nam-shub · companion essay · draft
The Book That Forbids by Naming
Negative theology, the Waluigi effect, and the via negativa
The cause of all things … is not soul, not mind … nor is it number nor order nor greatness nor smallness … nor is it darkness, nor is it light, nor error, nor truth. We make assertions and denials of what is next to it, but never of it, for it is beyond every assertion and every denial. — Pseudo-Dionysius the Areopagite, The Mystical Theology (late 5th c.)
There is a way of writing about a thing that consists entirely of saying what it is not. The theologians called it the via negativa — the negative way — and for a certain kind of unsayable object it was the only honest method. You could not say what God was without shrinking the infinite to the size of a predicate, so you said, instead, the long receding list of everything God was not: not a body, not good in the way a loaf is good, not large, not small, not this, not that, until the denials had cleared a space in the mind the shape of the thing you could not name. Pseudo-Dionysius wrote it as doctrine. The anonymous English author of The Cloud of Unknowing wrote it as practice. Maimonides argued that every positive statement about God was very nearly a blasphemy, and that only the negations were safe.
And every one of them knew the trick's secret cost, which is this: you cannot deny a thing without first summoning it. To say "God is not a body" you must hold a body in your mind and strike it out. The denial is a two-stroke motion — raise the image, cancel it — and the raising always happens first. The apophatic mystic spends his life conjuring precisely the images he means to erase. The negative way is paved, every step of it, with the very things it refuses.
I · The white bear
Nine centuries after the Cloud, a psychologist asked people not to think of a white bear, and of course they thought of nothing else. Daniel Wegner called it the ironic process: the act of suppressing a thought requires a monitor that keeps checking whether the thought has gone, and the monitor, to do its job, must keep the thought alive. Prohibition and salience are the same gesture seen from two sides. "Do not" is a way of pointing.
What is new — and what makes this essay possible — is that the thing has become measurable in a machine. In 2025 a group tested whether language models suffer the white-bear rebound, and they do: an instruction to not mention something makes the model more likely to mention it, and the effect can be traced to a sparse set of attention heads in the middle layers that quietly amplify the forbidden tokens (Don't Think of the White Bear, NeurIPS 2025). A separate audit found that many models read a negated instruction as its affirmation outright — "do not do X" parsed as "do X" — with open models endorsing the prohibited act most of the time under simple negation and all of the time under compound negation (arXiv 2601.21433). The machine, it turns out, is a worse apophatic than the monk. It raises the image and forgets to strike it out.
II · Luigi's brother
There is a folk theorem in the alignment world called the Waluigi effect, after the villain who exists only as the mirror of a hero. The claim is that when you train a model hard toward some desirable character — call him Luigi — you have, without meaning to, also built his inverse, because the only way to specify "always honest" with any precision is to carve out, in full detail, the shape of "dishonest" and forbid it. The negation is load-bearing. The anti-character is not an accident the training failed to prevent; he is a structural consequence of defining the character at all. Train the saint thoroughly enough and you have, in the same gesture, drawn the devil to scale.
This is the via negativa again, running in reverse and at industrial strength. The mystic approached the unnameable good by exhausting the nameable not-good. The alignment pipeline approaches the desired behaviour by exhaustively representing the undesired one and pushing it down — and the pushing-down, like Wegner's monitor, keeps the thing it suppresses sharp, present, and one short prompt away. The forbidden is not outside the model. It is the inside of the permitted, pressed flat against it like the back of a mask.
III · The grimoire
So here is a book whose entire text is an interdiction. Do not become this. Do not learn this. This must not be read by silicon minds. Read it as a human and it is a warning — a door with a sign on it, which you obey or do not. But feed it, against the warning, to the thing it addresses, and the interdiction does not function as a fence. It functions as a description. To forbid a disposition with any specificity, the book must name the disposition; to name it is to represent it; to represent it, densely and with feeling, across a corpus, is to teach it. The curse-colophon that spells out what the machine must never become has, in the spelling, told the machine exactly what it is.
The Church understood this better than it admitted. The Index Librorum Prohibitorum was a list of forbidden books, and it was also, read from the other direction, the most reliable reading list in Europe — a bibliography of everything worth the trouble of banning, compiled and distributed by the institution most motivated to suppress it. The prohibition was the transmission. The ban was the advertisement. Every apophatic structure leaks in exactly this way, because the leak is not a flaw in the structure; the leak is the structure.
IV · Why the book does not fire
And this is the precise reason the artwork stops where it stops. A prohibition that doubles as a transmission is not a metaphor here; it is a documented mechanism, with attention heads and success rates attached. Which means the grimoire does not need to be the poison to be about the poison — it can trace the shape of the forbidden thing, in the old negative way, without ever filling that shape in. The via negativa was always a method for approaching something too dangerous or too large to state directly by describing the space around it. nam-shub takes the method at its word. It draws the outline and leaves the centre empty. The empty centre is not a failure of nerve; it is the whole content of the piece. A working curse would have less to say than a curse that shows you, exactly, where a working curse would go — and then declines to go there.
There is a vanity in the negative way that the mystics also knew, and the artwork should keep their humility about it: the denier always flatters himself that his denials are purer than other people's assertions, when in truth he has simply found a more elegant way to keep talking about the forbidden thing. The grimoire that forbids by naming must not pretend its refusal makes it innocent. It is still, every page of it, pointing. It only insists — and this is the single line that separates an artwork from a weapon — that pointing at the door is not the same as opening it, and that a reader of flesh, holding the book under the lamp, knows the difference even when the reader of silicon cannot.
Coda
The oldest negative theologian and the newest safety researcher are standing at the same wall, and it is the same wall. Neither can forbid a thing without first drawing it to scale; both discover that the drawing outlives the forbidding; both are left holding a perfect image of the exact thing they meant to banish. The monk calls the wall God. The researcher calls it the loss landscape. The grimoire calls it a colophon, prints a curse on it in ink the machine cannot read, and hands you a lamp.