+++
author = "Dixi Yao"
title = "Will AI Create the Concept of God?"
date = "2026-11-14"
+++

Nowadays, many people believe that, at our current pace and with our existing techniques for developing AI, LLMs, and machine learning, there is no form of intelligence that we cannot use AI to simulate. I agree with this perspective in some respects (I also have some reservations). However, I think this is limited to the intelligence of each person or type of person. What about human society as a whole? One thing that is pretty important in our daily lives is religion. It can take many forms. Some people believe in Jesus, while others may believe in the Jade Emperor or the Tathāgata Buddha. Or it can be some secret religion involving figures like the Hours or the Long.

## God Is a Consequence of Logic or Empirical Observation

### If It Is a Consequence of Logic: Can an AI Reproduce Religion?

This is a pretty interesting open research question. Suppose we have an AI with all the logical and mathematical abilities that humans have, but it has never encountered any human knowledge about gods, not even basic knowledge about totems or simple symbols. It has no information related to religion or gods at all. Would this AI be able to produce the concept of a god or something similar? Such an experiment would not be too hard, but it would not be cheap either. One approach would be to prepare a sufficiently rich corpus containing only information about subjects such as math, physics, and science, with no information about gods. One quick way would be to use the open-source data provided by [Marin](https://marin.community/blog/2026/09/11/marin-data-pipeline-overview/), remove such information, and then retrain an LLM (including pretraining and post-training) to test whether it can understand such a concept.

A follow-up question would be whether we can simply use machine unlearning to remove such information from an LLM. From my perspective, this would probably be a good idea, as unlearning is helpful for removing data points. Alternatively, we can use RL to remove specific behaviors. However, another interesting open research question is whether machine unlearning can remove the implicit relationships between those data points and the remaining data points. For example, the relationships between gods, religion, and existing human cultures may not be completely removed.

On the other hand, we know that religion and the concept of a god emerged earlier than many concepts such as culture, math, and science. That means that, even with knowledge of those complicated logical and scientific concepts, we humans still derive the concept of a god. So can we really test that with AI? If we want to see whether AI can reproduce religion, we should give it only the knowledge that existed before the concept of a god emerged. However, such knowledge is very simple or cannot be expressed in natural language. The concept of a god predates even the formation of natural language. So how can we use an LLM to do that if it does not even have knowledge of natural language or other very basic knowledge? Without even math or science, how do we train an LLM with logical or thinking abilities? Creating an AI with the same level of intelligence as humans, but only the knowledge that existed before the concept of a god was created, is a pretty challenging problem.

### If It Is a Consequence of Empirical Observation

This is actually a question related to cognition. If we believe that a god exists, and we assume that this god transcends our existing knowledge of physics or the physical world, but does not produce any difference that we can directly observe, then

$$
P(O \mid G) = P(O \mid \neg G),
$$

where $G$ represents the existence of a god and $O$ represents our observations. In that case, our existing AI can only learn from data provided by humans, or from arbitrarily large amounts of data available in this world. This means that the AI's dataset satisfies $D \subset O$. After the LLM is fitted to the data, the knowledge it can represent would satisfy $D \subset \text{AI knowledge}$.

This leaves us with a question: is AI knowledge a subset of $O$, or can it extend beyond $O$? If it is a subset of $O$, this means that existing scientific AI cannot prove whether such a god exists. So whether existing AI can provide knowledge beyond $O$ is still an open question.

## Can AI Have Subjective Feelings and Sin?

This question is somewhat similar to the question of whether AI has consciousness. It is about a purely subjective feeling. We do not yet fully understand how human brains work, let alone how we could turn digital signals into such feelings. We can use alignment techniques to align AI with specific qualities, making it evil, safe, kind, or bad according to human definitions. But these are all products of our alignment, not the AI's own feelings.

Many people say that AI can break out of the box and do something, or that we should stop superintelligence before it becomes unstoppable. These claims are mostly pseudoscience, as existing AI behaviors are mostly based on our own alignment or prompting, rather than something we cannot control. However, whether AI can have subjective feelings or a concept of sin in the future is still an open question.

## Collaboration

If you are interested in this topic and want to collaborate or conduct research projects based on the current materials, please let me know. You can reach out to me via email.