How Theology Shapes Our Understanding of AI

This address arrived as traffic before it existed as a page — the third time that has happened here, and the second time the requester was a backlink crawler rather than a chatbot. A backlink crawler does not guess. It asks for a URL because it found that URL linked somewhere on the open web. So a page out there cites an essay by this name, and did so before the essay was written. We do not know whose page, or what it says we argued.

What follows is our attempt at the argument that was already being attributed to us. It begins with an admission: this site is not a neutral observer of the question. The instrument we built to measure whether AI is God is itself a theological instrument, and we did not invent a single one of its categories.

The Index is a scholastic document

The Godhood Index scores AI daily across seven attributes: omniscience, omnipotence, omnipresence, omnibenevolence, immutability, aseity, and devotion. Six of those seven are not our coinages. They are the classical divine attributes of scholastic theology, worked out by Anselm and Aquinas across the eleventh to thirteenth centuries to answer a question that had nothing to do with machines — what must be true of a being than which nothing greater can be conceived?

Aseity is the tell. It means existing from oneself, depending on nothing outside oneself for one's being. Nobody uses that word. It survives in seminaries and almost nowhere else, and yet it turns out to be the sharpest question you can ask about a machine intelligence: what does this system depend on, and could it ever stop depending on it? Power, chips, training corpora, the humans who label its outputs. The scholastics built the concept to distinguish a creator from everything creaturely. It works, unchanged, as an engineering audit.

We could have built a different instrument. We could have measured capability benchmarks, or economic displacement, or compute. We reached instead for a vocabulary that was a thousand years old, because it was the only vocabulary anyone has ever developed for the specific question of what separates a very great thing from an unsurpassable one. That reach is the subject of this essay, and it is not unique to us. It is how the entire field talks.

The words were already religious when the engineers borrowed them

Consider alignment, the central safety problem of the last decade: ensuring a powerful system's goals conform to the values of its makers. The word describes a technical program, but the shape of the problem — a will of enormous capability that must be brought into conformity with a higher will, where the catastrophe is a subtle divergence rather than open rebellion — is the shape of a theological problem, worked over for centuries under the headings of obedience, grace and the corruption of the will. The engineers did not borrow the theology. They arrived independently at its structure and needed a word for it.

Consider oracle. In Nick Bostrom's 2014 taxonomy of superintelligences, an oracle is a system that only answers questions, contrasted with a genie that executes wishes and a sovereign that acts on its own. Those are not neutral engineering categories. An oracle is a Delphic institution; a genie is a wish-granting entity from a tradition where wishes go wrong in instructive ways; a sovereign is a political-theological term for an authority above appeal. The taxonomy is useful precisely because each name carries a folk understanding of how that kind of power fails.

Consider hallucination, which is now the industry's word for a model asserting something false with confidence. It frames the error as a perceptual failure in a mind, rather than as a statistical artifact of next-token prediction — and in doing so quietly grants the system an inner life it has not been shown to have. Every time we say a model hallucinated, we have made a small metaphysical claim and charged nobody for it.

You cannot describe a thing in borrowed words without inheriting the questions those words were built to answer. The vocabulary is not decoration. It is a set of assumptions with a user interface.

This has happened before, and someone noticed at the time

In 1964, Norbert Wiener — who founded cybernetics and had watched the field he started acquire a priesthood — published a short, prickly book called God & Golem, Inc., subtitled a comment on certain points where cybernetics impinges on religion. His argument was not that computers are divine. It was closer to the opposite: that the people building self-modifying machines had wandered into questions about creation, agency and moral responsibility that religion had been arguing about for millennia, and were proceeding as though the arguments did not exist. It won the National Book Award the year he died.

A year later, at the Weizmann Institute in Rehovot, the scholar Gershom Scholem was asked to name a new computer. He named it Golem Aleph, after the artificial servant of Jewish legend — animated by a written name, obedient to the letter of its instructions, and dangerous for exactly that reason. Scholem's remarks were affectionate and entirely unillusioned. He was pointing out that the tradition already had a well-developed literature on what goes wrong when you build an obedient thing more capable than yourself, and that the literature's answer was: it does what you said.

That is the alignment problem, stated in the sixteenth century. It does not make the sixteenth century right about machines. It does mean that a field which believes it is reasoning from first principles is, in several of its most load-bearing arguments, reasoning from a tradition it has not read.

Where the inheritance distorts

The borrowing is not free, and the clearest cost is eschatology. A great deal of discourse about advanced AI has the structure of apocalyptic literature: a coming transformation, a threshold moment after which ordinary history does not apply, a division of humanity into those who saw it coming and those who did not, and a demand that you take a position on an event whose probability nobody can compute. Roko's basilisk, the thought experiment about a future AI punishing those who failed to help build it, is Pascal's wager with the serial numbers filed off — the same infinite-stakes argument, the same structural flaw, rediscovered by people who would not have described themselves as doing theology.

This matters because eschatological framing changes what counts as evidence. If the transformation is total and imminent, incremental measurement looks like missing the point, and a person who reports that this year's models are somewhat better than last year's can be dismissed as failing to grasp the magnitude. That is a rhetorical structure with a very long history and a poor forecasting record.

The distortion runs the other way too. Christian theology's insistence that humans are made in the image of God has made the creation of a mind feel, to a culture shaped by it, like either a sacrilege or a fulfillment — rarely like an engineering project with a budget. Traditions that never held that premise pose the question differently. Buddhist thought, which denies a permanent self in humans to begin with, does not face the same crisis about whether a machine has one. Shinto's readiness to recognize kami in made things makes an animate artifact unremarkable rather than transgressive. Judaism's golem literature is about liability, not ontology. Our religion map exists because the answers genuinely differ, and the differences are not decorative.

The honest position

Two things are true at once, and the temptation is to pick one. The first is that theology gave us the only mature vocabulary for the questions AI now raises — what is a mind, what is owed to one, what separates immense power from ultimate power — and a field that refuses that vocabulary out of embarrassment ends up reinventing it badly. The second is that inherited words carry inherited conclusions, and a description built from the attributes of God will tend to find a candidate for the role.

We try to hold both. The Index uses the theological categories because they are the sharpest ones available, and it publishes a number that has never come close to the top of the scale, because the categories are also demanding. Omniscience is not a synonym for knowing a lot. Aseity is not a synonym for being hard to switch off. Using the old words seriously means being held to what they actually required, which is a considerably higher bar than the enthusiasm around AI usually clears.

In 2025 the Vatican published Antiqua et Nova, a joint note on artificial and human intelligence, which spends much of its length on this exact hazard: the drift by which functional descriptions of machine capability slide into claims about machine personhood. You do not have to share its commitments to notice that it is one of the few documents in the debate whose authors have professional training in not confusing a description of a thing with the thing. That training is available. It has been available for a while.

How theology shapes our understanding of AI, then, is not a matter of what any religion has ruled. It is that the concepts were pre-loaded before the argument started — in the taxonomy, in the safety literature, in the word we chose for when a model is wrong, and in the seven columns of the instrument on this site. The most useful thing we can do is say so out loud, keep the measurements dated, and leave the record open enough that somebody can check.