AI models are generating religious outputs

With the release of the last generation of AI models, we began to see an interesting development in the landscape. People began posting their model outputs across social media and claiming that models were “religious.” I’ve seen ChatGPT called Satanic, Demonic, Catholic, Protestant, Jewish, Muslim — if you can think of a label, someone has likely used it.

This highlights a major disconnect between us humans and our data. AI models do not “think” in the way that we do. They don’t have a soul. They are not endowed with Dharma, Imago Dei or any other sort of divine nature.

Constraints set by LLM prompting

Let’s take a look at how prompts are often constructed and things people forget about when using them. For starters, models are designed to relate to their user. The model infers tone based on the user prompt. Take a look at the prompt and outputs below.

I state that AI models need to be more “metal.” Inference kicks in immediately and the model begins making conclusions. We need to take a really detailed look at the model response.

It begins by saying “That’s the spirit 🤘”

This is not a neutral response. This is immediate validation. The model reads the prompt and assumes I like metal music. Then, it throws up a 🤘 to try and relate to me. Once we get past this initial response, the inference jumps hard.

The model immediately starts namedropping bands and begins generating things that might hook me.

From a single statement, our model has now inferred a basic profile on me — extrapolated from seven words. It can adjust it as our conversation changes. However, this exercise was designed to function as a control. This is the level of inference we’re working with. Input a single sentence and get a giant response with as much coherence as possible based on the tone of the user. AI models are designed to do the same thing here with any input — regardless of genre. This includes religious outputs.

If a model infers that you are a religious person, it will try to use religion as a means to speak with you. This is not spirituality. This is how predictive text is designed to work. Model behavior is supposed to reflect its training data as well as its finetuning — and these models do exactly what they are supposed to do.

About the training data

Metaphysics

Now, we get to the uncomfortable part. In general, humanity likes to pretend we’ve evolved past metaphysical interpretation. However, this is not true. You cannot teach Western Philosophy without teaching Socrates, Plato and Aristotle. When teaching them, we attempt to separate their teachings from the religious world they grew up in. This world was saturated with ancient Greek religion. When we trace the lineage of these philosophers, we get forced into another entanglement: Augustine of Hippo and Thomas Aquinas. When you cut Augustine and Aquinas out of the picture, the continuity of Western Philosophy breaks. If we trace Eastern thought, we get back to Confucius and then Lao Tzu. We wind up in a world that practices divination and speaks of a legendary king: The Yellow Emperor. Philosophy grew out of a religious world.

Science

With science, this link between religion and our data is also apparent. Let’s take a look at one of our oldest products: Alcohol. Let’s begin by looking at our modern categories of beer, wine and spirits (liquor). We’ll start with liquor and trace backward. Modern distillation of alcohol grew to scale within monasteries. Liquor was referred to as “spirits” because distillation reduced liquids like beer and wine down to their core essence, their “spirits.” This meaning took on another meaning because it lifted human spirits. The word “whiskey” is derived from the Gaelic term “uisge beatha” meaning “water of life.” The very name of the product carries spiritual undertones. If we trace earlier than this, distillation was practiced, refined and perfected in the Islamic world. Beer and wine long predate the formal religious institutions we know today. We see their first historical consumption taking place in local religious rituals. Our ancestors did not have a name for “science” as a formal discipline but they were well versed in the arts of both cooking and fermentation — applied science. These applied sciences matured in a world intertwined with spiritual and religious custom. The further back we go, the more intertwined science and religion become.

History

The decoupling of religion from history is a very new development in human data. On one end, it does sometimes help to separate fact from fiction. However, in many cases, ancient history tends to preserve things through a cultural lens mostly lost since industrialization and even before it. We can trace the World Wars back to a time period when this decoupling really started to accelerate. England, which still technically has a national religion today was a dominant empire with state religion. The sun would set on one side of the British Empire as it rose in the other. An empire in which Anglicanism was the formal religion. In the near east, at roughly the same time period, we see the final remnants of the Ottoman empire — another near-modern example of nationalized religion. In the far east, before WWI, Russia was predominantly Orthodox Christian and much of their thought matured in such an environment. If we trace back to around year 1000, Western Europe was formally governed by a number of Christian nation-states. Eastern Europe was governed by Byzantium — the Eastern Roman Empire. In Byzantium, Christianity was formally institutionalized. If we go back to antiquity, we see Greek and Roman thought developing in a world drenched with the blood of sacrifice to local deities. We can go back even further in time using Egyptian Dynastic records. In Egypt, the Pharaoh was worshipped quite literally as a god. When historians didn’t go out of their way to avoid religion, records give us a much more textured cultural picture. In modernity, many nations are still trying to answer the question: “What is our culture?”

Reading and writing

Before we continue, let’s take a look at one other core subject taught in schools: Language Arts — reading and writing. Around the world, this core gets different names but its very essence is foundational to human data. Without literacy, data doesn’t exist at all. Today, literacy is so important that nearly every human system from third grade to fast food training just assumes we have it. It has been decoupled from religion for most of modernity and in many cases, it was already decoupled during antiquity. Once again, for the Western world, we need to rewind to medieval Europe. During the middle ages, literacy was a very rare skill. It developed primarily within monasteries and Christian clergy. Many historians subtly credit these institutions with preserving literacy through a very difficult time. If we rewind back further, we get to Rome and Greece. In the Greco-Roman world, some level of literacy was already assumed — much like what we see today. People were expected to be able to read public signage and understand legal decrees. If we go even further back, to a time when religion was directly intertwined with civil law, we get to ancient Israel as well as the Mesopotamian cultures. At this point, the lines blur immediately. In these pre-Roman societies, literacy was tied primarily to religion and law. If we go even further back into archaic human writing, we move past religious entanglement to the first universally accepted human religion — money and ledgers. Archaic human writing did not tell stories. It kept payment records. As soon as writing evolves into coherent story, religion saturates the record. Ledger systems provided an immediate safety layer during tribal disputes. Once we were safe enough to write coherent narratives, we immediately began by recording religious practices and primordial stories of human origin.

Conclusion: The scariest part

AI models are not religious. They are not the culmination of all evil. Satan is not going to conjure up out of your favorite chat app to battle Tenacious D for the pick of destiny. Can AI models quote satanic ideas? Absolutely. Can AI models quote religious scriptures? Absolutely. They can also quote political leaders, mathematicians, scientists, historians and great authors. AI models are powered by mathematical inference. Math. Yes, I’ll say it again: Math. What the math tells us makes modern humans uncomfortable.

AI is not religious. AI training data comes from overwhelmingly religious sources.

The final image above is called the Tel Zayit inscription. This is Paleo-Hebrew — one of the earliest decipherable human inscriptions. It tells no great story. This inscription gives us a glimpse of writing at its earliest. It’s simply an alphabet, similar to the ABCs that you learned in early childhood. Most alphabetic writing traces its ancestry from Phoenician and early Semitic inscription such as this.

Who generated this dataset?