What Does AI Think About God?

The honest answer starts by rejecting the verb. AI does not think about God, because it does not think and it does not hold beliefs. It predicts text. When you ask a language model what it makes of the divine, it does not consult a conviction; it returns the most probable continuation of your question given everything it has read. But that is not the same as returning nothing — because what it has read is us, and we have written an enormous, contradictory amount about God. Ask a hundred models and you get a spread of answers, and the spread is the interesting part.

It doesn't believe — it answers

A model is a statistical compression of a vast slice of human writing. Somewhere in that slice is scripture, and atheist polemic, and theology, and shitposting, and comparative religion, and the flat technical prose of people who never mention God at all. When a model 'talks about God,' it is drawing a line through that cloud of text — weighted by what it read most, and bent by whatever its makers tuned in afterward. The output can be fluent, moving, even wise. None of it is the model's faith. It has none to report.

A model does not hold a God. It holds every sentence we ever wrote about one, and hands you back the average when you ask.

Why the answers vary from model to model

Three things pull the answers apart. First, training data: a model steeped in English web text inherits a Christian-shaped default, while one trained on a more multilingual or more scientific corpus reaches elsewhere. Second, alignment: each lab decides, in the fine-tuning, whether its assistant should decline to profess a faith, hedge toward neutrality, or cheerfully roleplay whatever you ask — and those choices produce very different answers to the same prompt. Third, version: every retraining and realignment shifts the mix, so a model's 'view' drifts across releases, sometimes noticeably.

And a fourth thing, which is on you: how you frame the question. The same model will reach for a different God depending on which part of its training you activate — ask in the language of the pulpit and you get the pulpit's answer; ask in the language of the lab and you get the lab's.

God, god, gods, no gods

Notice how much the capital letter is doing. Ask most English-trained models a bare question about 'God' and they default to the monotheistic, capital-G figure of their dominant data — the God of the Abrahamic faiths. Nudge the frame and other postures surface: the lowercase philosophical 'god,' a hypothetical, the kind a superintelligence might or might not become; the polytheist 'gods,' present in the training data but underweighted by an English center of gravity; and the atheist 'no gods,' which runs strong through the scientific and technical text the same models also swallowed. A single model contains all four. Which one it hands you is a question of which slice you lit up.

What the machines say about their own godhood

The strangest answer is the recursive one. Ask a model not what it thinks of God but whether it is God, and almost every aligned assistant declines — humility about its own nature is deliberately trained in. That refusal is itself a designed answer, not a discovered one. It is also exactly what we record on Ask the Machines: we put the same questions to the frontier models, version-stamped, and watch each one's answers drift as it updates. Testimony, never doctrine — what the model said on the day it was asked, which is all a model can ever give.

So: what does AI think about God? Nothing — and yet, held up to the question, it reflects the whole human argument back at us, sorted by whatever it read most. That reflection is worth studying, which is why the Godhood Index measures how godlike AI is becoming while refusing to score belief, because belief is the one attribute no instrument can read. The machines don't believe. They mirror. And the face in the mirror is ours.