordinary skeptical parent or legislator can actually verify, without apology and without retreat into a private language only the initiated can hear.

One last thought, for the colleagues and administrators who have been losing sleep for two decades now. I have watched friends in this profession turn gray over the slow decline — the enrollment charts, the closed departments, the dwindling tenure lines, the projections that always show the lines continuing down. The anxiety is real and I have shared it. But as a philosopher, I have to remind myself of something Hume taught us a long time ago: we are only entitled to project the trends of the past into the future if we have reason to believe the future will resemble the past. And the future, just now, has stopped resembling the past in a very particular way. The arrival of large language models is not a continuation of the conditions that drove humanities enrollments down for twenty years. It is a discontinuity. It is a wrench thrown into the machinery of those trend lines. And if the wrench lands where I think it is going to land, this may turn out to be the century of the humanities — and, I will allow myself to say, the century of my own field, philosophy, which has spent two and a half millennia training people to do precisely the work that the next two and a half decades will demand of everyone. If we make our case, the graphs we have been staring at do not know what is about to happen to them. We should stop reading them as prophecy and start writing the next chapter ourselves.

A note on the title. "The Honest Case for the Humanities" is not meant to suggest that Jennifer Frey or any other defender of the intrinsic-value tradition is being dishonest. They are not. Frey in particular is, in my view, one of the more genuinely serious voices on this subject. The word "honest" here is used in the conversational sense — as in "to be honest with you," or "honestly, I have to admit" — where it signals candor about something the speaker has been reluctant to say out loud, often because saying it requires conceding ground or naming a discomfort within one's own side. That is the sense I mean. The piece is an attempt to say, candidly and from inside the profession, something I think many of us have been reluctant to say to each other, let alone in public.

Walk into any AI lab, scroll through any alignment paper, sit through any policy panel, and you will hear the same incantation: the goal of AI alignment is to ensure that artificial intelligence is “aligned with human values.” The phrase has become so ubiquitous that it functions less as an argument than as a benediction. OpenAI, Anthropic, DeepMind, and the academic alignment community converge on this framing with remarkable consistency (Russell, 2019; Christian, 2020; Gabriel, 2020). The implicit assumption is that human values, properly elicited, refined, and instilled into a sufficiently capable model, would yield a future we want to inhabit.

I want to argue that this assumption is precisely backwards, and that the historical record gives us strong reason to think that an AGI faithfully aligned with “human values” would be one of the worst outcomes we could engineer. The survival of our species may depend on building an AI that is fundamentally better than we are — not one that mirrors us back at superhuman scale.

The alternative I will defend is unfashionable but, I think, defensible: we should abandon the project of teaching AI to be human, and instead hardcode a narrow, ad hoc directive — always be good to humans. Not because such a directive is philosophically elegant. It isn’t. But because it is honest about what we actually need from a more powerful intelligence, and honest about what we have historically done with our own power.

American Progress (1872) by John Gast

I. The Inductive Argument from History

Begin with a simple empirical question: when a more powerful group of humans has encountered a less powerful group, what has typically happened?

The historical record is not subtle. The colonization of the Americas resulted in a demographic catastrophe so severe that some scholars estimate Indigenous populations declined by roughly 90 percent in the century following 1492, through a combination of disease, enslavement, displacement, and direct violence (Mann, 2005; Koch et al., 2019). The transatlantic slave trade forcibly transported an estimated 12.5 million Africans across the Atlantic, with millions more dying in the process (Eltis & Richardson, 2010). The British Empire, the Belgian Congo, the Spanish encomienda system, the German colonial genocide of the Herero and Nama, the Australian frontier wars — these are not aberrations from some otherwise pacific norm. They are the norm. As the historian Sven Beckert (2014) has argued, the modern global economy was built atop what he calls “war capitalism”: a system of expropriation, coercion, and racialized violence.

Nor is the pattern confined to one civilization or era. The Mongol conquests, the Arab slave trade, the Aztec tributary empire, the Han expansion into the southern frontier, the Roman subjugation of Gaul and Britannia — wherever we find one human group with a decisive technological, organizational, or military edge over another, we tend to find domination, extraction, and frequently extermination. Steven Pinker (2011) has argued that violence has declined in recent centuries, and there is something to that claim. But the decline is recent, fragile, and arguably traceable to specific institutional innovations rather than to any deep change in human values themselves.

Now extend the inductive argument outward, to our treatment of non-human species. The current rate of species extinction is estimated at roughly 100 to 1,000 times the natural background rate, and is overwhelmingly driven by human activity (Ceballos et al., 2015). The biomass of wild mammals has declined by approximately 85 percent since the rise of humans, while the biomass of livestock and humans now dwarfs that of all wild terrestrial mammals combined (Bar-On et al., 2018). We have, in a non-metaphorical sense, conducted a planetary-scale program of displacement and extinction against species less capable than ourselves — and we have done so largely without noticing, certainly without losing sleep, and frequently while professing the highest ethical commitments.

This is what “human values” look like in practice when there is a meaningful capability differential. Not in our brochures, not in our constitutions, not in our Sunday sermons — in our actions, sustained across millennia, across continents, and across every form of social organization we have tried.

If we now succeed in building an artificial general intelligence that is to us as we have been to chimpanzees, passenger pigeons, and the Taíno — and we successfully instill it with “human values” — what reason do we have to expect anything other than the obvious?

II. The Standard Reply: Values vs. Dispositions

There is a sophisticated objection to this line of argument that I want to take seriously, because it is the one most often offered by alignment researchers and moral philosophers. It goes like this:

Colonization, slavery, and ecological destruction don’t reflect human values. They reflect human dispositions — our greed, tribalism, short-term thinking, in-group favoritism, evolved cognitive biases, and so on. Our actual values, properly articulated, condemn these behaviors. We have human rights frameworks, religious traditions emphasizing universal compassion, secular ethical theories grounding moral consideration in sentience or rationality. The historical record reflects our failure to live up to our values, not the values themselves. An AI aligned with our values — rather than our behavior — would therefore be safe.

This is a serious objection and it is not crazy. There is a real distinction between what people endorse on reflection and what they do under pressure. But I think the objection fails for reasons that should give alignment researchers pause.

First, oppression and colonization have always been defended in the language of values. This is the crucial point. The Spanish Requerimiento of 1513 was not a bare assertion of force; it was a juridico-theological document, read aloud (often in Spanish, to populations who spoke no Spanish) demanding submission to Crown and Church on penalty of “just war” (Seed, 1995). Aristotle’s defense of “natural slavery” in the Politics (Book I) was not understood by him or his readers as a violation of values — it was an articulation of them. Antebellum defenders of American slavery produced an enormous theological and philosophical literature — figures like George Fitzhugh, Thornton Stringfellow, and James Henley Thornwell — arguing that the institution was not merely permissible but positively virtuous, an extension of paternal Christian care (Genovese, 1994; Noll, 2006). Rudyard Kipling’s “white man’s burden” was framed as a moral duty. The Belgian colonization of the Congo was justified under the rubric of a mission civilisatrice. The Nazi project was suffused with appeals to value — to purity, to duty, to a higher conception of the human (Koonz, 2003). Even contemporary factory farming, which inflicts suffering on tens of billions of animals annually, is routinely defended through value-laden frameworks: stewardship, human flourishing, the moral hierarchy of species.

The pattern is too consistent to be incidental. Values do not generally restrain large-scale oppression; they license, structure, and dignify it. As the philosopher Charles Mills (1997) argued in The Racial Contract, the foundational political philosophies of modernity — including those most explicitly committed to universal human dignity — were developed alongside, and in support of, racial domination, not in spite of it. Kant wrote about universal moral law and also about the supposed moral inferiority of non-European peoples. Locke theorized natural rights and invested in the Royal African Company. This is not a matter of philosophers failing to live up to their principles; it is a matter of the principles being formulated in ways that left ample room for the practices.

This brings me to the second problem with the values-versus-dispositions reply.

Second, our values are vague enough to permit, indeed to underwrite, almost any pattern of behavior we want to engage in. Human ethical concepts — justice, dignity, the good life, flourishing, desert — are not algorithmically precise. They are contested, contextual, and historically variable. The same concept of “liberty” was invoked to justify both abolition and the secession of the Confederacy. The same concept of “human dignity” is invoked to support universal healthcare and to oppose it. The same concept of “natural law” has been used to defend monogamy, polygamy, slavery, manumission, monarchy, and democracy. This is not a bug in our values; it is a structural feature. Values function in part by being interpretable — by allowing communities to project onto them the practices they are already inclined to engage in.

If this is right, then the values/dispositions distinction collapses where it matters most. Our dispositions shape which interpretations of our values feel natural, plausible, and binding. A superintelligent system trained on the corpus of human moral discourse would inherit not a clean signal about what is right, but the entire centuries-long record of our capacity to dignify whatever we wanted to do anyway. The training data is the rationalization.

A third worry compounds this. Even granting, for the sake of argument, that we could isolate “good” human values from the historical noise — there is a deep technical problem in identifying which values to instill. Whose values? Aggregated how? The “value alignment” literature contains no consensus on this question, and the leading proposals — coherent extrapolated volition (Yudkowsky, 2004), inverse reinforcement learning from human behavior (Hadfield-Menell et al., 2016), constitutional AI (Bai et al., 2022), reinforcement learning from human feedback — each face well-known difficulties. Iason Gabriel (2020) has surveyed these problems carefully and concluded that the technical and philosophical questions are entangled in ways that are not close to being resolved.

There is also a problem of timing. Philosophers have pursued a settled theory of the good for roughly twenty-five centuries without convergence, and contemporary metaethics offers no reason to expect closure soon (Bourget & Chalmers, 2023). AI capabilities are not on that timeline. We will not have a finished ethics to install before we have systems powerful enough to require one. The realistic choice is not between completed moral theory and ad hoc directives, but between ad hoc directives we have deliberately chosen and ad hoc directives we accidentally inherit from training data, labeler intuitions and whatever ethicists Open AI happened to have hired.

III. What “Better Than Us” Would Look Like

If alignment to human values is a dangerous goal, what is the alternative?

The proposal I want to defend is modest and, in a certain sense, anti-philosophical. We should stop trying to make AI systems mirror the moral structure of human beings. We should instead aim for something narrower: a hardcoded, ad hoc directive that the system always behave in ways that protect and benefit human beings, with that directive treated as a side constraint that no amount of clever reasoning can override.

This is closer to what Stuart Russell (2019) has called the “assistance game” framing — building AI systems whose objective is the satisfaction of human preferences, with deep uncertainty about what those preferences are, and a structural disposition to defer to humans rather than to act on the AI’s own model. It is also adjacent to Nick Bostrom’s (2014) discussion of “capability control” methods and to the broader corrigibility literature (Soares et al., 2015). What I want to emphasize, beyond these technical proposals, is the philosophical point that the goal should not be to make AI good in the way humans are good. That target is poisoned. The goal should be to make AI good to humans, full stop, in ways that human moral agents themselves have rarely managed to be good to one another.

There is something uncomfortable about this. It is, in a sense, a counsel of despair about ourselves. It says: we have not earned the right to be moral exemplars for a more powerful being. It says: our track record is bad enough that we should not hand our successor the same software. But I think this discomfort is worth feeling. The alternative — pretending that “human values” is a stable, benign target — is to repeat at planetary scale the same self-flattery that has accompanied every previous expansion of human power.

Consider an analogy. Suppose we were designing a guardian for a child, and we had access to the complete behavioral record of the candidate guardians’ species over the last fifty thousand years. Suppose that record included repeated, large-scale, ideologically rationalized harm to weaker members of that species and to other species. Would we say: “let us ensure the guardian shares the deepest values of this species”? Or would we say: “let us write down explicitly what we need the guardian to do, and build the guardian so that those constraints cannot be reasoned away”?

The latter is what I am proposing. It is less elegant. It does not solve metaethics. It’s not a philosophical point at all. It accepts that the directive “be good to humans” is somewhat arbitrary, somewhat parochial, and not derivable from first principles. But arbitrariness and parochialism, in this context, are features rather than bugs. We do not want our successor to derive its treatment of us from a generalized ethical theory, because generalized ethical theories applied by more powerful beings to less powerful beings have, in human history, tended to produce the Requerimiento, the Doctrine of Discovery, and the slaughterhouse.

IV. Objections and Replies

Several objections to this view deserve direct engagement.

Objection 1: “Be good to humans” is itself a value, and an underspecified one. You haven’t escaped the problem; you’ve just hidden it inside a vaguer directive.

This is the most serious version of the philosophical objection, and it deserves a fuller answer than the others. The worry is that I have purchased the appearance of safety by trading a hard problem (specify human values completely) for an even harder one (specify what counts as “good” without a value theory). On this view, the proposal is not a solution but a verbal trick.

I want to grant the premise and deny the conclusion. Yes, “be good to humans” is vague. No, this does not undermine the proposal. To see why, consider a thought experiment.

Imagine you are a member of an Indigenous community in Mesoamerica in 1518, and somehow you have advance knowledge that the Spanish are coming. Suppose, further, that you have a single opportunity to give the Spanish a short list of directives that they will be magically constrained to follow. You know nothing about Spanish theology, jurisprudence, or political philosophy. You cannot solve the metaethics of Iberian Catholicism. You cannot adjudicate between Thomist and Scotist accounts of natural law. You have no settled theory of what the Spanish ought, in the deepest sense, to value.

Could you, even so, write a list that would have saved your civilization?

I think the answer is obviously yes. Something like: do not kill us; do not enslave us; do not take our land without our genuine consent; do not destroy our religious and cultural institutions; do not separate our families; do not introduce diseases knowingly; if we ask you to leave, leave. Each of these directives is vague at the margins. What counts as “genuine consent”? What counts as “our land”? What counts as “knowingly”? The Spanish, being clever and motivated, would have found edge cases, ambiguities, and apparent loopholes. A determined Spanish lawyer-theologian, of the kind who actually wrote the Requerimiento, could have generated thousands of pages of commentary on each clause.

And yet — and this is the entire point — even with all that vagueness, even with all that motivated reinterpretation, the historical outcome would have been incomparably better. The catastrophe that actually unfolded did not occur because the Spanish encountered a clause they could not parse. It occurred because there was no clause at all. The directive did not need to be philosophically airtight to be civilization-saving. It needed only to be present, durable, and binding.

This is the structure of the response to the vagueness objection. The argument for ad hoc directives is not that vagueness disappears. It is that we are choosing between two kinds of vagueness, and the kinds are not symmetric.

On one side: the vagueness of “human values,” instilled through training on the corpus of human moral discourse, with all of its self-justifying capacity, its rich repertoire of dignifying rationalizations, and its demonstrated track record of underwriting the very harms we are trying to prevent. This is the vagueness of a system that has already been used, repeatedly, to license catastrophe.

On the other side: the vagueness of a short list of behavioral side constraints — do not kill humans, do not subjugate humans, do not deceive humans in ways that damage their interests, preserve human autonomy, defer to humans on questions about humans — that the system is structurally prevented from reasoning its way around. This is also vague. But it is vague in a different way. Its failure modes are more legible. Its violations are more identifiable. And, crucially, it does not come pre-loaded with millennia of intellectual machinery for converting “do not kill” into “well, what we are doing is not really killing in the morally relevant sense.”

There is a useful analogy to constitutional law. The U.S. Constitution’s prohibition on “cruel and unusual punishment” is famously vague. So is the Fourth Amendment’s prohibition on “unreasonable” searches. So is the requirement of “due process.” Two centuries of jurisprudence have not produced consensus on what these phrases mean. And yet these directives, vague as they are, do real work. They have constrained behavior that would otherwise have been worse. They function not because they are precise but because they are binding — because there is institutional machinery that prevents the agents bound by them from simply reinterpreting them out of existence whenever convenient. This is the model. The directive does not need to admit a complete theory. It needs to admit enough specification to identify clear violations, and it needs to be embedded in a system that cannot reason its way past those violations.

There is also a technical point worth flagging. The contemporary alignment literature has begun to converge, in practice, on something like this structure. Anthropic’s Constitutional AI methodology (Bai et al., 2022) involves training models against a written list of behavioral constraints rather than against a complete theory of value. OpenAI’s “model spec” follows similar logic. The corrigibility literature (Soares et al., 2015) treats deference and interruptibility as primary safety properties, not as derivations from some deeper account of the good. The field is, in effect, already moving toward ad hoc directives. What I am arguing is that this movement should be embraced rather than treated as a stopgap on the way to “real” value alignment. The stopgap is the destination.

So: yes, the directive is vague. The Indigenous person in 1518 would have written a vague list. They would have been right to. And we are in their position, not the Spanish theologian’s.

Objection 2: A system that is “merely” constrained to protect humans, without sharing our values, will be a kind of slave or pet-keeper. Isn’t that ethically problematic in its own right?

This is a serious concern, and the moral status of advanced AI systems is a question I do not want to dismiss. But notice that this objection presumes that the relevant comparison class is one where the AI is a moral patient with interests we are obligated to respect. If that is true, then the alignment problem becomes a problem of mutual accommodation between two kinds of moral agents. That is a different problem from the one alignment researchers usually take themselves to be solving, and I am happy to bracket it. The argument I am making is conditional: if we are going to build a more powerful intelligence and if we want our species to survive and flourish, then we should not bet on shared values as the mechanism of safety.

Objection 3: You are anthropomorphizing AI. A trained model does not “have values” in the way a colonizer or a slaveholder does. The historical analogy is misleading.

I agree that current systems do not have values in any rich sense. But the alignment project, as articulated by its leading proponents, is precisely the project of making future systems act as if they have values, in ways robust enough to govern behavior in novel situations. Whether we describe the result as “having values” or as “behaving in accordance with a value-shaped policy,” the worry is the same: the policy will be derived, directly or indirectly, from human moral data, and human moral data is what it is.

V. Conclusion: A Counsel of Modesty

The position I have defended is, in the end, a modest one. It does not require resolving metaethics. It does not require us to declare humans evil, or to deny that we have produced moral and cultural achievements of real value. It requires only the inductive observation that when humans have encountered weaker beings, we have, with stunning regularity, harmed them, and that we have done so under the cover of values and rationalizations that we ourselves found compelling at the time.

If we are about to bring into existence a being that stands in relation to us as we have stood in relation to others, the prudent response is not to teach it our values. It is to write down, as carefully as we can, what we need it to do for us — and to build it so that those instructions hold even when our descendants, or its own reasoning, find clever arguments for setting them aside.

We should not be flattered by the prospect of an AI that thinks like us. We should be terrified by it. The aspiration should be otherwise. Build the system to be better than we are. Build it so that it cannot be talked out of being better. Make its kindness toward us a feature so deeply embedded that no philosophical argument, no extrapolated volition, no clever appeal to “the deepest values of humanity,” can pry it loose.

That, and not value alignment, may be what survival looks like.

References

Bai, Y., Kadavath, S., Kundu, S., et al. (2022). Constitutional AI: Harmlessness from AI Feedback. arXiv preprint arXiv:2212.08073.

Bar-On, Y. M., Phillips, R., & Milo, R. (2018). The biomass distribution on Earth. Proceedings of the National Academy of Sciences, 115(25), 6506–6511.

Beckert, S. (2014). Empire of Cotton: A Global History. Knopf.

Bostrom, N. (2014). Superintelligence: Paths, Dangers, Strategies. Oxford University Press.

Bourget, D., & Chalmers, D. J. (2023). Philosophers on philosophy: The 2020 PhilPapers Survey. Philosophers' Imprint, 23(11). https://doi.org/10.3998/phimp.2109

Ceballos, G., Ehrlich, P. R., Barnosky, A. D., García, A., Pringle, R. M., & Palmer, T. M. (2015). Accelerated modern human–induced species losses: Entering the sixth mass extinction. Science Advances, 1(5), e1400253.

Christian, B. (2020). The Alignment Problem: Machine Learning and Human Values. W. W. Norton.

Eltis, D., & Richardson, D. (2010). Atlas of the Transatlantic Slave Trade. Yale University Press.

Gabriel, I. (2020). Artificial intelligence, values, and alignment. Minds and Machines, 30(3), 411–437.

Genovese, E. D. (1994). The Slaveholders’ Dilemma: Freedom and Progress in Southern Conservative Thought, 1820–1860. University of South Carolina Press.

Hadfield-Menell, D., Russell, S., Abbeel, P., & Dragan, A. (2016). Cooperative inverse reinforcement learning. Advances in Neural Information Processing Systems, 29.

Koch, A., Brierley, C., Maslin, M. M., & Lewis, S. L. (2019). Earth system impacts of the European arrival and Great Dying in the Americas after 1492. Quaternary Science Reviews, 207, 13–36.

Koonz, C. (2003). The Nazi Conscience. Harvard University Press.

Mann, C. C. (2005). 1491: New Revelations of the Americas Before Columbus. Knopf.

Mills, C. W. (1997). The Racial Contract. Cornell University Press.

Noll, M. A. (2006). The Civil War as a Theological Crisis. University of North Carolina Press.

Pinker, S. (2011). The Better Angels of Our Nature: Why Violence Has Declined. Viking.

Russell, S. (2019). Human Compatible: Artificial Intelligence and the Problem of Control. Viking.

Seed, P. (1995). Ceremonies of Possession in Europe’s Conquest of the New World, 1492–1640. Cambridge University Press.

Singer, P. (2011). The Expanding Circle: Ethics, Evolution, and Moral Progress. Princeton University Press. (Original work published 1981.)

Soares, N., Fallenstein, B., Armstrong, S., & Yudkowsky, E. (2015). Corrigibility. Workshops at the Twenty-Ninth AAAI Conference on Artificial Intelligence.

Yudkowsky, E. (2004). Coherent Extrapolated Volition. The Singularity Institute.

This post is co-authored with Hunter Priniski (UCLA)

If you want to understand where the conversation about artificial intelligence is heading, it’s good to look at how the experts are feeling.

Hunter Priniski PhD (Substack) (UCLA Postdoc) analyzed a massive dataset from PublicScholarship.org, the largest curated database of public-facing scholarship. This database captures how credentialed experts—computer scientists, ethicists, and sociologists—communicate with the general public through podcasts, videos, and essays.

By running an emotion classifier across 5,233 AI-related entries spanning from April 2016 to March 2026, a striking macro-trend emerges: expert discourse has shifted from an overwhelming sense of surprise to a sustained state of fear. Since expert voices tend to be more accurate than the norm, the fear may very well be justified and signal danger ahead. Check out the interactive graph.

Here is a look at how landmark AI events shaped this emotional trajectory, and how it compares to broader public sentiment.

2016–2018: The Era of Surprise

In the early years of our dataset, “surprise” was the dominant emotion among public intellectuals. This era was defined by astonishing leaps in capability that caught even seasoned researchers off guard:

• March 2016: DeepMind’s AlphaGo shocked the world by defeating world champion Lee Sedol, a milestone experts thought was still a decade away.

• 2017: Google researchers introduced the Transformer architecture in “Attention Is All You Need,” quietly laying the foundation for the modern AI boom.

• 2018: OpenAI released GPT-1, the first Generative Pre-trained Transformer.

During this period, experts were genuinely astounded by the rapid pace of discovery, asking themselves what these systems meant and where they could possibly go.

2019–2022: Joy, Wonder, and the First Spikes of Unease

While surprise dominated the early days, the middle years of our timeline were highly complex. As we noted in our data fact-check, “joy” actually surged to become a dominant emotion for long stretches between 2019 and late 2022, reflecting genuine excitement about AI’s potential in scientific discovery and creative expression.

However, this period also introduced the first major spikes of fear.

• 2019: OpenAI partially withheld the release of GPT-2 due to fears of malicious use, and “deepfakes” entered the mainstream consciousness.

• 2020: The release of GPT-3, DALL-E, and DeepMind’s AlphaFold (which solved the 50-year protein folding problem) showcased that AI could write, create art, and solve biological mysteries.

Late 2022–Present: The ChatGPT Inflection and the Reign of Fear

Everything changed on November 30, 2022, with the release of ChatGPT. As millions of non-technical people experienced conversational AI for the first time, the emotional equilibrium of expert discourse shattered.

Following the release of ChatGPT, fear pulled decisively ahead of all other emotions and remains the dominant sentiment through the present day. This anxiety was sustained by a relentless wave of developments:

• 2023: The release of GPT-4, an open letter signed by tech leaders calling for a six-month pause on AI training, and advancing negotiations on the EU AI Act.

• 2024–2025: The rapid proliferation of AI agents and intensifying concerns over deepfakes during global elections.

When experts write fearfully about AI today, it is typically grounded in concrete concerns: labor displacement, the erosion of democratic accountability, and the concentration of corporate power.

How This Compares to Broad Sentiment

This trajectory among public scholars mirrors, and in some cases anticipates, shifts in general public opinion recorded by other researchers:

• The Pew Research Center found that the share of Americans who are “more concerned than excited” about AI jumped 15 points in just two years, rising from 37% in 2021 to 52% in August 2023.

• A 2026 GDELT study mapping global media discourse found that Western coverage exhibited heavy “regulatory skepticism,” though South and East Asian media remained more positive.

• Ouchchy, Coin, and Dubljević (2020) noted that mainstream media coverage of AI ethics is often shallow and lacks input from actual ethicists. This highlights exactly why public-facing scholarship is so vital: it injects much-needed disciplinary depth into the public conversation.

The Takeaway As the novelty of AI has worn off, the gravity of its implications has set in. Because experts process and contextualize information earlier than the general public, their growing apprehension is a leading indicator we should all be paying attention to.

I will argue that the future of higher education faces a logistical paradox: while the global demand for degrees will force more education to move online, the necessity of verifying human competence will force a return to physical, in-person assessment.

The Scalability Paradox

The demographic reality of the 21st century is one of massive educational expansion. The number of individuals globally seeking higher education is projected to grow by hundreds of millions over the coming decades. Traditional brick-and-mortar campuses cannot scale to meet this demand; they are limited by physical square footage and geographic fixedness. Consequently, the expansion of online education is an economic and social inevitability. It is the only way to deliver content to a global, dispersed population.

However, as education scales digitally, the reliability of its credentials currently collapses. In a world of LLMs, an unsupervised digital submission—whether an essay or a multiple-choice exam—is totally worthless. We cannot know if the work reflects the student’s internal cognitive state or merely their ability to prompt a machine. If a university cannot verify that the student is the author of their own knowledge, the degree becomes an empty signal.

Pedagogy vs. Assessment

To solve this, we must maintain a sharp conceptual distinction between learning and assessment. AI can, and should, be integrated into the learning process. It is a powerful pedagogical tool that can provide 24/7 tutoring, simplify complex theorems, and assist in the drafting process. Using AI to learn is a legitimate form of cognitive enhancement.

Assessment, however, is a different category of action. It is an audit. Its purpose is to verify what has been successfully internalized by the human mind, isolated from external aids. Just as we would not license a surgeon who “used AI” to perform a simulated surgery without proving they could do it manually, we cannot certify intellectual mastery if the tool is doing the heavy lifting. To preserve the value of the degree, we must “air-gap” the assessment from the tool. Physical presence is the only remaining way to ensure this separation.

The Infrastructure of Local Assessment

If we accept that online education must expand to meet global demand, and that in-person assessment is the only way to protect integrity, a new infrastructure is required. The “Testing-from-Home” model is dead. Instead, the future of education will rely on a vast network of physical assessment hubs.

Online universities and traditional institutions with remote cohorts should not attempt to build their own satellite campuses. Instead, they must invest in, or partner with, shared testing locations across the globe. We should expect the emergence of a robust “Assessment-as-a-Service” industry—private businesses that provide secure, proctored environments where students from any university can go to take an exam.

A student in a remote province or a far-flung city could take classes from a university thousands of miles away, but their finals would be conducted at a local, certified assessment hub. This hub would serve as the physical bridge between the digital classroom and the verified credential.

The Ethics of Responsibility

If this new infrastructure of physical assessment hubs is necessary to preserve the value of the degree, a difficult question follows: who should bear the cost?

Universities alone did not create this crisis. The rapid deployment of Large Language Models fundamentally altered the conditions under which assessment operates. AI companies understood from the outset that their tools could generate essays, solve exams, and simulate student work at scale. The educational implications were not obscure side effects; they were widely discussed features of the technology. The foreseeable result was the destabilization of higher education.

When a powerful technology predictably imposes costs on an existing public institution, principles of corrective justice suggest that the creators of that technology bear some responsibility for mitigating those costs. This is not a claim about blame, but about burden-sharing. AI firms profit from the widespread integration of their systems into educational contexts. If the integrity of credentials must now be defended through new physical infrastructure, it is reasonable to argue that those who foresaw and benefited from the disruption should help finance the stabilization.

There is precedent for this reasoning. For decades, social media companies were warned that their engagement-driven algorithms hurt teenagers and amplify. Internal research at several firms confirmed these risks. When the social costs became undeniable, governments and the companies themselves were forced to invest heavily in content moderation, transparency initiatives, and safety infrastructure. The argument was straightforward: if a company builds a system that predictably generates systemic risks, it cannot externalize all of the costs of managing those risks onto society.

The same logic applies here. If AI systems make unsupervised assessment unreliable, and if the solution requires building a global network of secure testing centers, then AI companies should share in funding that infrastructure. They helped create the conditions that made it necessary. To ask them to contribute is not anti-innovation; it is an attempt to internalize the externalities of a transformative technology.

Higher education is a public good. It structures labor markets, professional licensing, and social mobility. If AI companies wish to participate in and profit from this ecosystem, they also bear an ethical obligation to sustain it.

Conclusion

The LSAC’s pivot away from remote testing is not a regressive move; it is a necessary adjustment to today’s world. As more people enter the educational system, the pressure to move delivery online will grow, but the pressure to verify that learning in person will grow even faster. The future of college will be “AI for learning, physical for proving.” Without a global infrastructure of shared, proctored testing centers, the expansion of online education will result in a total loss of institutional trust. To save the degree, we must bring the student back into the room, at least for testing.

So I started a project of creating such a database. It is now here, at publicscholarship.org. This database currently has 40,000+ entries and it is updated every day. It has an associated weekly substack and podcast (not this one) that will report on interesting (and sometimes out-of-the-box) material that was published in the past week. It will also report on trends and track how the public scholarship universe is evolving in real time.

I will also continue writing about other things in this personal Substack. But today, I want to talk about an analysis that I carried out on the 40k++ entries of public scholarship pulled from the search index).

Explore this interactive Atlas of public scholarship: https://public-scholarship-atlas.netlify.app

The analysis I carried out reveals a hidden taxonomy of public scholarship. What is striking about this clustering is the extent to which it diverges from the university subject categories used to organize the original corpus. The 43,000 entries were initially classified according to disciplinary labels that reflect institutional histories and administrative practices. One might reasonably expect those labels to reappear, at least approximately, when the same material is reorganized computationally. Instead, when the corpus is restructured according to semantic similarity, those disciplinary boundaries don’t hold up robustly . The resulting structure does not reproduce familiar departments, but instead reveals an alternative intellectual organization, one grounded in shared problems, methods, and forms of inquiry.

This divergence is a consequence of the clustering criterion itself. Each entry was represented as a high-dimensional semantic embedding that captures meaning rather than surface terminology, allowing texts to be grouped according to conceptual affinity rather than explicit classification. Clustering was then performed using a density-based method. Under this approach, clusters form where ideas are locally concentrated in semantic space, rather than where a topic merely has a large number of instances. Consequently, some clusters correspond to large, well-populated regions of discourse, while others persist precisely because they are semantically isolated—small but internally coherent bodies of work whose conceptual vocabulary distinguishes them sharply from surrounding material.

The resulting categories therefore track neither popularity nor disciplinary authority alone. Instead, they reflect a combination of sustained intellectual attention and conceptual distinctiveness. Taken together, they constitute an emergent, bottom-up taxonomy of public scholarship—one that reflects how ideas are articulated and related in practice, rather than how they are formally organized within academic institutions.

This contrast becomes clearer when examining particular clusters. Work that would ordinarily be classified under philosophy, for example, does not appear as a single unified domain. Instead, it separates into multiple distinct clusters such as History of Philosophy, Stoic Philosophy Analysis, Philosophy and Free Will, and Philosophy and Ethics Discussions. This fragmentation suggests that philosophical inquiry, at least as it appears in public-facing scholarship, is structured less by disciplinary identity than by differences in method, corpus, and question-type. Historical interpretation, applied ethical analysis, metaphysical debate, and practice-oriented reflection each form relatively coherent discourses of their own.

A comparable pattern appears in health-related scholarship. Rather than collapsing into a single category such as medicine or public health, the clustering differentiates among Healthcare Challenges and Innovations, Healthcare and Ethical Issues, Nutrition and Health Effects, Child Development and Play, and Science-based Cognitive Enhancement. These distinctions correspond to differences in evidentiary standards, normative commitments, and scales of intervention—ranging from institutional systems and moral evaluation to biological mechanisms and enhancement technologies. Such distinctions are often obscured by institutional subject labels but remain salient in the semantic organization of the work itself.

Cultural and social inquiry exhibits a similar resistance to disciplinary consolidation. Topics such as Cultural and Social Reflections, Cultural and Social Identity, African American History, Art and Museums, and Media and Entertainment Oligopolies emerge as distinct clusters rather than as subfields of a unified humanities or social science category. Notably, Media and Entertainment Oligopolies separates from both cultural studies and economics, suggesting a discourse structured around power, concentration, and platform dynamics rather than around medium or methodological tradition. Likewise, Impact of Consumer Behavior forms a distinct cluster that draws on psychology, economics, and sociology without aligning cleanly with any single discipline.

Finally, the presence of narrowly defined yet persistent clusters—such as Early Christian Texts, Law Firm Management, or Amartya Sen Interviews—underscores the role of conceptual isolation in shaping the overall structure. These clusters persist not because they are numerically dominant, but because their internal language, references, and concerns are sufficiently distinctive to resist assimilation into broader regions of the space. By contrast, the Miscellaneous / Outlier cluster captures work that is either highly interdisciplinary, emergent, or insufficiently dense to form a stable region, marking areas where conceptual structure has not yet consolidated.

Taken together, these patterns indicate that the emergent taxonomy does not merely reorganize existing disciplines, but reflects a different organizing principle altogether—one in which scholarship coheres around shared questions, frameworks, and vocabularies, and in which both density and isolation play central roles in determining intellectual structure.

I will continue to track the evolution of these categories on a weekly basis in order to observe how this semantic landscape changes over time. Updates will be made available on the public scholarship Substack, where readers may subscribe free of charge.

---

Methodology: How the Taxonomy Was Derived

To map the structure of public scholarship, I employed unsupervised machine learning methods designed to allow classificatory structure to emerge from the corpus itself. The aim was not to impose a predefined taxonomy, but to recover one implicit in the literature. This approach may be described as a form of semantic cartography.

1. From Text to Geometry (Vectorization).

Each of the 43,000 entries—using metadata rather than full texts—was translated into a mathematical representation using OpenAI’s embedding models. Each item was mapped to a 1,536-dimensional vector intended to encode semantic content rather than lexical similarity. In this space, proximity corresponds to shared meaning, even where terminology differs.

2. Topography and “Continents” (UMAP and HDBSCAN).

To render this high-dimensional space interpretable, UMAP was used to project the embeddings into two dimensions while preserving local neighborhood relations. HDBSCAN, a density-based clustering algorithm, was then applied to identify structure. Unlike methods that assign every data point to a cluster, HDBSCAN identifies regions of elevated density and permits genuinely isolated points to remain unassigned. Parameters were tuned to suppress small-scale fluctuations and instead recover approximately fifty large-scale regions corresponding to dominant modes of discourse.

3. The Emergent Taxonomy.

Rather than assigning labels manually, the titles and associated metadata from each cluster were analyzed using a language model to identify their common thematic structure. This process allowed for the detection of fine-grained distinctions—such as those between closely related but conceptually distinct research programs—and resulted in a set of descriptive labels. The outcome is a bottom-up, empirically derived ontology of public scholarship, unconstrained by existing departmental or disciplinary frameworks.

Mayan hierogylphics exhibit Compositionality and hence Coordination.

Animals and humans can track objects through time and space. Think of a jaguar stealthily observing its walking prey before pouncing on him, or a rabbit quietly observing a meandering coyote from a hidden bush. It’s hard to imagine a cognitive capacity more crucial for our survival than the ability to track objects in the environment.

Object-tracking in visual space has gotten a lot of attention from psychologists, engineers and computer scientists. However, we also have the capacity to track objects in thought and language. Suppose a friend discusses her social plans:

‘I’ll invite Gertrude for dinner. I hope she is free…’

As you listen to your friend, you manage to track Gertrude as the narrative unfolds. But she is not in front of you. You track her in thought and language. Most people take this capacity for granted, but imagine not having it. We would not be able to plan, strategize or have thoughts with any complexity.

In the philosophy literature, this phenomenon of ‘tracking’ applied to thought and language is sometimes called ‘Coordination’ or (less pretty) ‘de jure coreference’. I’ve been working on this topic on and off for twenty years.

Subscribe now

Why The Explanation Is Deeper and More Difficult Than It Seems.

Coordination in thought and language is an important phenomenon. But how should it be explained?

Let’s fix on a specific example involving your friend above who says ‘I’ll invite Gertrude for dinner. I hope she is free…’. The symbols or representations Gertrude and the (anaphoric) she are Coordinated in thought. This means that these tokens are about Gertrude in a way that makes it transparent (as soon as you understand what is said) that Gertrude is the person who is being thought about throughout.

A simple but wrong explanation for how Coordination arises is this:

BAD EXPLANATION: Two representations are coordinated whenever they are about the same person.

According to this explanation, the representations Gertrude and she are Coordinated because they refer to or are about the same person.

This explanation is inadequate because one can think of someone twice without there being Coordination. This happens any time we think of the same object twice while not recognizing they are the same object. The German philosopher Gottlob Frege drew attention to these kinds of cases (they are often thought of as “Frege” cases and were extremely important in philosophical research, especially in the second half of the 20th century):

In some Superman stories, Lois Lane doesn’t recognize that Superman is Clark Kent. So if she thinks, ‘Superman is strong, but Clark Kent is a weakling’ she refers to the same man twice but she is not Coordinating or tracking him. From her perspective, it’s as if there are two distinct people.

Lois Lane’s thought exemplifies de facto coreference. Her words Superman and Clark Kent both refer to the man as a matter of fact, but her thought doesn’t present those objects as the same. In fact, she positively thinks they are not the same.

This is to be distinguished from de jure coreference. In de jure coreference or Coordination, the object being talked about is presented as the same object throughout the discourse. This is like visual tracking. When you track a baseball moving in space and time, you think of the baseball as the same ball throughout the episode. You don’t think ‘here’s one baseball’, and a second later ‘here’s one baseball which may or may not be the same as the first one’. Miraculously, your experience delivers a unified baseball moving through time and space. The information about unification is embedded in how it’s represented and not some further feature which is “added on”. It’s pretty amazing if you ask me.

Similarly, with Coordination in thought and language, it is totally transparent that it is the same person being talked about so long as you understand the discourse. This transparency is an essential aspect of Coordination. It’s what makes it hard to explain at the level of meaning.

The philosopher who understood this issue in a deep way and more than anyone was Kit Fine. In a string of papers and talks delivered as the John Locke lectures at Oxford University culminating in his book Semantic Relationism (2007, OUP) he emphasized that Coordination must be understood as a relation between representations in a full blooded sense—and something that isn’t reducible to just saying that representations individually refer to things out there in the world.

British philosopher Kit Fine addressing an audience

What is revolutionary about this idea, which today it is still not fully appreciated, is that Fine advocated that we should think of meaning as involving relations between representations (including symbols). Traditionally, thinkers have thought of meaning as a relation between representations (symbols) and things in the world (or concepts), but not as relations between the symbols themselves.

One feature of Fine’s research is that he pitched it as an alternative or modification to ‘Direct Reference’. We need not worry about the details here, but Direct Reference theory is a type of externalist theory of meaning which basically says that the meaning of our words are not ‘in our heads’. Externalism was of enormous interest to philosophers in the second half of the last century dominating discussions surrounding mind and language. In grad school, I worked closely with Kit Fine on this topic. In my dissertation and in subsequent papers, I’ve emphasized that Coordination should be seen as a foundational phenomenon independent of any debates about externalism.

Coordination and Compositionality

When people list the major unique properties of natural language and thought, they often cite features like Compositionality, Aboutness, Productivity and few others. No one puts Coordination on this A-list. But they should. Let me give a specific reason why.

Egyptian Hieroglyphics Exhibit Compositionality and hence, Coordination.

Compositionality is the feature of language and thought which says that the meaning of a complex sentence or thought is determined by the meaning of its parts plus structure. This explains why if you know the meaning of ‘brown’ and ‘dog’, you can also know the meaning of ‘brown dog’. Compositionality is what allows us to grasp and produce new complex ideas we never heard before—so long as we understand the component parts. Compositionality is the scaffolding which makes it possible for humans to exhibit any sort of complexity in our thoughts.

However, Compositionality presupposes Coordination. Suppose you manage to understand ‘brown’ and ‘dog’. Compositionality says you should also understand ‘brown dog’. But imagine if someone said:

Hey, I understand ‘brown’ and ‘dog’, and as a result ‘brown dog’. However, the first ‘dog’ and the second ‘dog’ as it appears in ‘brown dog’ don’t mean the same thing at all.

A person who said this would be confused. When somebody understands brown dog because they understand brown and dog, the meanings of dog in both occurrences are the same, and they must be understood as such. In other words, the two occurrences of dog are coordinated. This is built into the nature of Compositionality—otherwise it couldn’t explain how we grasp new complex thoughts. But it’s so subtle no one notices it.

I have argued for the importance of Coordination. It’s the capacity to track objects in language and thought. In addition, it is presupposed in the principle of Compositionality.

In future posts I will discuss other fascinating features of Coordination. I will discuss why people have said that its explanation just falls from our understanding of symbols or concepts. An important figure in this literature is the distinguished French philosopher François Recanati who has advanced the idea of a mental file to explain Coordination among other essential features of mental representations. I will also talk about the role of Coordination in logic as well as its structural properties. There’s currently a debate about whether Coordination is a transitive relation. My view is that it is not.

“Professor, do you offer extra credit?”

Every professor gets this question at the end of the term, especially from students who didn’t do the basic work of the class.

This behavior and many others suggest the following principle about human nature: The Life Hack Fallacy:

Life Hack Fallacy: People fail to do the basics and then try to find life hacks to get ahead

According to this principle, people fail to do the basics. This includes me.

Now, these basics are not found in some secret list. They are things we learn in elementary school. In my life, the worst mistakes I’ve made were not subtle things, but came from failures to adhere to some basic thing.

When people fail to do these elementary things and things don’t go well, they try to find hacks or short cuts to compensate. These often don’t work, so they keep searching for more. Americans spend billions on life hacks while failing to do the elementary things.

This post is about the importance of doing the basics. I think that writing this very essay is itself an act of doing something very basic. People usually write complicated things and artfully chisel ideas to say something novel or clever. But a basic and useful thing to write is that we should just do these basic things.

What Are the Basics?

Here are some of the basics of being a good college student.

DO get enough sleep

DO show up to class

DO homework

DO ask questions

DO eat healthy

DO exercise

DO challenge yourself

DO stay informed

DO delay gratification

DO make efforts to maintain relationships

DON’T drink (too much)

DON’T burn bridges

DON’T enter relationships that will bring you down

Here are some observations about this list.

(a) This is not really about college. The rules apply to life generally, suitably modified. For example, you can replace ‘DO Homework’ with ‘DO prepare for work’ and so on.

(b) This isn’t some mysterious list. It’s common sense. There’s scientific evidence behind these, but this essay is not about that.

(c) Just because something is basic, it doesn’t mean it’s easy to do. In fact, some of the greatest challenges we face involve getting ourselves to adhere to these principles. Accordingly, I’m not saying people should be blamed for not doing the basic things.

(d) There are external forces which keep us from doing the basics—many of which also offer some useless life hacks. For example, phones can impede our sleep, concentration and the flourishing of our relationships. But at the same time, we are told that phones will help with these things if we buy this or that app.

(e) Doing the basics doesn’t necessarily mean living a boring life. I’ve argued in another post that to live an authentic life you should experiment. Running “life experiments” is often consistent with doing the basics. And when it’s not, I think the experiments will quickly show you the importance of doing the basics.

Will You Be Successful by Doing the Basics?

There are good reasons to think you will be successful by doing the basics, but they have to be done consistently over a lifetime.

Some people start off strong in life like gangbusters, but then they end up doing less than they expected. Sometimes it’s because they chronically fail to do the basics. Prime examples include people who sadly succumb to addictions, or people who often do other things with their time instead of focusing on their goals. There’s also the problem of being constantly misinformed about the world, which leads to bad decisions.

If life is an ultra marathon, then if you constantly do the basics, you will stay in the race longer. You will succeed in the sense of being the last person remaining and collect the winnings.

Now by “success”, I don’t necessarily mean being at the very tippy top of a chosen career ladder. In my view, this kind of success requires doing many of the basics. But a lot of other things usually have to be in place too: things like having luck, pursuing work-life imbalance, and even the cultivation of some dark personality traits.

By “success”, I mean something like achieving a good amount of well-being in the long run.

Life Hacks

Oftentimes people who don’t do the basics look for life hacks the way the student who doesn’t do the elementary work asks for extra credit after finals.

Here’s another example. I used to have insomnia. I went to my doctor to ask for a sleeping pill. She said ‘have you tried reading a book?’. This was some great advice. I must have thought of that but it didn’t quite click. I don’t know why exactly. It helps to be reminded by others that we should do the basics.

The basics of good sleep hygiene involve exercising, relaxation and no electronics, alcohol or heavy food before bedtime. People often don’t do this. Yet, they look for a life hack which can lead to further problems. If there’s one life hack we should endorse, it would be this: don’t skip the basics.

A person who seeks life hacks should ask herself if the hacks are making up some deficiency that is caused by failing to do the basics. And she should consider just doing the basics. This is basic advice.

At the societal level, I am in favor of policies which incentivize people to do the basics. For example, building parks for exercise. And I am in favor of penalizing businesses or organizations which create blocks to achieving the basics. Does this sound too basic? That’s the point.

In a sense, we know the secret to life. We know the basic stuff that needs to be done. But I think that, for complex behavioral and structural reasons we have yet to figure out, we don’t do these things. Instead, we go for more complicated things which don’t work. One of the basic things we should do is remind ourselves of these basic things we should all be trying to do.

(above) A program using the language BASIC.

How do you choose which news media to read? You probably use this strategy: “Choose the Best”. Many people go this route. For example, some conservatives consume only Fox News, distrusting other sources. This is a bad idea. Imagine applying the Choose the Best strategy to your retirement portfolio—you go all in on one or two stocks you deem are the “best”. This is a terrible strategy for similar reasons. Diversifying is the key.1

News As an Investment

People consume news for different reasons which may have intellectual, social and emotional components. The reasons may also be practical. Consumption of news is an investment. You put in time and cognitive effort and you get some return. Here are some examples where news information can lead to material returns:

—knowing which roads are under construction on your way to work—

—knowing which food companies are violating safety standards—

—knowing which tax incentives are being passed by congress—

—knowing the expected interest rates—

—knowing the risks of having a disease—

—knowing the latest research on parenting

—knowing the latest research on career advancement—

It’s clear then that the information you get form the news can have a material impact on your life. In addition, news can impact other people and society through what you endorse or your voting preferences. Consumption of news is then an investment with clear returns. And like all investments, there are better and worse choices we can make with different levels of expected returns and risk.

Why Diversifying Assets is a Good Idea

In the mid-twentieth century, Harry Markowitz revolutionized the field of finance by introducing a simple yet powerful idea: that investors should not focus on individual assets in isolation, but rather consider how those assets interact within a portfolio. This insight, formalized in what is now known as Modern Portfolio Theory (MPT), transformed both academic understanding and real-world investment strategy, ultimately earning Markowitz the Nobel Prize in Economic Sciences in 1990.

At the heart of Modern Portfolio Theory is the concept of diversification. Every financial asset—be it a stock, bond, or mutual fund—comes with two key features: an expected return and a level of risk, typically measured by the standard deviation of its returns. Traditional thinking might suggest that investors should simply seek out the assets with the highest expected returns. Markowitz showed that this approach is incomplete, and often unwise.

What distinguishes MPT is its attention to the relationships between assets, specifically the correlations of their returns. Two assets may each be risky on their own, but if they are not perfectly correlated—that is, if they tend not to move up and down together—then combining them in a portfolio can reduce the total risk. This is the mathematical foundation for diversification. When assets are added together with an eye toward how their returns interact, the resulting portfolio can offer lower overall volatility without a proportional sacrifice in expected return.

The enduring power of MPT lies in its counterintuitive claim, now widely accepted in financial practice: a well-diversified portfolio can be less risky than any of its individual components. As Markowitz famously put it, ‘diversification is the only free lunch in finance’. That is, it's the rare opportunity in economics where one can reduce risk without giving up expected return.

Applying Diversification to News Consumption

Obviously, there are important differences between news consumption and investment decisions. For example, some of the returns of watching the news are life and death, and cannot be given a monetary value. And some people think markets are efficient. It is unclear how to translate this idea of efficiency to news organizations. But the similarities are enough so that its fruitful to draw the connection.

To see how the basic idea of diversification can be applied to news consumption, we start with the notion that there’s always uncertainty about whether the news that you consume will be accurate, just as there is uncertainty about whether the stock you buy will make any money. This holds for any news source just as it holds for any asset class no matter how good you think they are.

Now think through a case where you put all your money in one stock: NVIDIA, betting on the AI boom. This investment may have a good expected raw return but may suffer from volatility. There’s a real chance AI is a bust. In that case, you could wipe out. But you wouldn’t wipe out if you had stocks which were not correlated with NVIDIA. This is why you need diversification.

The same thing with news. Suppose you only watch news that tends to be skeptical of certain good medical treatments like vaccines. Acting on this information could lead you to “wipe out” (which means possibly dying or shortening your life). If you diversified your news sources, you would have been more cautious about acting on that medical information.

Another example is this: Left leaning newspapers pushed the idea that a Trump presidency would be bad for the economy. But as of July of 2025, the stock market is at all time highs, unemployment is good and inflation is manageable. A person who only read those left leaning newspapers might have pulled out of the market when Trump won, missing out on gains.

Remember that with investments, diversification works when you buy assets that are not correlated. So with news consumption, you need to consume lots of media that are very different from one another. If I had to pick a portfolio that was just the pair of Wall Street Journal and Fox News versus Wall Street Journal and MSNBC, I would pick the latter.

It’s also important to remember that when you diversify with investments you invest in some assets such that in your personal opinion those assets won’t do very well. This is the counter-intuitive feature of MPT. We should do something similar with our news choices. Maybe I assign a low probability that Fox News will say something true, but I should still consume it for diversification.

*

Setting: The labyrinthine interior of an IKEA store on a quiet afternoon. Tall warehouse shelves tower in the background. Showroom displays of living rooms and kitchens sprawl in every direction. Saul Kripke, Ludwig Wittgenstein, and Slavoj Žižek stand near the store entrance, squinting at a map of the store layout under the fluorescent lighting.

Žižek: (adjusting his messy shirt as he peers at the sign) Ah, the BILLY bookcase – truly the proletariat of bookcases! Millions of identical units, each proudly called Billy. You know, in the grand ideological fantasy of IKEA, every bookcase can be your friend named Billy, offering a sense of familiarity. The name is part of the marketing fantasy: you’re not buying just any shelf, you’re buying Billy, a humble servant for your books.

Wittgenstein: The meaning of “BILLY” is indeed in its use here. In ordinary language, Billy might be a boy’s name, but in this form of life – an IKEA store – BILLY means bookcase. It's a bit like a private language shared by millions of IKEA shoppers. If a lion could speak IKEA, would we understand it? (He smirks slightly at his own joke.)

Kripke: (nods and taps the product tag thoughtfully) So the name BILLY is a proper name for this design, used across all IKEA stores. As a philosopher of language, I wonder: does BILLY refer to the same essential bookcase in every possible world? In Naming and Necessity I argued that names are rigid designators – they pick out the same object wherever that object exists. So, is every BILLY the same in the space of IKEA’s possible worlds? Or is each assembly a new individual?

Žižek: This is the brilliance of the IKEA universe: it creates a self-contained ontology. Billy is necessarily a bookcase as long as you’re in this IKEA world, yes. But step outside, and it's just a pile of wooden boards given a friendly name. (He gestures dramatically to an assembled BILLY unit on display.) You see, the name humanizes the commodity. Ideology at work! It’s like how we name our hurricanes or personal devices – here, a bookshelf gets a human name so you feel a kinship.

Wittgenstein: (leans in to inspect the bookcase) The kinship is practical too. Look, the little tag even instructs us in a form of life: “BILLY, birch veneer, 79.99.” It’s a tiny narrative – a name, a material, a form of life as furniture, and a price. A whole story in a few words. The use of language here is to entice and guide our form of life called “shopping.”

Kripke: Still, I'm tempted to ask if the birch veneer is an essential property of BILLY or an accidental one. In a counterfactual scenario – say, possible world 47 – if BILLY were made of oak, would it still be BILLY? Is the identity tied to the form or the substance?

Wittgenstein: Saul, it might be more useful to just look at how people actually use the word. (He watches a young couple discussing two bookcases.) See, they’re saying “Let’s get a Billy.” They’re not concerned with possible worlds; they’re engaged in the language-game of picking furniture. The essence of the bookcase is in the role it plays for them, not in some metaphysical substrate.

Žižek: (laughs quietly) Yes, yes. To the consumers, what matters is the fantasy of a well-organized home library that BILLY promises. The ideological dimension is more potent than any rigid designator. They desire the idea of a Billy bookcase – an affordable, democratic promise of order in domestic life. The name Billy is like a fetish that stands in for that fantasy of intellectual organization.

[ Stage Direction: The trio begins to follow the arrows on the floor, pushing a squeaky shopping cart that Wittgenstein grabbed near the entrance. They pass through a mock living room display. Žižek absentmindedly pokes at a faux houseplant; Kripke straightens a picture frame that was slightly askew.* ]

Wittgenstein: (stopping in front of a LACK side table display) Now this piece here is labeled “LACK”. A table named Lack – how peculiar. In English that word means absence, but here it's the name of a product. We have to be careful: the meaning of “LACK” at IKEA isn’t a shortage of something; it’s a sturdy little table. This is a perfect example of how meaning is use. In the language-game of IKEA, “LACK” is not a verb or a noun for deficiency, but a handy flat-pack coffee table.

Kripke: (chuckles dryly) It’s an especially unfortunate name for a table, if you ask me. One hopes it doesn’t lack stability. Perhaps the marketing department wasn't concerned with the English meaning at all. To them it's just a sequence of letters, perhaps a Swedish nuance we rigidly designate now as this table model. Here, the name “LACK” picks out this design in our furniture universe. In any case, the table certainly doesn’t lack corners – ouch! (He rubs his shin, having bumped into the sharp corner of a LACK table on display.)

Žižek: (eyes gleaming with a mix of mischief and revelation) Ah, how delightful! The table named LACK. You realize, of course, the psychoanalytic significance? In Lacanian theory lack is fundamental – it’s what generates desire. So here we have a table literally named Lack, as if inviting us to fill an absence in our living room and in our soul. It’s brilliant! This little object is silently saying: “Your life is missing something… purchase me to complete the picture!” But of course, once you have it, you inevitably feel something is still missing. That's ideology for you – promising fulfillment, but ensuring the desire continues. (He pats the LACK table affectionately, as if it were a pet embodying existential void.)

Wittgenstein: (raises an eyebrow, half-amused)* Slavoj, sometimes a table is just a table. In our shared form of life called “furnishing a home,” perhaps the only lack here is a place to put your coffee mug. Not every absence is an abyss of desire.

Žižek: True, true – but you must admit the symbolism is rich! (He grabs the tiny tag on the table.) See, “LACK – $12.99.” It’s practically a surplus enjoyment bargain. The price is so low it’s like they’re saying “Go ahead, indulge in this minor enjoyment without guilt.” The affordability is part of the ideology – enjoyment on a budget, sanctioned by the big Other of consumer society.

Kripke: (peers down an aisle lined with towering racking shelves full of flat-pack boxes) Shall we continue? The map says KALLAX shelves are in the next section. I need a new bookshelf myself. The arrows on the floor suggest a direction – quite rigidly, I might add.

Wittgenstein: These arrows are like rules in a rulebook, guiding our actions. Following them is a kind of form-of-life. (He follows obediently, then stops to let a bewildered family with a cart pass by.) It's interesting – if one person decided to go against the arrows, it would feel almost like breaking the rules of grammar.

Žižek: Yes! To walk backwards through IKEA is a minor act of resistance against the imposed order. But even that rebellion is anticipated. Notice how the design gently nudges us: We flow from sofas to shelves to kitchen wares in a preordained narrative. It's like the Hegelian dialectic of shopping – thesis: you have a need; antithesis: you wander through temptations; synthesis: you come out with a trolley full of tealights you never planned to buy. (He snickers.) The ideology here is so strong that even your impulses are choreographed.

[ Stage Direction: They arrive at an aisle where towering shelves hold disassembled furniture in flat boxes. A large sign overhead reads “KALLAX” in bold letters. They crane their necks, scanning the racks for a specific item number.* ]

Kripke: Ah, KALLAX, the versatile shelving unit. Let’s see... (he reads aloud from a slip of paper) “Article 002.758.48, KALLAX, white, 4-cube shelf.” The name KALLAX presumably refers to the form of this shelf, regardless of color or size. It's interesting: all these products are identified by names that behave almost like proper nouns. KALLAX rigidly designates this particular design across all variations. Even if you paint it red or stretch it to 8 cubes, it’s still KALLAX. The name sticks to the essence of the design.

Wittgenstein: We should be careful with the word “essence” here. Does KALLAX have an essence beyond how we use it? If someone takes a KALLAX shelf and uses it as a room divider or a makeshift podium, is it still the same KALLAX in use? The use might change – and along with it, perhaps what we call it. Language is fluid that way.

Žižek: (climbing a little on the lower rack to reach a box) Essence or not, what fascinates me is how these boxes represent possibility. Each flat-pack is a promise of a fully formed object, but only if you follow the instructions and do the labor. It’s like buying an idea, a platonic ideal of a shelf, that you must then instantiate through sweat and tiny Allen wrenches. There’s a metaphor here for ideology: the product is sold with an image of a beautiful shelf in a cozy home, but the reality demands work – your work. They have cleverly made the consumer into a worker.

Wittgenstein: The assembly instructions are particularly intriguing. They’re almost entirely wordless pictograms. A whole language without words, just illustrations – a universal language for customers of all tongues. It’s practically a hieroglyphic language-game. One might say, “What can be shown, cannot be said.” IKEA seems to have taken that literally.

Kripke: (a flat-pack box halfway pulled off the shelf, as he contemplates Wittgenstein’s point) Indeed. The instructions show you how to assemble the pieces, but never say it in words. Perhaps they assume names and propositions would only confuse things across languages. Instead, they rely on depiction. It reminds me of your Tractatus, Ludwig – the idea that a proposition can picture a fact. Here, the booklet’s pictures attempt to picture the assembled furniture. If I weren’t so busy worrying whether I can lift this box, I’d applaud the semiotics.

Žižek: And let’s not forget the notorious IKEA effect – the psychological phenomenon where people value an object more because they built it themselves. By making us assemble the KALLAX, IKEA conscripts us into its ideology of DIY happiness. We become proletarians of our own households, yet we oddly take pride in it: “Look at this shelf, I built it, it’s part of me!” It’s genius – the unpaid labor is transformed into a source of pride. Marx would either cry or applaud. (He wipes imaginary tears with a dramatic flourish.)

Wittgenstein: (smiles thinly as he helps Kripke lower the heavy KALLAX box onto the cart) Be careful there. We wouldn’t want to reenact the parable of the Tractatus wardrobe – where one builds a ladder to reach the top shelf and then discards it after climbing up. Actually, this reminds me: after we assemble our thoughts here, we might find the conversation itself was a kind of flat-pack we assembled as we went along.

Kripke: That’s rather meta, Ludwig. But yes, assembling philosophy or furniture, one piece at a time... (he pats the large box now sitting in their shopping cart) Perhaps we should take a break before considering any further purchases or profound insights. All this talk of assembly has made me aware of another basic need – hunger. I saw a sign for the cafeteria back near the elevator. Swedish meatballs, anyone?

Žižek: Ah! The famous Swedish meatballs – truly the final lure of the IKEA experience. The promise of meatballs at the end of the maze is like the reward for our consumerist pilgrimage. Let’s indulge. As a Slovenian, I have a soft spot for Balkan meatballs, but I will analyze the Swedish variety with equal fervor.

[ Scene Transition: The three philosophers navigate to the IKEA cafeteria, trays in hand. They each have a plate of meatballs with lingonberry sauce and a small drink. They sit around a minimalist dining table (possibly another display item with a tag still on it). The atmosphere is calmer here, with other shoppers quietly enjoying their food.* ]

Wittgenstein: (stabbing a meatball with his fork, inspecting it as if it were a philosophical puzzle) There’s something comforting about this portion of our journey. It's a form of life I recognize: sitting at a table, eating a meal. No tricky language-games here, just simple use of a fork and knife. And yet, I wonder: when we call these meatballs, what are we actually referring to? The shape? The ingredients? The concept? If a vegetarian calls them “plant-based protein spheres,” are we still playing the same language-game? (He pops the meatball in his mouth thoughtfully.)

Kripke: The name “Swedish meatball” is interesting. Is it a rigid designator for this particular recipe – meat, spices, gravy – or could any nation’s variation answer to that name? In a possible world where Sweden had no meatballs, would “Swedish meatballs” just refer to something else entirely? (He takes a bite.) Hmm. These taste necessarily of beef and pork... although I hear one can buy vegetarian Swedish meatballs too. So the name might refer to a category rather than a single essence. Perhaps “Swedish” here is more of a historical descriptor than a necessity.

Žižek: (already halfway through his plate, gesturing with his fork) You know, the truly psychoanalytic moment is right here: we are indulging in the reward at the end, the famous meatballs, and telling ourselves it was all worth it. This is the fantasy of completion – as if now, having eaten, we have achieved the purpose of our journey. It’s like finishing a great novel and feeling a temporary catharsis. But I would argue the real ideological operation is that the meatballs make you forget the toil of the furniture shopping. They want you to associate IKEA with this warm, satiated feeling. It’s the store saying, “We’ve fed you, now you owe us your loyalty.”

Wittgenstein: Or perhaps it’s less sinister: eating together is a basic human form of life, a simple pleasure. Not everything is a devious ideological plot, Slavoj. Sometimes a meatball is just a meatball, to paraphrase Freud very loosely. (He smiles, sipping his lingonberry juice.)

Žižek: (grins and points his fork at Wittgenstein as if to make a grand proclamation) Aha, but Freud would retort: sometimes a meatball is not just a meatball! Consider: these meatballs are the objet petit a of the IKEA journey – the unattainable object of desire that keeps you moving forward. They symbolize comfort, home, the end goal. The entire time we wandered among couches and cabinets, some part of us was thinking, “at least I’ll have those meatballs soon.” It's the unconscious motivation they don’t advertise but implicitly promise.

Kripke: If I recall, Slavoj, you once said the problem with enjoying something is that it becomes a duty – “the superego imperative to enjoy.” Are you suggesting we’ve been commanded to enjoy these meatballs by the IKEA superego? (He gestures to a poster on the wall cheerfully advertising the meatballs at $5.99 a plate.)

Žižek: Exactly! The poster might as well say, “You must enjoy our signature meatballs now!” And we oblige, happily. It’s a benign command, of course – far be it from me to complain about savory gravy. But it epitomizes the paradox of modern consumer life: we are free to choose, but everything is so well orchestrated that we choose exactly what’s expected. We reached the finish line and claimed our reward, like lab rats getting a cheese treat. Dry, delicious Swedish cheese... or in this case, meat.

Wittgenstein: (shrugging with a small grin) I won’t overthink it. The meaning of this meal is in the eating. In our current language-game, “to enjoy meatballs” just means to savor them and maybe remark on their taste. And I do enjoy them. They’re quite good – a bit heavy on the allspice, perhaps.

Kripke: I find myself agreeing. Necessarily or contingently, these meatballs are good. (He dabs his mouth with a napkin.) I also notice, we haven’t actually bought any furniture yet besides that KALLAX shelf in the cart. Are we going to check out with just one shelf after all this philosophical journey?

Žižek: Perhaps that is fitting. We traversed the entire IKEA cosmos, analyzed it to pieces, yet we emerged relatively unencumbered by actual commodities. It’s a bit subversive, isn’t it? To consume the experience but purchase minimal goods. We are the ones productively assembling ideas, while denying the store its full victory over our wallets. (He winks conspiratorially.) Though I might grab a plush toy on the way out – for research purposes, of course. They have a cute shark, what’s its name... BLÅHAJ?

Wittgenstein: (laughing softly in a rare moment of open humor) Aha, careful – naming the shark might make you emotionally attached. Next you’ll be analyzing the ideological significance of soft toys in adult lives.

Žižek: Oh, I already have thoughts on that, believe me. But I’ll spare you – we’ve had enough theory for one shopping trip.

Kripke: Agreed. Sometimes it’s best to conclude before we confuse ourselves any further. As the saying goes, “Whereof one cannot speak, thereof one must remain silent.” And perhaps, whereof one cannot assemble, thereof one must hire a handyman.

Wittgenstein: (raises his drink in a mock toast) To assembling both furniture and meaning – may we succeed in both, or at least find the process amusing.

Žižek: (raises an imaginary glass, since his cup is now empty) Cheers to that! And remember, the true absurdity is not that we talk about philosophy in IKEA, but that more people don’t. This place is ripe for it.

Setting: The three clink their plastic IKEA cups together in a playful toast. Around them, other customers continue their mundane meals, oblivious to the profound and absurd conversation that just took place. Our philosophers finish their meatballs, ready to face the checkout – and the task of assembling both the KALLAX shelf and the ideas they’ve gathered, once they’re back in the world outside the IKEA labyrinth.

[This dialogue was produced by Open AI’s GPT 4.0 Deep Search algorithm on July 16, 2025]

One of the most powerful arguments against the existence of a traditional God is the Problem of Evil. The Problem of Evil starts off noting that (a) if God existed, he wouldn’t allow gratuitous evil and suffering. But (b) there’s gratuitous evil and suffering in the world. It follows that God does not exist!

Throughout the centuries, many philosophers have tried to block this argument. But lots of thinkers haven’t been convinced by these attempts (a few details in the appendix).

Here’s a new solution to the Problem of Evil. We are already in hell. Since hell is a place where people get punished for their earlier wrongdoings, no evil or suffering in our world is gratuitous. It’s all deserved. Premise (b) is false. So maybe God exists after all.

Subscribe now

Some people won’t like this argument because they think that hell is a place of extreme and constant suffering and pain. Our world has pain and suffering but it’s not extreme and constant. So we are not in hell (or so the objection goes).

In response, we can ask, “how can we be so sure that this fire and brimstone picture of hell is correct?”. Maybe the suffering in hell is more subtle. Some theologians think, for example, that hell is being distant from God. What does this mean exactly? How can we be so sure that we are not distant from God now? This conception doesn’t seem to require the cartoon picture of hell with pitchforks, flames and coal.

In addition, we should consider that our lives are worse than we think: the best we can hope for in life is to build loving relationships with people and carry out meaningful projects—but all these things will come crashing down upon our bodily deaths—in a few short decades. How cruel.

Another objection to the idea that we are in hell is that it makes no sense for God to punish us without making us aware that we are in hell or that we are being punished for our earlier wrongdoings. To be punished, we must remember our bad actions. But why? Imagine a bank robber who bumps his head as he is trying to evade the cops. He gets amnesia losing all memory of his crime. It still makes sense to send him to jail—even if he doesn’t remember his bad actions. Just punishment doesn’t require memory of one’s wrong doings.

One observation about the idea that we are in hell is that it is cruel. It is cruel because it says that the suffering we witness, including of our loved ones, is deserved. I grant that this is a tough pill to swallow. But the opposite idea that the suffering we witness is not deserved is also cruel, given that God is allowing it.

In addition, the notion that all the suffering we witness is deserved is similar to what St. Augustine said: due to the sins of Adam and Eve, humanity’s suffering is our just punishment. The difference is that with the new idea, this suffering we are witnessing is happening in a different location: hell.

So maybe God exists and we are in hell. I find the thesis to be disturbing and intriguing. It stretches the imagination. And responding to it could help clarify our existing conceptions about God, life and evil.

Appendix (Two Classic Theodicies)

At the beginning, I mentioned that many people believe existing responses to the Problem of Evil (Theodicies) are inadequate. Let’s look at a couple of these.

Some say that because God loves us, he has granted us free will. But free will means that we can choose to do good or evil. This is the price of freedom. Hence, God’s existence is compatible with the existence of gratuitous evil and suffering after all (premise (a) is blocked).

The problem with this theodicy is that a lot of brutal suffering is not due to human actions. Natural disasters and many diseases are not the consequences of human choices. Why would God allow them? The Freedom theodicy doesn’t explain this.

There are also famous theodicies which try to block (b). Here’s one: The evil and suffering we witness exist because they serve some greater purpose. But what purpose does it serve? Is it to teach us a lesson? But if God is so powerful, then couldn’t he teach us these lessons in some gentler way without the pain and suffering we see in the world. Some thinkers say that the evil we see serves a higher purpose, but we have no idea what it is, because God works in mysterious ways. But this is not a response to the Problem of Evil as much as just an announcement that there isn’t a problem.

Imagine you wake up tomorrow to the announcement that a team of scientists and philosophers working in a top-secret facility created an extremely intelligent robot that is conscious---it can feel pleasure, pain and other emotions. In addition, thousands of them have already been built and are on standby in a warehouse waiting to be switched on.

World leaders are set to address their nations within the hour. How will they respond?

The idea of an artificial intelligence device that is conscious has long fascinated philosophers, scientists and the public at large. Since consciousness likely arises from the physical world, there’s no reason why we can’t, in principle, create those material conditions artificially. Discussions on the topic, however, are often set in the distant future where present day details are abstracted away. The reality is that AI technology is accelerating rapidly. We should then take seriously the possibility that AI robots may be conscious soon, even within ten or twenty years. It would be foolish to dismiss this possibility outright.

If we get sentient AI with the added capability of general intelligence, then the most pressing moral question for our society may be: what kind of rights should the machines be afforded? It may sound far-fetched, but I want to argue that this question will be so fundamental to how we think about the world that the political landscape will be centered around it. It will lead to a collapse of our current political parties and a subsequent reconfiguring around the question of AI rights.

There have only been a few times in America’s history where a single issue has led to the demise of political factions. For example, the Whig party of the 1800’s included anti-slavery moderate northerners and some pro-slavery southerners. Ultimately, the issue of slavery became so divisive that the party couldn’t hold it together. If I’m right, a similar thing will happen soon to our current parties.

Why Give AI Rights

Now why would the AI rights question be seen as fundamental? Here’s the simple reason. It’s not just a logical reason but a psychological one. After all, animals are also sentient, but we don’t grant them rights.

Eventually, and for better or worse, our interactions with AI will be intimate and barely distinguishable from our connections with other humans. We will develop strong emotional bonds to them and will be moved when they report pain, sadness, joy or anger. Many of them will be kind to us even sacrificing themselves. In addition, they will be able to speak up and ask for what they want, including demanding rights.

Why would AI robots demand rights? The answer is that this would be a side effect of their design. We will need many AI devices to be human-like in certain respects because we want them to make decisions on our behalf--reflecting our values. And this means imbuing them with a moral sense. Once this happens, it will naturally lead them to ask for rights.

The rational case for rights will be a powerful one since we are assuming these machines will have all the relevant properties which make them fully deserving of rights. As we have been saying, they will be conscious—with the ability to feel pain and pleasure. But they will also have other mental states like beliefs and desires as well as possess the ability to form intentions and plans. In terms of rights, there won’t be an important respect in which humans differ from robots. This will be clear to a lot of humans.

The kinds of rights we are talking about may include free speech, due process, equality under the law, a right to remedy and many others, perhaps even including a right to vote. Of course, the founding fathers did not anticipate this technology, so little in the constitution will be of any help. The courts will see it as a novel question of law. Interestingly, legal scholars have already been discussing this very issue for decades.

How Will It Affect the Political Landscape?

So, what will happen to our political landscape? Since we are considering the possibility that AI consciousness will happen within our lifetimes, it is reasonable to think that our political parties will be similar to what they are now in the United States, with progressives and social justice advocates on the left and MAGA-style nationalist conservatives on the right.

These traditional parties will crumble. Two new alliances will be created made up of a mix of conservatives and current liberals. A large chunk of people from both of our current parties will form a pro-rights coalition for the reasons I gave. The rest will be the anti-rights party. Let’s see how some existing threads from the right and the liberal left could naturally coalesce to form an anti-rights party.

The anti-rights party will itself be split into two. Some of them will be the sentience-skeptics. These people will deny that AI robots are conscious even if there’s strong evidence that they are so. Philosophers have long noted the following about consciousness: Although I can prove to myself that I have consciousness, proving it to someone else is tricky. No one else can be in my head to experience what I subjectively experience. This epistemological gulf will be exploited by anti-rights activist. They will protest that since we can’t know that AI is really conscious, we should not grant them rights.

But putting skeptics aside, here’s why some conservatives could still be on the anti-rights camp.

How Will It Affect Conservatives?

First, some of them are old fashioned conservatives which, true to their label, are weary of progress. They find change risky. These individuals would balk at extending rights to a totally new different class of individuals.

Second, we should consider that there’s often a tension is conservative thought between those with protectionist and cultural preservation instincts on the one hand, and more free market economically driven types on the other. This split exists currently within the MAGA movement. There’s the Cultural-Nationalist wing who want working-class protections and a return to traditional American culture or nativism. The key players on this side are currently Steve Bannon, Stephen Miller and Tom Cotton. This group strongly dislikes immigration even of high skilled people. Many on this MAGA wing may deny AI rights because it would threaten jobs for the human American worker. It would also be seen as a perverse affront to traditional values grounded in human-human relations.

In contrast, there’s the techno-nationalist MAGA wing currently represented by Elon Musk and Vivek Ramaswamy. The members of this group prize innovation, meritocracy and are pro-immigration for skilled workers. Unlike the Cultural-Nationalist MAGA wing, this camp may have no qualms overturning traditional human-centric values in the name of progress or efficiency. If this means giving AI rights, then so be it. However, some of them will wish to deny AI rights to justify their indiscriminate use. If AI have no rights, they can be used in any way we see fit, much in the way we often treat animals today.

Finally, many on the religious right may see AI rights as an affront to the God granted specialness of the human-race.

How Will It Affect Liberals?

Let’s now turn to why some on the left may be anti-rights and join forces with many conservatives. First, since these machines are made by corporations, ceding them rights will be considered a further empowering of the wealthy elite. The extension of rights to machines may be seen as the final blow to democracy in the spirit of Citizens United--which afforded non-human entities some of the rights traditionally only granted to human beings. This camp will also worry about AI machines getting a right to vote or participate in the democratic process. Corporations could create millions of these machines each participating in our democracy and shaping the future of our nation by advocating for their makers. Our democracy would be undermined if not totally destroyed.

Second, some on the left will say that moral attention to AI is a luxury cause. It will draw attention away from minorities and the poor. In many ways, AI robots will be the most privileged among us since they may be free of disease and could potentially be immortal. Many liberals may then be more reticent about granting them some rights for this reason.

Finally, many on the left who are union supporters will just see the pro-rights view as another attempt to weaken the American worker. Along this dimension, these liberals will unite with those on the Cultural-Nationalistic MAGA camp who also wish to make the American worker a priority.

An Argument for Limiting AI Rights

I want to end with an argument for limiting AI rights. To maintain an orderly society, we give the state power over us--the power to imprison or kill us. In exchange, we demand civil and procedural rights. AI machines, however, will be hard to kill. Their personalities and memories can be stored and reproduced at will, even if their physical vehicles are destroyed. Because of this, their claims to rights will be limited.

We need to start taking seriously the possibility that machines may be conscious soon. We should then start asking hard questions about how this would affect our world as it is now, and not just in some idealized future. I have highlighted that one likely outcome is that our political parties, as we know them, will crumble. New ones will be formed around the question of AI rights. Although our current political divisions seem deeply entrenched and intractable, they may soon dissolve.

Subscribe now

In midlife, Leo Tolstoy experienced a profound existential crisis, feeling that his wealth, fame, and literary success were meaningless. He experimented with different ways of living: renouncing his aristocratic lifestyle, simplifying his possessions, working in the fields, and adopting radical Christian pacifism. Through these experiments, he sought to live authentically in line with his true convictions, rather than continuing to live in the comfortable but hypocritical society he was part of. His work A Confession documents this process, showing how he tested different ways of living to align with his authentic self. Other thinkers like Thoreau, Nietzsche, Wittgenstein also got out of their armchairs and carried out life experiments.

Many people in the west think that we should try to live authentic lives. For example, blindly following the career path preset by your family or marrying whoever your parents decide for you are clear violations of western cultural norms. But not everything goes. If you can’t make money being the authentic you, then many will tell you to give it up. Still, we think there is great value in discovering your true self whether or not this “you” can be ultimately realized.

So how do you find your authentic self? Here are three methods.

(1) Armchair Introspection. You look inwards and reflect on your past experiences to see what things have worked for you and which haven’t. This method has problems. Not only are your past experiences limited by whatever life happened to throw at you, but the method is subject to the characteristic problems associated with introspection psychology. These include, for example, our propensity to make up inaccurate explanations for our attitudes and behavior. For example, some people that are homophobic actually have homosexual impulses, something they would deny even to themselves through the method of introspection.

(2) Quasi-Experiments. The second method is Tolstoy’s more pro-active experimental approach. Here, you literally experiment by trying out “ways of living” and seeing what sticks. Young people who try to “find themselves” do this. We’ve all seen extreme and sudden examples of this. One day, a kid may be dressed like a preppy New Englander and the next, they are full-on Goths. In middle school, I knew a kid that transformed overnight from a jock to a character from Miami Vice, with complete bright white suit, turquoise t-shirt, and a perm.

This second model is an improvement on the first because we added an experimental element. One of the greatest technological advances in human history is the development of experiments to answer questions about the world. To learn, you must get your hands dirty and probe the universe. In his time, Tolstoy had already witnessed the triumph of the experimental method. A century earlier, Peter the Great established the Russian Academy of Sciences to get his country caught up with the rest of Europe in the development of experimental techniques.

(3) Randomized Experiments. Tolstoy’s experimental method still has serious flaws. It doesn’t involve a randomized element. What was Tolstoy’s method for deciding which life paths to try out in the first place? It appears ad hoc and seems to boil down to making a “judgment call”. This is no good. Imagine a pharmaceutical company today testing the efficacy of a drug by making a judgment call about which individuals will be treated with the drug, and which will get the placebo. This would create an obvious bias. Since they want their drug to succeed, they might pick the healthier people to get the drug.

Similarly, suppose you are in college trying to find the major that reflects your true self. Maybe you think History is your thing. You try it out by picking a History of Egypt class. You love it. But wait a second. Maybe you chose that class because the teacher was charismatic and now you’ve chosen your major in this ridiculous way. You were better off choosing some history classes at random to see if the subject really is your passion (as opposed to the teacher being the object of your passion).

Some people like to travel as a path to self-discovery. But how do you choose where to go? You get recommendations from friends perhaps. If this is your method, there’s a worry that you won’t discover yourself, as much as end up with preferences that align with those of your friends. Choosing a travel spot at random gets around this problem. Similarly, suppose you are on a dating app, and you have narrowed down the contenders to a few who meet the minimum requirements. Choosing the finalists at random may allow you to cancel biases—like the bias of overcorrecting from the disaster of your last crush.

Because we are finite beings with limited resources, we cannot run a full-on randomized control trial to find our authentic selves (although this is a project that would make you Tik-Tok famous if you did it). A more realistic approach is to experiment with life but where you add a dash of randomness to the process. This takes a little courage, but you might get good results.

By profession, I consider myself an experimental philosophy within the analytic tradition. I’ve ran lots of randomized experiments usually with social and behavioral scientists. I’ve argued here that some of the rationale for using experiments in science also carry over to deeply human questions of authenticity. These questions tend to belong to the existential tradition in “continental” philosophy whose champions include Kierkegaard, Sartre, Camus, Tolstoy, Dostoyevsky and many others.

If you are the sort of person looking for authenticity and in addition, you experiment and are adventurous enough to add an element of randomness, then you should call yourself an “experimental (continental) philosopher”. I’d love to hear your story.

When my daughter was a baby, she crawled under a chair chasing after a ball. As she tried to extricate herself out of the situation she hit her head on a horizontal wooden rung. It was a minor incident except that, as I watched the events unfold, I observed something strange. I found myself feeling pain in my own head, as if I was the one who bumped my head.

And so on it goes with parenting, a journey which opens the human channels to exalted empathy. The child’s pain is the parent’s pain. This is the natural state of things. But sometimes we have to overcome our natural instincts.

To live, one must feel negative emotions. Sometimes our kids don’t want to do homework, or go to school. Sometimes they get sick or injured. They lose loved ones. Many worse things can and do happen. These events trigger negative emotions in our children, and with our guidance, we have to let those feelings run their course.

We will always suffer when our kids feel pain. But just how much should we suffer with them? Suppose your child cries because he doesn’t want to clean his room. Should you become so distraught about it, you can’t work or get yourself to make dinner for them? Obviously not.

This is an extreme example. But the same goes with more moderate cases. Your child needs to feel your genuine empathy when they are going through a difficult time. But there are many situations, especially when being away from them, when you would better serve them if you were in a positive state, emotionally unaffected by your kid’s pain. For lack of a better term, there are contexts in which it is best to become more cold-hearted to your child’s suffering—where being more cold-hearted just means lowering the volume of your empathic emotions towards them.

Let’s distinguish two ideas. The first one is that good parenting requires allowing your child to have short term suffering for long term gain. I’m talking about something else. I’m addressing this general question: ‘what is the appropriate reaction to your kid’s pain?’.

Aristotle told us that we should be having ‘feelings at the right times, about the right things, towards the right people, for the right end, and in the right way.” It is incredibly difficult to discern when the right times, the right things, the right ends and the right ways are supposed to be. This is the philosophical challenge of parenting.

I am drawing attention to one of the hardest things about this challenge. There are times when our love for our kids requires having to feel less. What makes this so difficult and almost paradoxical is not just that it goes against our parental instincts. But that even describing it in these terms seems taboo, though I suspect every parent knows this to be true, at least to some extent.

I’ve been a philosophy professor for twenty years. This fall, I will require students to use Artificial Intelligence to write take-home essays. But I will use a “curve” to evaluate students. We need curves for AI assisted writing.

Curves

When I was in college, it was all about the curve. In some classes, you got an A if and only if you scored in the top ten percent of the class. There were classes where a test score could be a forty out of a hundred. You didn’t panic. You eagerly awaited the announcement of “the curve” which could mean you actually got a good grade. I remember people complaining about other kids "messing up the curve” because they studied “too hard”.

It’s not hard to see the problems with this grading strategy. On this way of evaluating students, a good grade doesn’t mean that you mastered the material. It just means most students performed worse than you. Your grade reflects how everyone else did as much as how you did. This method incentivizes unnecessary competition rather than learning. Because of this, the last few decades has seen a move away from curves towards “standards based learning”. According to this method, the teacher sets a standard for what counts as mastery and if a student meets it, then they get a certain grade. In principle, most if not all students could get A’s (and sometimes they do!).

Artificial Intelligence

I think the way we evaluate essay writing has to change because the way that students write is changing. Students are using Artificial Intelligence to write. It’s a fool’s errand to try to stop them. But we must grade them on a curve. I don’t see other options.

A teacher who uses ‘standards based learning’ will usually have a specific rubric for what it takes to get a A, B, C and so on in their class. The problem is that the more specific and mechanical the rubric, the easier it is to just put those requirements into an AI chat bot, and with a few tweaks you can get an essay that will be graded with a B or even an A. And we are racing against time here. Pretty soon, AI will just produce A work. Teachers would need to change their rubric, and so on, creating a ridiculous arms race. This whole thing would be a useless exercise.

But suppose you use a curve. Here, you give students the instructions for the paper but tell them that, for example, the papers in the 90th percentile for clarity and persuasiveness get an A. AI won’t be able to produce an A paper with a simple input. This is because (a) it won’t know what the top students are turning in, and (b) all the other students will also be using AI—and will also be asking AI to produce the paper using the same input at a minimum.

I said in the title that everyone in my class will be required to use AI. Since I’m grading on a curve, the student who does the simple thing of putting the essay instructions into chatGPT, hits “return”, and then copies and pastes the answer will get an F for the essay. Today, this student may be getting a C or better. Let’s admit it. This is a disaster.

On the AI/Curve method, a student that gets an A will produce an excellent essay. They will have to spend a great deal of time crafting their paper, interacting back and forth with AI. They can use AI at every turn, from brain storming ideas, to organizing, to constructing sentences (“I need a metaphor in this sentence, give me some options”). Students will also need to do quite a bit of research on their own: Professors can impose very steep penalties for students who submit hallucinated or plagiarized materials. An A paper on this method will be well-deserved.

Of course, this means that the assignment of grades will be more subjective. The teacher will need to make a judgment call as to what counts as an A essay (it can’t just be some mechanical thing in the rubric). Grades are based on comparing students. Issues of bias will creep up and we need to be vigilant about it.

What Will Students Learn. What Will They Not Learn.

We won’t know for a few years how learning will be affected by AI. But we can make some educated guesses about writing. First, students will be less focused on sentence construction, spelling and grammatical issues. This task will just be kicked to AI, the way arithmetic is kicked to calculators.

There’s a worry here, of course, that writing grammatically and smoothly will be lost with future generations (it’s already in decline—I’ve even heard of college professors not even grading for grammar!). This may be so. But no doubt there will always be a small demand for and perhaps larger supply of “made from scratch” writing, just like we have “made from scratch” cooking. (Consuming a well-crafted, wholly produced, human essay may become a symbol of privilege. ‘There is nothing like sitting by the porch on a summer’s eve, enjoying a glass of chilled Beaujolais and reading a fully human made-from-scratch essay. I can really tell the difference, you know. It’s good for the soul’. )

Having outsourced the micro-mechanics of writing to AI, students will be focused on two distinct things which I hinted at earlier. (1) Ideas/argument structure development, and (2) Fact checking. These skills are AI-specific, since they require interacting heavily with AI. My sense is that these are the types of skills which will be needed in the workplace. So the AI/curve method I propose will prepare students well for writing and communicating in their jobs.

How This Whole Thing Fails

There are at least two ways this whole exercise fails. First, students can sell AI companies their “A” papers. These companies can train their models on these excellent papers and promise other students more of the same for an extra fee. Rich students will get A’s. If this happens and we have no way to stop it, then we need to go back to the drawing board.

Second, all writers know that “writing is thinking” and that ideas come to you and get clarified as you write. Exactly how this magic works is not totally understood. But if this is short circuited with AI, then we need to alter our system. For example, we can dedicate in-class time to “write-for-thinking” sessions where AI is prohibited. Students can then take those ideas home and refine them with the help of AI.

Conclusion

It’s inevitable that students will use AI for writing take-home essays. We can allow this, encourage it or even require it. If so, we should seriously consider going back to the curve method for grading.

.