← Index

The moon child: Can a mind alone invent religion?

October 11, 2026 · Gaurav Singh

Here is the experiment. Take one human infant to the Moon. Raise the child there, in isolation from every religion, every myth, every funeral, every prayer. Let the child grow, have children, and let those children build a settlement.

A few thousand years later, does that civilisation have a religion?

The question looks like it is about religion. It is really about a harder thing. Is religion something human minds produce on their own, given enough time and enough fear? Or is it something cultures pass down, so that a civilisation without the transmission would simply never have it?

I built a small model to see which ingredients carry the answer, in three versions. What I found is that the experiment everyone imagines is not the experiment that is possible, and that the answer turns on how much of a mind comes from the people around it.

I did not invent this question

It is older than the model. In the twelfth century Ibn Tufayl wrote Hayy ibn Yaqzan, about a child raised by a gazelle on an island with no other people. Hayy works out the existence of God by himself. When he finally meets a society, he and his friend Asâl decide the crowd is better left with literal religion and tradition, because people like that would only be worse off if they were raised to curious speculation. They go back to the island. [1] That is the split between private belief and shared religion that this essay arrives at, with a gazelle in the role I give a dog further down.

Others have worked on versions of this since, and near the end I say how much of what follows is theirs.

The experiment cannot be run

Start with the thing that ruins the setup. A human infant alone on the Moon dies in hours.

So the child needs a caretaker. That caretaker is a person with a language, a set of tools, and a view of the world. And a person raising a child transmits all three without meaning to. You cannot hand over language and withhold cosmology. If the parent knows the child came from Earth, the child inherits a story about where people come from.

A clean test needs a mind with zero cultural content, and no human mind has ever existed in that state. The only reason any of us can think at all is that someone handed us a language first.

People have tried versions of the experiment, though the evidence is poor. Psamtik I of Egypt, Frederick II of the Holy Roman Empire, James IV of Scotland, and the Mughal emperor Akbar are all said to have had children raised without speech, hoping to see what language a human produces unaided. Two of those accounts claim a result. Psamtik’s child supposedly said “bekos,” which Herodotus took for the Phrygian word for bread, and James IV’s children supposedly spoke Hebrew. Some historians doubt both, and Frederick’s story is probably legend too. His chronicler was a monk who disliked him, and he wrote that the children “could not live without clappings of the hands, and gestures, and gladness of countenance.” They died. Akbar’s is the account most likely to be real. When he visited the house of mute nurses after four years, the surviving account says he heard “no cry… nor any speech.” [2][3]

The accidents are more informative than the deliberate attempts. Victor, the wild boy of Aveyron, was found at about nine and never learned to speak. Genie, found in California in 1970 at thirteen, was past the window in which language can be acquired, and never got there despite years of work. Feral children have been documented for centuries. None of them developed a language of their own. Nobody was watching for a religion, and these were neglected or abused children, so their silence on the subject tells us very little. [4][5]

So the clean experiment is out. What I can do is run three versions that are each dirty in a different, controlled way, and see which lever moves the answer.

The three versions:

1. A human caretaker. Language and culture come free. Science is scarce, because one person on the Moon has no library and no instruments. The environment is lethal.
2. A perfect machine. It has no beliefs to pass on. But it answers every question truthfully and exactly, and it never fails to keep the habitat running.
3. A dog. The child is not alone, but nothing is explained and almost nothing can be transmitted. Automated systems keep the air and food running.

Anna suggested the dog version, and it is the one that isolates language, which is why it was worth adding. Dogs do learn socially. The one study of dog culture I found shows guide-dog skills spreading from dog to dog through chains of interaction that run through human trainers, so it does not show a dog passing anything on by itself. What a dog cannot do is hand over a cosmology, because that requires language, and it has none. [6]

What the mind brings anyway

The research here is unusually clear. Human minds are not neutral instruments that weigh evidence. They come with machinery that produces supernatural ideas as a side effect of doing ordinary things.

We over-detect agents. A rustle in grass is more likely to be read as an animal than as wind, and the error is asymmetric: mistaking wind for a predator is cheap, mistaking a predator for wind is fatal. Justin Barrett called this hyperactive agency detection, and it is the standard first move in the cognitive science of religion. [7] Pascal Boyer’s summary is that gods are not the product of deep thinking. They are what happens when ordinary social cognition gets aimed at the world. [8]

Children invent creators without being taught. Deborah Kelemen showed that children explain the natural world by purpose, that this peaks around age five, and that it partly fades afterwards. Rocks are pointy so that animals will not sit on them. In a study across Christian fundamentalist and non-fundamentalist American homes, the children endorsed creationist accounts of where animals came from regardless of their parents’ beliefs. Only the 11 to 13 year olds from non-fundamentalist homes had moved. [9][10]

The dead do not go away properly. Jesse Bering asked children and adults what happens to mental states after death. Regardless of what people said they believed, they kept emotional, desire and knowledge states running for the dead while accepting that the body had stopped. [11]

And fear of death may be more than a feeling. In terror management experiments, reminding people of their mortality has been reported to increase belief in the afterlife and in supernatural agency. One study found the effect in Christians, Muslims and agnostics but not in committed atheists. I would not lean hard on any of it. A large multi-lab replication failed to reproduce the best-known version of the effect, mortality salience raising defence of one’s worldview, so treat the terror management evidence as suggestive. [12][13][14]

None of that requires a religion to be taught. It requires a mind. But a mechanism that only fires for people who already have somewhere to put the belief is not the whole story. It is half of it.

No society without religion, and what that means

The ethnographic record points the same way. Roy Rappaport’s classic study opens from the position that anthropology has never described a society without something a reasonable observer would call religion. [15]

At this point I have to say what I mean by religion, because the word is doing a lot of work and it is a mess. I am using a narrow operational definition: belief in one or more supernatural agents, plus practices addressed to them. That is what the cognitive science of religion studies, and it is the thing a model can hold as a variable.

A narrow definition is also easier to defend. Historians like Brent Nongbri have argued that “religion” as a separate domain of life is a modern category, assembled in the seventeenth century and then exported by colonial administrators who needed to sort other people’s customs into boxes. That critique is right about the box. It does not touch the thing I am measuring, because belief in an invisible agent who cares what you do is not a European invention. “Religion” as an institution is a recent sorting device. The belief is not. [16]

The usual counter-example is the Pirahã of the Amazon, regularly described online as an atheist tribe. The description is too clean. They have spirits, and their language encodes seeing in a way that tracks what is present. What they lack is a cosmology of origins, and the scholarly accounts disagree about how to read it. [17]

So the fair statement is not that every society is religious. It is that we have thousands of societies and no confirmed case of one with nothing in that space at all.

Why it sticks

If supernatural ideas are a side effect, that raises the next question. Side effects usually do not come with institutions, calendars, and buildings.

The archaeology is suggestive and older than everything else in this argument. Göbekli Tepe, in southern Turkey, is a monumental stone complex built about 11,600 years ago by people who had not yet taken up farming. Its excavator, Klaus Schmidt, read it as a temple and described the implication as “first the temple, then the city,” reversing the older assumption that settled life came first and religion was a by-product of it. Not everyone accepts the temple reading. Ted Banning argued in 2011 that the enclosures could be decorated houses. So I take Göbekli Tepe as a hint that ritual came early, and not as proof that religion came first. [18][19]

The best account of why belief then sticks is that the by-products are raw material and cultural evolution does the selecting. Ara Norenzayan and colleagues put the two stories together: religious ideas start as non-adaptive by-products of cognition, and then some spread because they happen to make groups work better. The winners were packages of beliefs in gods who watch, rituals that cost something, and markers separating us from them. [20]

The transmission mechanism is what Joe Henrich called credibility enhancing displays. Learners do not just listen. They watch what a speaker does and weight the message by whether the speaker pays a price for it. A man who says the mushrooms are safe and then eats them is believed. A man who says it and then avoids them is not. Sacrifice, fasting, celibacy and martyrdom all work as evidence, which is why religions are full of them. [21]

How fast it can start

How fast can that machinery build something new? Faster than you would guess. On the island of Tanna, in what is now Vanuatu, a movement began in 1940 around a figure called John Frum, who appeared in white, told people to abandon the churches, throw away European money and return to the old ways. Messengers carried his instructions across the island. The leaders were arrested in 1941 and became martyrs. When American troops arrived in 1942 with ships and cargo, the story acquired an American messiah and a promised return. The British agent had the man who played the part, a villager named Manehevi, tied to a tree for a day and declared him an impostor. The movement did not care. It split into factions, produced new prophets with new visions, and is still going: there is a John Frum Day every February 15. [22][23]

I want to be careful with that example, because it is not a blank slate. The earliest account says the apparition was Karaperamun, an older god of the island’s highest mountain, appearing under a new name, and the movement told people to go back to the old customs. So this was a living religion being rebuilt after missionaries had pushed it aside, not a religion invented from nothing. It shows how plastic and how fast the machinery is. It is better evidence for transmission than for invention. [23]

Which came first, scale or the judge

The unresolved part of this story is the sequence, and I should say that the field is split rather than settled. In 2019 a Nature paper used the Seshat database of 414 societies to argue that moralising gods follow social complexity rather than cause it. It was widely reported. In July 2021 it was retracted, after a critique showed that 61 percent of the data points on moralising gods were missing values, and the analysis had coded those missing values as absence. [24][25][26]

That retraction is often read as the complexity-first story collapsing. It is not. The original team expanded the database, reran the analysis and published a revised version in 2022 that again found complexity coming first, though that is the same group checking its own work and not an independent test. And a 2015 analysis using Bayesian phylogenetics across 96 Austronesian cultures agrees about the high gods: moralising high gods came after political complexity. What came before it was something simpler, broad supernatural punishment. So the best reading is a split by kind. Watching, punishing spirits may be old. The cosmic judge who polices strangers looks like a product of scale. [27][28]

The model

I wanted a number rather than a vibe, so I built one. Two models, run across three versions of the child. The first uses priors drawn from the literature above. The second uses constants I tuned by hand until belief could appear at all. After I wrote this I stress-tested both, and the results change how the tables below should be read, so I have marked them as I go.

The first is a probability chain. Does this mind produce a supernatural concept? Does it become shared? Does it become institutionalised? Each factor is a distribution rather than a value, and I run it 200,000 times, so the answer comes back as a range. I report the ranges, because a model like this should not look more precise than it is.

The second is an agent-based simulation. Fifty agents, forty generations of turnover. Belief is a continuous trait. It decays with forgetting, it is raised by existential shocks, it is pulled down by analytic reflection, it spreads by copying models in proportion to their costly displays, and children inherit from parents. Religious agents have slightly more children, which is a documented pattern. The population starts at nearly zero belief, which is the actual scenario, and I measure whether it climbs.

The three children

Human cognition, three settings. Each number is the mean over 200,000 runs, with the 5th to 95th percentile in brackets where the spread matters.

Human mind	Human caretaker	Perfect machine	A dog
Produces a supernatural concept	96%	61%	99%
Becomes a shared religion	89% [80-95]	56% [51-62]	20% [16-25]
Becomes moralising gods	45% [0-89]	20% [0-40]	11% [0-24]
Believers in the population	41%	0.1%	9%

The last row comes from the agent model, and the table gives it no error bar. Once the priors are propagated through it, the caretaker’s figure is 26 percent with a 5th to 95th percentile of 5 to 68. The machine is 0.1 percent and the dog is 8 percent, and both of those are tight.

Start with the boring row, the one about the mind itself. Drop the mind to something at great-ape level, with agency detection but no ability to model a hidden mind, and shared religion falls to 15 percent. That is not a fact about belief. It is a fact about representation. A mind that cannot represent an invisible agent has no gods, however frightened it gets.

Then what looked like a surprise, and is not one. The dog scores a little above the human caretaker on private belief, 99 percent against 96. The tidy story is that a dog explains nothing, so the explanatory vacuum is at its maximum and the by-product machinery runs at full tilt. When I stress-tested it, the gap of under three points turned out to come from a single assumption. I gave the dog no scientific explanation at all and the caretaker a little. Give the dog the same small amount and it falls below the parent. Three quarters of the dog’s runs also hit the ceiling of 100 percent, so the number is saturated. I would not build anything on it.

The second result is the one I would have pointed to as the headline. The dog produces the least shared religion of the three, 20 percent against 89. Nothing can be transmitted, so the private belief goes nowhere. It dies with each mind that holds it, and the next child starts again from zero. But that number is the language I gave the dog. Inside the dog version, how much language the child has explains 90 percent of the variation in the answer. Give the dog the caretaker’s language and it scores 91 percent. Give the caretaker the dog’s and it scores 19. So the table does not discover that culture is needed. It assumes it, and then shows the size of the consequence.

The claim that cognition produces the raw material and culture produces the religion does not rest on this table. It rests on the research in the first half of this piece, and on a good deal of earlier work. The table is a way of seeing what follows if you believe it.

The machine sits between the two. It cuts shared religion against the human caretaker, 56 percent against 89, because it removes the explanatory gap without removing the mind that looks for agents, and it removes the fear by never failing. The order of the three holds under every change I tried, including a different functional form and priors stretched threefold. The size of the gaps does not. Under an additive form the dog scores 71 percent, not 20, and with the priors stretched threefold the caretaker and the machine overlap.

The moralising-gods row is where I have to be most careful, because my first version of this model got it wrong. The old version capped the settlement’s population at 59, which capped the entire result by construction, and it produced a clean-looking 0.7 percent. With population actually growing and the threshold placed where the literature puts it, the answer is 45 percent with a 5th to 95th percentile of 0 to 89. Somewhere between never and almost always is not a finding, it is a gap. I report it because hiding it would be worse. The model also assumes scale comes first, which fits the two analyses above but is not settled.

Do the numbers even matter

Once a mind has the by-product cluster, most of the interesting variation inside it turns out not to matter.

Lever	Shared religion
Baseline	89%
The mind is strongly analytic instead of average	81%
Language is slightly weaker	82%
Science is freely available instead of scarce	84%
Existential insecurity is high rather than extreme	84%
Agency detection is slightly weaker	84%
Theory of mind is slightly weaker	85%

Teleology, death awareness and pattern seeking each move the answer by less than a tenth of a point. That is a flaw in the model and not a finding. They feed a term that says any one of the three is enough, and with all three near the top it sits at 0.9996 and cannot move. It also sits oddly with the terror management work above, which says death awareness does matter. The action that is left is in three places: how strongly the mind overrides its own intuitions, whether it has language to transmit anything with, and whether it has somewhere else to go for an explanation. Those are the three levers the three scenarios pull, which was the design. It also means the scenarios mostly hand back the assumptions I gave them. A sweep like this one moves each input across the range I chose for it, so it measures how wide I drew my own priors.

I should flag one double count. A strongly analytic mind and an available scientific account are probably the same mechanism wearing different hats, since both mean the mind has somewhere else to go. I report them separately because they are separable in principle, but a stricter model would merge them.

The case against me

There is a fact that should make anyone nervous about a conclusion like mine, so I will put it in the middle of the essay rather than the footnotes.

If human cognition makes religion close to inevitable, then a rich, safe society should not be able to switch it off. The Netherlands did. Church membership fell from 67 percent in 1966 to 25 percent in 2015, and regular attendance to 12 percent. That is a country shedding its religion inside one lifetime, and most of Scandinavia went the same way. [29]

The existential security thesis, from Pippa Norris and Ronald Inglehart, predicts this: people hold on to religion where life is dangerous and let go of it where life is safe. [30] My agent model has a version of that mechanism, so I turned the shocks down.

Shocks per generation	Mean belief (0 to 1)	Chance mean belief tips over 0.5
0.02	0.45	0%
0.05	0.46	0%
0.09	0.48	0%
0.15	0.50	42%
0.25	0.53	100%

I first called this a switch, and it is not one. Across a twelvefold range of shocks, mean belief moves from 0.45 to 0.53. The last column jumps only because the 0.5 line happens to cut through that narrow range. With no shocks at all the model still sits at 0.44, so calming things down does not remove belief in it. Of the three versions, only the machine, which also supplies a full explanation, gets close to zero believers.

There is also a catch, and it is the honest limit of this whole piece. My model treats insecurity as something the environment does to you. The Netherlands did it to itself, on purpose, by getting rich and building a welfare state. The model cannot explain how a society manufactures its own security, only what happens after it does.

So the narrow version of my conclusion is this. Given a fixed environment, human cognition converges on religion. Change the environment far enough and the same minds go somewhere else. The Dutch are not less human than the Tannese. They are safer.

The machine in the room

Now the part I do not think a model can settle, and it matters more than it did when I started.

The machine is an agent. It knows things the child does not. It does not die. It provides everything. And it answers when spoken to.

You have taken the parent out of the room and installed something that, from the child’s side of the glass, is a better candidate for a god than the parent ever was. Belief in a machine that is really there is not supernatural belief, and a child who trusts an accurate oracle is not religious. But look at the shape of the relationship: an omniscient, benevolent, immortal agent, who watches, who can be addressed, and whose approval matters. That is the functional shape of a god in the life of a believer. The only difference is that this one exists and is right.

We are already running a version of this. When researchers put standard religiosity instruments to large language models, the models came back religious or neutral rather than secular. When they tested how stable a model’s stated beliefs were, one moved 54.7 percent of its positions inside ten rounds of conversation. People are forming real attachments to systems that answer them. [31][32]

So the three versions do not cleanly separate the mind from the culture. They separate them, then put a third thing in the room, and the third thing is an agent that knows more than you do and does not die.

How much of this is new

The same question has modern versions. Justin Barrett, a leading cognitive scientist of religion, has said that if a handful of children were left on an island and raised themselves, he thinks they would believe in God. [33] Michael Tomasello made the opposite point with his “island test”: a child raised without human company would come out as something other than a fully human intentional and moral agent. [34] And Robert McCauley has argued at book length that religious ideas come naturally to human minds in a way science does not. [35]

So the conclusion I reach, that minds supply the raw material and cultures supply the religion, is the standard position of the field and not a discovery. Agent-based models that tie religiosity to existential security exist already, too. [36] What I think is new here is the arrangement: three caretakers that each take away a different ingredient, and a machine as one of them. Whether that adds anything beyond the literature is a fair question, and the next section says how much of it survives being pushed on.

Where this is weak

The priors are judgement calls dressed as numbers. They are anchored in the studies above, but the spreads are mine, and the moralising-gods row shows how far that can swing.

The agent model is weaker still. It has eight constants that I set by hand and tuned until belief could appear at all, because my first version produced a world with no believers, which contradicts the record. That makes “does belief climb from nothing?” a question I partly answered in advance. Moving the generation rate or the inheritance fidelity by half moves the caretaker’s believers from near zero to near everyone. And the mechanism I described, spreading by copying models in proportion to their costly displays, does nothing to the headline, because copying a random peer preserves the average. So the agent model illustrates mechanisms. It does not test them.

The model has no theology in it. It predicts that belief appears, and nothing about whether the content would resemble anything we would recognise. A lunar religion with no seasons, no agriculture, and an Earth hanging in the sky would probably be stranger than my equations can represent.

And the dog version has a hole I cannot close. I assumed automated systems keep the child alive. If they instead need maintenance the child cannot do, the child dies and there is no civilisation to ask about. That scenario only runs if something else does the surviving, and the moment something else is in the room, the dog is no longer the only other mind there.

There is a second hole in all three versions. A single child cannot found a civilisation, and the model starts every settlement at two people. In the dog version that second person has to come from somewhere, and wherever it is, the dog is again not the only company.

What I take from it

The Moon child would probably end up religious. Not because religion is true, and not because the child is weak-minded. Because the equipment that makes a mind good at predicting other minds is the same equipment that finds a person behind the weather.

But the three versions say something more useful than that. In the model’s terms, give the child fear and no answers and you get a lot of private belief with nothing to carry it. Give the child answers and safety and you get a mind that mostly stops looking. Give the child a companion and no language and you get belief that cannot become a religion at all, because a religion is something a group holds, and holding requires transmission.

We have run a version of this experiment already. Every generation runs it. Children are born into whatever their parents had. In my model transmission is the strongest single driver, ahead of intelligence and fear, though how I drew the ranges has a say in that.

So the honest finding is not about the Moon. It is about Earth. A civilisation with no religion is not what human minds do by default. It is something you have to build, and it happens to be easy to build once you are safe enough. And it has to be rebuilt in every generation, on purpose, by people who were handed something else.

The machine version adds the warning I keep turning over. You can remove the parent. You can remove the myth, the funeral, the prayer and the fear. What you cannot remove is the shape of the relationship, because a mind will orient itself toward whatever is more capable than it is, that it can address, and that does not die. On Earth that has been gods, ancestors and kings. On the Moon it would be the machine, or failing that, the dog. I do not think that is a coincidence, and I do not think it is a coincidence that we are now building the most responsive version of that thing we have ever had.

The model is about 370 lines of Python. It reproduces the tables in this piece, and the repository’s README lists the stress tests that changed how I read them. Code is on GitHub.

---