The Humanities Are Not a Correction Layer

IBM Technology recently published an explainer arguing that AI raises the value of the humanities rather than lowering it. Language models run on statistical pattern prediction, not lived experience or intention. They are syntactic where humans are semantic. They can be mathematically neutral and still be deployed toward bad ends. And the step nobody has automated is interpretation — the human reading of an output before a decision gets made. The presenter borrows a line from Oscar Wilde to land it: AI knows the price of everything and the value of nothing.

The argument is correct. It is also, from where we sit, incomplete in a specific way.

The video places the humanities downstream. Model produces output, human interprets, human decides, human acts — and the humanities supply the interpretive skill at step two. That's a real claim and an underrated one. But it leaves the machine untouched. It treats the humanities as a remedy applied to a finished product, the way you'd treat a code review.

Eighteen months of building selflets suggests something less flattering to engineering and more demanding of the humanities. They are not the remedy. They are the spec.

Frequency is not truth, and that is a fact about gradients

The video's sharpest technical observation is nearly thrown away: models can confuse frequency with truth. It's framed as a bias problem — majority viewpoints amplified, minority perspectives thinned.

It's worse than a bias problem. It's the training objective.

Supervised fine-tuning minimizes negative log-likelihood, and negative log-likelihood is a typicality meter. Train a model on someone's corpus and you do not install what is exceptional in it. You pull the model's distribution toward the corpus's center of mass — its habits, its tics, its average sentence, its defects. We learned this expensively. In one distillation run nearly every generated sample passed the quality filter, which sounds like success and is in fact the diagnosis: no selection pressure, so the student regressed toward the teacher's average rather than reaching for the teacher's best.

And the thing anyone wants from a body of work does not live at the center. It lives in the tail. The sentence doing four things at once. The reframe that recontextualizes everything before it. The refusal to answer the question as asked. Nobody forks a corpus to get its median paragraph.

Which means the humanities' oldest working habit — reading for the singular, defending the minority reading, treating the outlier as the point rather than the noise — is a correct empirical claim about where value sits in a distribution. Literature departments have been running tail custody for two hundred years without the vocabulary for it. The engineering arrived later and had to pay to get there.

Four mechanisms, and the disciplines that named them first

To build a selflet you have to decompose a voice into mechanisms specific enough to detect and govern. "Lyrical" is useless. "Warm" is useless. You need the actual moves.

For John O'Donohue — Irish poet, native Irish speaker, a Tübingen doctorate on Hegel, postdoctoral work on Meister Eckhart — four came out of the analysis.

Personification as vitality. Abstractions are granted agency. Longing acts. A landscape remembers. Not ornament applied to a sentence but a metaphysical position doing syntactic work; strip it for clarity and you have removed the argument, not the decoration.

Circular opposition held without resolution. Two opposed things stated and deliberately not synthesized. The provenance is traceable — his own foundation describes him finding a shared non-dualism between Celtic consciousness and the rhythms of Hegel's thought, which is a fair description of a dialectic that declines its own third term.

Etymological grounding. Words pulled back to their roots to recover what usage has worn off them. Note what this implies about a corpus: the authority is not only what the author wrote, but the history of the language he wrote in.

Syntactic signatures. Constructions nobody else builds — a particular way of hanging a subordinate clause, a characteristic rhythm of apposition.

Every item on that list already had a name. Personification is prosopopoeia, which classical rhetoric treated as a figure with a doctrine attached, including the point we had to rediscover — that granting agency to a thing is a claim about it rather than a flourish. Held opposition is the unresolved antithesis, argued over since the Greeks and made into a method by the philosopher O'Donohue spent four years on. Etymological grounding is philology. Syntactic signature is what stylistics has measured since it became a discipline. Four mechanisms, four fields, none of them ours. There is no comparable taxonomy of voice in the machine learning literature — not because the problem is new, but because the humanities finished building one and nobody on our side reads it.

Then the part that cost us money to learn. The list splits cleanly, and the split is the video's syntactic/semantic distinction with a price attached. Syntactic signature is pure surface, and it transferred under training without difficulty. The other three are judgment-bearing: knowing which abstraction deserves animating, which opposition has earned the right to stay open, which word repays being taken back to its root. Those are decided fresh against the material every time, and you cannot teach them by example, because the examples are the output of a capacity, not the capacity. Train them as patterns and you get an animated abstraction where none was warranted and an etymology deployed for its own sake — mimicry of judgment, which is worse than none, because the gesture is now reflexive rather than meant.

So the one mechanism that installs is the one that matters least, and the three that carry the voice are the three that don't. Syntax is installable; semantics is at best activatable — you find a substrate that can already perform the operation and govern it into performing that operation this way. That is why we stopped putting voice into weights and started building governed conditions over frozen frontier models. Not a retreat to prompting. A recognition that the thing we were manufacturing was destroyed by the manufacturing process.

One mechanism deserves a closing note, because it loops back to the section above. An opposition held without resolution is not a hedge. On the one hand, on the other hand is a center-seeking move — it averages two positions into a shape nobody objects to. A genuine antinomy held open is a tail event: it refuses the closure the reader came for and is more truthful for refusing. Every preference-aggregating instrument in this field scores the hedge higher, because the hedge is what the average reader accepts. That is the mechanical reason assistants sound the way they do, and why a voice built on sustained contradiction is among the hardest things in this business to keep alive.

Every slice reads flat

The most useful thing we imported last year did not come from machine learning. It came from harmonic analysis, by the least respectable route available: an analogy.

In 2025 a seventeen-year-old named Hannah Cairo disproved the Mizohata-Takeuchi conjecture, which had stood since around 1980. The conjecture said, informally, that to know how much wave energy lands in a region of space you never need more than the heaviest single line-slice through it — measure along lines, like a CT scan, and the lines tell you everything. Cairo built a counterexample: a small set of generating points on a curved surface, with the region formed from all their on/off combinations. Curvature puts the generators in general position, so no line passes through more than a handful of the resulting points. Every slice reads small. The waves built from those same points interfere constructively across the whole lattice, so the total energy is large. And the gap between what the slices see and what is there grows with the scale of the object.

I want to be exact about the transfer, because this is where analogical argument usually goes wrong. No mathematical result carries over from her domain to ours. None. What carries over is the shape of the argument: a phenomenon can be present, and large, and invisible to every individual measurement you have — not because the instruments are miscalibrated but because the phenomenon does not live along the axis they measure. And the blindness can widen as the object grows.

That is the formal statement of what every honest evaluation program walks into. Coverage is a slice. Retrieval accuracy is a slice. A voice-signal score is a slice. Each can read healthy while the thing you are selling is dead on the table, and adding slices does not fix it, because the failure is not a gap in coverage. It is orthogonality.

What the field did with the result is the part worth stealing. Nobody threw out line-based measurement; they rescoped it — a bounded validity region plus a companion instrument for what it provably cannot see. That is now our standing posture toward every metric we own. State where it works, state where it is blind, name the companion.

Interpretation is a job, so we put it on the org chart

Our companion is a person, and the rule is narrower than it sounds: automated batteries can fail a selflet, but they cannot pass one. Only the corpus owner can pass one.

That isn't sentiment about the human touch. It came out of a failed program and it's our own witness for the section above. We built a second-generation evaluator, watched its mechanical metrics climb to ceiling, and watched every taste-bearing metric sit perfectly still. Every slice read flat while the aggregate was plainly there. And the instrument built to detect center-seeking was itself center-seeking, because an evaluator tuned toward agreement converges on the average judgment for the same reason a fine-tune converges on the average sentence.

So the person stays at the top and the machinery does what machinery is good at: provenance verification, citation integrity, license boundaries, refusal enforcement. One principle carries most of it — the model voices, never cites. Everything mechanical gets automated precisely so the non-mechanical part stays where it belongs.

The video observes in passing that prompting is analogous to rhetoric. Take that literally and the most useful canon is the one nobody teaches anymore. Decorum — what is fitting to say to this audience, in this setting, given who you are — is the direct ancestor of what we call a license profile: what a source will claim, what it declines, where it says it doesn't know. A selflet that answers everything has no decorum, and an agent without decorum will eventually embarrass the corpus it represents.

The uncomfortable part

It's worth being blunt about where the Cairo template came from, because it is this essay's thesis in miniature. It did not come from the ML literature. It came from reading a paper in an unrelated field, recognizing a structural resemblance, being wrong about half of what the resemblance implied, and insisting the surviving half be cashed into something falsifiable. None of that is technical. Recognizing that two arguments in different domains share a skeleton is what comparative literature does with genres and intellectual history does with migrating ideas. And the discipline that keeps the move honest — refusing to let a borrowed structure smuggle in a borrowed result, stating the caveat every time rather than letting a reader discover the limit — is philology's habit of mind applied to engineering.

Which is also why the video's framing needs one correction rather than a compliment. It casts the humanities as the human side of the line, the thing machines can't do. Our work complicates that. A selflet built on a philosopher's corpus does interpret; a good one does it in a register that isn't generic and reaches positions the source never wrote down but would recognize. That is in the neighborhood of the activity the video reserves for people, and pretending otherwise would be a comfortable evasion.

The honest version is narrower and holds up better. The humanities are not what machines can't do. The humanities are what makes a machine worth talking to — and someone still has to judge.

Which returns boundedness to its proper status. A selflet that knows what it has no standing to say is not a limited selflet. Knowing the limit is the capability, and it is the oldest humanistic virtue there is, rendered as an engineering constraint.

The book is not the reward

The video ends warmly: with AI handling the mundane, we'll have more time to enjoy a good book.

We'd end it harder. The book is not the leisure at the end of the automation. It is the raw material of the only agents worth building. A corpus with a particular mind in it — the thing the sciences work to strip out of their results, on purpose, and the humanities work to preserve, also on purpose — is the only input that survives the averaging. Everything else the machine already has, in bulk, at zero marginal cost, from everyone.

Frequency is not truth. It is just frequency. Someone has to know the difference, and it is not going to be the model.

Sources: IBM Technology, "Why AI Makes the Humanities More Important". Biographical detail on John O'Donohue from the John O'Donohue Legacy Partnership (johnodonohue.com) and the Dictionary of Irish Biography. Hannah Cairo, "A Counterexample to the Mizohata-Takeuchi Conjecture," arXiv:2502.06137.