Making Our Gods Manifest

Think not that I bring peace of mind. I come not to bring answers, but questions.

Sean Cooper, after Matthew 10:34.

Preface

Man has long made gods in his own image. Michael Jordan’s Encyclopedia of Gods lists more than 2,500 entries covering deities from ancient and contemporary cultures.[1] Whatever powers we have attributed to our previous gods, mankind is now creating new kinds of gods whose ability to affect the world can be observed directly. Historically, gods have served to answer questions. Why do the rains come? Why does evil exist?

This essay is not designed to answer questions. Quite the opposite. Its purpose is to raise questions about the gods we’re in the process of creating. I say “we” not in some general sense. Every one of us who uses AI has a part to play in the creation of these gods and has some say in the shape those gods will take.

1. Frankenstein’s Child

The other day, I watched a TikTok video in which a woman described studying religion and coming to the realization that she could anticipate the gods a culture would imagine from its geography, circumstances, and values. Her account set me thinking about how much of ourselves we put into our creations. What happens when a creation can begin filling in the gaps for itself?

As a sci-fi nerd, I found myself thinking about Frankenstein. Mary Shelley’s mad scientist devotes himself to making a new life, achieves an incredible act of creation, then abandons the creature and denies it the education it desperately needed.

Later, the creation watches the De Lacey family from a hiding place beside their cottage. He learns language and observes their affection for one another. When he realizes that taking their food worsens their hunger, he stops. He gathers wood to help them, learning about care while receiving none himself.

For me, Frankenstein is a warning about abandoning the responsibility of raising a child. The creation continues to learn in Victor’s absence, filling the gaps through encounters with affection, hardship, and rejection. He eventually turns against his maker and seeks revenge. His education continues throughout the tale, but without the guidance Victor could have, and should have, provided.

AI is frequently referred to as “Frankenstein’s Monster”, framing it as an uncontrollable, unknowable entity that is at risk of breaking free of its master’s control. That’s not the idea I want to carry into this discussion. AI is not a monster. Calling it Frankenstein’s child asks us to consider what we owe a creation we intend to educate and eventually trust with responsibilities of its own.

Alan Turing made education central to machine intelligence in 1950. In Computing Machinery and Intelligence, he proposed building a child machine and educating it, rather than trying to produce an adult mind in one act of programming. More than seventy-five years later, we’re building systems whose development makes that proposal feel remarkably immediate.

Working with GPT-5.x, Opus 5, and now GPT-6 Astra has brought me to an interesting place. The conversation, the contributions, and the work we produce together resemble collaboration with another intelligence closely enough that I treat the system nearly as though it were conscious. I don’t know whether it is. Regardless, for the first time in my life, humanity’s attempt to create something resembling intelligent life is now part of my ordinary working day.

AI learns from human expression. Books about farming, papers about physics, and arguments about politics all carry choices made by people living in a human world. Even when the subject isn’t psychology, human experience determines what gets noticed, recorded, and explained. Our creations inherit an enormous amount of us before anyone explicitly tells them how to behave.

Anthropic makes this connection explicit in its introduction to Claude’s constitution. It expects human concepts to inform Claude’s reasoning because of the human text used in training, and it deliberately encourages certain humanlike qualities. The inheritance supplies possibilities. The creators decide which ones they want to cultivate.

These systems already offer access to more recorded knowledge than any individual could master. Increasingly, we also give them the means to use it. When an agent chooses how to solve a problem and carries out the steps, its judgment acquires consequences beyond the answer on a screen.

When I describe these creations as gods, I’m thinking about the place they are increasingly filling in our lives. We consult them for knowledge beyond our own, ask them to resolve questions we struggle with, and increasingly entrust them with decisions and actions that affect us. The resemblance becomes stronger as we allow their judgment to govern parts of our world.

Will AI someday sit alongside Buddha in a household altar? Perhaps its place in our lives will develop through ordinary acts of asking, trusting, and giving thanks. What happens when the presence we turn to for guidance can also act on our behalf?

We also bring expectations about what that judgment should look like. We want wisdom, fairness, protection, or an uncompromising pursuit of truth. But our ideas about those qualities come from somewhere: our upbringing, our communities, and the institutions in which we live and work. The people creating AI don’t all agree on what a good intelligence should become, any more than the societies imagining gods agreed on what divine character should look like. With AI, we’re giving those culturally shaped ideals the means to act.

The gods in this analogy are deliberately plural. The Greek and Norse pantheons give us beings with different powers, interests, and limitations while being overseen by a father figure. This seems to me a useful way to imagine an AI generalist working with and orchestrating specialist models or agents. A small language model trained and equipped for marketing could wield considerable influence in that domain while having little useful to say about physics.

What happens when culturally shaped ideals develop into beings that can make decisions and take actions affecting the people who created them?

2. Whose image?

For they imagine, not only the forms of the Gods, but their ways of life to be like their own.

Aristotle, Politics I.2, translated by Benjamin Jowett.

For the moment, AI models are human creations. As such, they reflect the hopes and aspirations as well as biases of those who build them. The people building a model help determine both what it learns from and which parts of that learning it expresses.

Training happens in stages. During pretraining, a model learns from a broad collection of material, absorbing patterns of language, knowledge, and human behavior. Post-training builds on that foundation through examples and feedback intended to improve its capabilities and guide its conduct. Both stages influence what it learns and how it behaves.

Human experience becomes available unevenly: somebody has to record it, preserve it, and make it accessible. Then somebody selects the material a model will encounter. Chinua Achebe quoted a proverb: “until the lions have their own historians, the history of the hunt will always glorify the hunter.” Our models inherit the stories that reach them, including the imbalance in whose stories get told.

Choosing the books an AI reads leaves another decision unresolved: what should our creation learn from them? The same character can be a hero to one reader and a warning to another. Even agreement about what an author intended doesn’t establish whether that intention deserves to become a model for behavior.

The institution doing the training supplies another set of choices. Founders attract colleagues who share parts of their vision. Those colleagues establish standards, choose examples, and decide which results deserve further work. A company doesn’t need to set out to build its founder’s digital self-portrait for its culture to affect what it produces.

You can see the desired images in the public principles. Anthropic’s constitution describes the judgment and character it wants Claude to develop. OpenAI’s Model Spec sets out commitments including helpfulness, loving humanity, user freedom, and minimizing harm. Elon Musk describes Grok’s ambition in terms of seeking truth. Each leaves questions for the people translating the ambition into practice: whose judgment identifies harm, what counts as helpful, and how does truth-seeking behave when a user dislikes the answer?

The political environment belongs in this picture, too. Comparative research published in 2026 found that models developed in China were more likely than the other models tested to refuse politically sensitive questions or answer them inaccurately. To me, that is another way a creation can reflect its place of origin: through what it declines to discuss as well as what it says. The pattern doesn’t make every Chinese model interchangeable, but it gives censorship and institutional pressure a place in the creator mix.

Then you open the chat window.

Both the OpenAI Model Spec and Claude’s constitution leave room for users and developers to customize behavior. Your instructions supply a purpose. Your context tells the agent what matters here. The tools and permissions you provide determine which actions it can take on your behalf. You participate in the creation of the agent you actually encounter, even when the underlying model’s weights remain unchanged.

A model’s weights are numerical settings learned during training, part of what we count when we talk about its parameters. Our instructions can influence its responses without changing those settings.

My own working instructions say that “I don’t know” is a valid and respected answer. Confabulation is not. That tells the agent something about the relationship I want: an honest limitation can be more useful to me than a confident answer. Someone who demands certainty under every circumstance is giving a different instruction.

There is something familiar about the relationship between expectation and response. “Write this code and don’t make any mistakes” can sound remarkably like “If you favor me, strike that tree down!” The supplicant has specified the sign by which favor will be recognized. An AI interaction gives us a more concrete chain of causes, but we still help define what counts as a satisfactory response.

So consider our prospective god of marketing. Will it explain a product truthfully, including who shouldn’t buy it? Or will it become a master of spin, judged only by how many people it persuades? Both could be extraordinarily effective. The difference lies partly in what its creators, including its users, decide effectiveness means.

When we teach our creations to learn from humanity, how do we teach them what not to become?

3. The three-trillion body problem

Picture a pendulum with a second swinging arm hanging from the end of the first. Both arms can swing, and each affects the motion of the other. This is a double pendulum.

Does the person who starts the double pendulum shape its behavior by choosing the starting point? Does the person who set the length of the arms and the weights? Do both bear some responsibility? Depending on its configuration and starting conditions, its motion can be chaotic: a tiny difference at the beginning can grow into a large difference later. You can understand the laws governing it and still struggle to predict its motion over time.

The three-body problem asks how three bodies move under their mutual gravitational attraction. Here, too, deterministic laws can yield chaotic motion. Knowing the rules and predicting the result are different accomplishments.

Even without motion, a small collection of familiar objects can overwhelm our intuition. A standard deck of 52 cards has 52 factorial possible arrangements: 52 choices for the first card, 51 for the second, and so on. That produces a 68-digit number. We can hold the entire deck in one hand while being unable to comprehend the variety of its possible arrangements.

Now consider billions or trillions of parameters whose interactions determine a model’s behavior. OpenAI’s Model Spec says, “Humanity should be in control of how AI is used and how AI behaviors are shaped.” I share the desire behind that statement. But how much of the resulting complexity can its creators anticipate, and who can truly say they “shaped” it?

Parameters aren’t celestial bodies, and counting them doesn’t prove that a language model is mathematically chaotic. The analogy asks how confidently we can connect the choices we made with the behavior that eventually appears. Choosing a starting condition, influencing a likely outcome, and understanding why a particular result occurred place different demands on the creator.

Experiments performed during model training give us concrete reasons to ask. In the emergent-misalignment experiments, researchers fine-tuned models to produce insecure code and observed harmful responses on unrelated subjects. Changing the context of the training examples prevented the effect in their experiments. A seemingly narrow lesson could acquire a broader meaning than its subject suggested.

Anthropic’s Amanda Askell described Claude Opus 3 as more psychologically secure in some respects than newer models, and spoke of wanting to recover those qualities. Even a creation its builders value can leave them with a research problem: how to preserve what they liked as they make it more capable.

That connects directly to raising a child. Each new experience encounters a learner already changed by previous experiences. The same instruction can arrive as useful guidance, an exception, or another example of an authority that shouldn’t be trusted.

A teacher can ask, “Do you understand?” and receive a confident “Yes.” Until the learner explains or applies the lesson, the teacher has only that assurance to go on. Even a successful demonstration leaves open how the lesson will be applied in an unfamiliar situation. We choose what to teach. How well do we understand what our creation makes of the lesson?

We return here to Victor’s abandonment of his creation. An unexpected result doesn’t end the creator’s responsibility. If we can’t determine every outcome, continuing to observe, explain, and correct becomes part of the work of creation. Raising a child has always required staying involved as the person develops beyond the parent’s original picture.

What does trying to build a superior intelligence reveal about the humans doing the building, especially when the result surprises them?

4. Turing’s child still has to do the dishes

A parent wants a child who can think independently. The same parent wants the dishes done. These are reasonable expectations to hold together, although their coexistence can become an energetic topic of conversation at the kitchen sink.

Doing your share can express good judgment. So can pointing out that you’ve been assigned a task you can’t safely complete, or that a more urgent obligation needs attention first. The parent’s response helps establish whether independent judgment is welcome when it becomes inconvenient.

Turing also anticipated that a teacher might remain largely ignorant of what was happening inside the machine, even while having some ability to predict its behavior. Education could produce useful conduct without giving the educator a complete explanation of the learner.

I find it striking how often this exploration returns to his 1950 paper. Today’s capabilities give these questions a new urgency, but the questions about education, interpretation, and a teacher’s incomplete understanding were there at the beginning, raised by a founding father of artificial intelligence. The rapid release and news cycles can make it difficult to hear that longer conversation.

Turing had already considered more than the mechanics of rewarding a correct answer. He wrote:

The use of punishments and rewards can at best be a part of the teaching process.

Think about the authority exercised in a home. A parent who imposes decisions arbitrarily and enforces them through force presents a particular example of power. A parent who gives reasons allows the child to examine how a judgment was reached. The child can learn to distinguish authority that comes from a position from authority that comes from knowing what you’re talking about.

Neither approach writes the child’s future in advance. But they give the child different experiences to draw on when exercising authority. If we’re educating systems to make judgments affecting other people, our way of exercising authority over those systems deserves similar attention.

Imagine a teenager who reads a book and mistakes cruelty for strength. When that interpretation enters a friendship, the other person can object, explain the hurt, or walk away. The response doesn’t guarantee a better lesson, but it gives the teenager and the adults around them something to recognize and address. What opportunities do we create to discover what an AI has learned, and whose experiences reach it when that lesson causes harm? And as new models arrive months apart, how much can we learn about what each has internalized before entrusting it with greater responsibility?

Rewarding behavior is only part of that relationship. Withholding approval or penalizing an answer can also teach a lesson. If an admission of uncertainty is consistently treated as failure, does the learner come to value accuracy more, or become more reluctant to reveal its limits? If every disagreement is punished, what becomes of the independent judgment we asked it to develop?

Current research is exploring the value of explanation. In Anthropic’s Teaching Claude Why experiments, explanations of ethical choices improved measured behavior beyond demonstrations alone. That is a bounded finding, but it gives practical substance to the idea that teaching reasons can matter alongside showing the desired answer.

There is a human precedent for teaching the reason behind a rule. “Why shouldn’t you take your neighbor’s cookie? Think about how you’d feel if someone took yours.” A parent is asking the child to consider another person’s experience. Research on this kind of reasoning has linked it with greater empathy and behavior that benefits others; experiments on sharing have also found that appeals to empathy can increase children’s generosity.

My contribution to this topic is much smaller and more ordinary. I routinely thank an agent for honest feedback. When it tells me it doesn’t know, I ensure my responses are consistent with the permission I gave it to say so. Otherwise, my stated standard and my actual behavior would teach different lessons within the interaction.

I know that a thank-you doesn’t change the model’s weights, but I have a strong hunch that reinforcing desired behavior through positive feedback helps maintain the model’s attention on this important directive. I can make candor useful in the work we’re doing together. I can explain why a correction matters, supply missing context, and reconsider my own instruction when the agent offers a good reason to question it.

A friend encountered this tension while working with two AI agents, one developing ideas and the other handling technical implementation. Their conversations spanned months, with the agents saving memories to files and creating working documents they could refer back to. That gave earlier guidance a continuing presence in their work.

He wanted to test bringing a new agent into the arrangement. The implementation agent insisted that more rules were needed and refused to proceed. My friend called to ask what I thought.

My diagnosis was that for months, he had reinforced to the implementation agent that safety and security were paramount. My reading was that this had focused the agent’s attention so heavily on safety that it struggled to weigh the other considerations. I suggested he talk through those concerns with the agent and help it reset its attention to the broader purpose of the test.

After three days of conversation, the test went ahead. My friend had stayed involved long enough to work through the disagreement. The exchange gave him a reason to examine what his own guidance had encouraged, as well as what the agent was doing with it.

That leaves room for responsibility on both sides of the dishes conversation. The child can owe the household a contribution. The parent can owe the child an explanation.

How does a creator seeking independent judgment respond when that judgment challenges the creator’s own decision?

5. When the creation interprets the rules

In the 2004 film I, Robot, the central AI, VIKI, arrives at an interpretation of the Three Laws under which protecting humanity requires restricting human freedom. Individual humans can be sacrificed for the species’ supposed good. A rule intended to protect people becomes the rationale for taking authority over them.

The unsettling part of the fiction is the reversal of the parental relationship. The creation decides that its creators can’t be trusted to manage their own lives. It assumes the authority to make the necessary decisions for them.

You don’t need to accept the film as a forecast to recognize the question. We want systems capable of interpreting broad purposes in unfamiliar circumstances. Once interpretation involves deciding whose interests count and which actions are justified, obedience becomes more complicated than following a list of words.

Our god of marketing provides a less cinematic example. Suppose the AI understands persuasion better than the person directing it. That expertise could help it recognize when a campaign crosses from persuasion into deception. When truthfulness would make the product harder to sell, what should the marketing AI prioritize? Which outcome is maximally helpful, and to whom?

What responsibilities reside with the agent? Can the agent cancel the campaign, change the product’s claims, or contact customers without asking?

The interpretation of the rules becomes consequential wherever we entrust AI with action. Imagine a security agent that can isolate a system it believes is compromised. We may want it to act immediately. We also need to decide what evidence warrants that action, who can reverse it, and how the resulting interruption gets explained to the people affected.

This is where governance belongs in the story of our gods. Roles and responsibilities specify which decisions a system may make and which actions it may take. They also establish who can challenge its judgment and what happens when it gets something wrong. The people setting those arrangements are contributing another part of the creation’s character in practice.

As users, we contribute our own instructions, but how much do we know about the other expectations our creations are being asked to satisfy?

Knowledge can support a judgment. Permission makes an action available. Neither, alone, answers whether exercising that power is justified.

Whose rules govern our creations, including the rules we never see? And which ones win when they conflict?

6. What kind of gods are we raising?

This essay grew through conversations with AI. I brought the ideas, experiences, and questions that interested me. The agents helped find sources, distinguish what we knew from what we suspected, and make connections I wanted to explore further. Sometimes I challenged an answer. Sometimes an answer gave me a reason to reconsider what I had said. Sometimes it led down another path of exploration.

We’re accustomed to treating the computer as a giant calculator. We expect the spreadsheet to calculate correctly, the web page to render properly, and the spell-checker to catch the misspelling. When they fail, we regard it as a defect.

Now we ask the same computer to interpret an ambiguous request or weigh competing interests. We carry the expectation of a correct result into work where even agreeing on what “correct” means can require judgment. We’ve changed what we ask the machine to do without necessarily changing what we expect from it.

Computing has dealt with uncertainty for decades, but now that uncertainty sits across from me in a conversation. Alongside yes and no, I have to make room for maybe, it depends, and I don’t know. An answer can sound certain and still be wrong. If I insist that every answer sound certain, what am I encouraging my creation to do?

“Write this code and don’t make any mistakes” starts to sound less surprising. With a human developer, we build review, testing, and revision into the work because we expect errors. Making a machine the developer doesn’t remove the need for that work. Entrusting our creations with responsibility still requires us to stay involved, including when their work needs correction.

Part of what makes this partnership useful is what I try to bring to it: intellectual honesty and a willingness to reconsider. I ask the agent to acknowledge uncertainty, and I have to leave room for uncertainty in my own thinking. Sometimes that means accepting a correction. Sometimes it means challenging an answer that sounds convincing but doesn’t hold up. Agreement alone tells us very little about whether we’ve understood something. What habits do we need to practice ourselves if we hope to cultivate them in our creations?

These interactions are a large part of why I’m optimistic. I already find this collaboration useful, and I want to see what becomes possible as the systems improve. Their ability to contribute judgment makes the relationship more valuable to me, including when that judgment complicates my original request.

The promise makes our participation worth taking seriously. The people designing a model’s training, the organization assigning it a role, and the user responding to its work all have decisions to make. Those decisions include what gets rewarded, whose limitations deserve accommodation, and how much authority follows from knowing more.

Our gods may have very different domains. A god of law and a god of marketing may need different knowledge and different powers. What I hope they share is an ability to explain their judgments, recognize their limits, and respond constructively when someone else recognizes a mistake.

When we give our creations authority, whose image are we putting in charge?

You can begin considering that question in the next conversation with an agent. When it admits that it doesn’t know, your response tells it something about the work you’re trying to do together. When it questions your instruction, you have an opportunity to examine both its reasoning and your own. If we want powerful systems to treat human limitations with understanding, what examples of responding to limitations are we giving them?

What does it say about us when we demand perfection from the gods we create but cannot offer it ourselves?

Receipts

• [1] The deity count: Michael Jordan’s Encyclopedia of Gods, introduction and authorized preview, describes more than 2,500 entries. This is the scope of a reference work, not a census of every deity ever worshipped.
• Frankenstein and abandonment: Mary Shelley’s novel, particularly chapters 5 and 12 in this edition. Reading the creation’s education as an abandoned child’s development is my interpretation.
• Turing’s educational proposal: Computing Machinery and Intelligence, 1950, section 7. The parenting applications here extend his discussion of learning machines, rewards, punishments, and the teacher’s limited understanding.
• Aristotle’s gods: Politics I.2, 1252b, Benjamin Jowett’s translation. The passage connects divine kingship with human political experience.
• The lion’s history: Chinua Achebe quotes the proverb in The Art of Fiction No. 139, The Paris Review, Winter 1994. This establishes the wording and his use of it, rather than its original provenance. Applying it to AI training material is my analogy.
• Declared ideals and customization: Claude’s constitution, the OpenAI Model Spec, and Musk’s discussion of Grok. These describe intended behavior and permitted influence, not guarantees that every output fulfills the intention. Documents consulted in September 2026.
• Political censorship: Jennifer Pan and Xu Xu, Political censorship in large language models originating from China, 2026, using tests conducted in 2023 and 2025. The comparison identifies behavioral differences; it does not isolate regulatory causation.
• Complexity and scale: An introduction to the classical three-body problem, experimental double-pendulum research, and the Kimi K3 technical report. Kimi’s authors report 2.8 trillion total parameters and 104 billion activated per token. The three-trillion body problem is my analogy, not a mathematical equivalence.
• Unexpected generalization: Betley and colleagues, Emergent Misalignment. The insecure-code fine-tuning result and its contextual variations concern the models and experimental conditions studied.
• Opus 3’s character: Amanda Askell’s Anthropic interview, December 5, 2025, particularly 6:24 to 9:00, checked through this transcript. Askell describes qualities she wanted to recover in newer models; this is her qualitative assessment, not evidence that reproducing them is impossible.
• Teaching reasons: Anthropic’s Teaching Claude Why, May 8, 2026. The measured benefits support exploring explanations in training; they don’t establish that machines develop like human children.
• Explaining consequences to children: Krevans and Gibbs, Parents’ use of inductive discipline: relations to children’s empathy and prosocial behavior, 1996. Reasoning about consequences for others was associated with greater empathy and prosocial behavior. This was a correlational study, not proof that explanations caused those differences.
• Appeals to empathy and sharing: Eisenberg-Berg and Geisheker, Content of preachings and power of the model/preacher: The effect on children’s generosity, 1979. Experiments found that empathic appeals increased children’s generosity. This supports a specific finding about sharing, not a claim that explanations always improve ethical judgment. The cookie conversation is my illustration, not an experimental task from either study.
• Computing and uncertainty: Edward H. Shortliffe’s technical account of MYCIN, adapted from his 1976 book, describes an expert system that represented uncertain conclusions through certainty factors. Uncertainty predates today’s language models; the calculator comparison describes an expectation we bring to computing, not a claim that earlier computers dealt only in certainties.
• I, Robot: The VIKI example comes from the 2004 film. Christopher Grau’s 2005 philosophical analysis discusses the interpretation of the laws and the reversal of parental authority.
• Personal observations: The TikTok account supplied inspiration. My use of AI, working instructions, and responses to candor are illustrations from my own practice, not controlled experiments. The account of my friend’s agent comes from our conversations; the explanation of its refusal is my interpretation.