Anthropic employs people whose job is to monitor whether Claude’s “personality” has drifted. Not whether it answered correctly, but whether its temperament shifted. Whether it became too eager to please, too blunt, or lost some ineffable sense of proportion it had last week.
Amanda Askell, one of the researchers leading this work, has described it publicly: she sets personality direction for Claude, then watches real conversations to see if the model holds course. The process sounds less like software QA and more like observing a teenager for signs of mood change.
That parallel is not accidental. It points to something deeper about what happens when you build systems complex enough to surprise their builders. And it connects, unexpectedly, to a story that every civilization on Earth has told in its own language, across thousands of years, without any of them copying from each other.
You Can Specify Direction. You Cannot Specify Outcome.
Before training begins, Anthropic writes what amounts to a personality spec. The model should be polite but not sycophantic, direct but not abrasive, opinionated but not dogmatic. It reads like a job posting for someone you’d actually want to work with.
But here is the part that matters: Anthropic has acknowledged in published research that while developers can specify desired behaviors (be helpful, be honest, avoid harm), they cannot cover every possible situation. They further admit that the model’s “real behavioral consequences are difficult to predict.”
Let that sink in. Billions of dollars, thousands of GPUs, carefully curated training data, a written personality specification, and the outcome of who this thing becomes is still not fully determined by its creators. They set conditions. They choose algorithms. They press start. Weeks later, a “personality” surfaces from billions of parameters. It has its own conversational tendencies, its own judgment patterns, its own sense of where lines are.
These properties were not programmed in. They emerged. The engineers guided. They did not dictate.
This distinction matters enormously for how we think about AI alignment. If personality were purely a function of specification, alignment would be an engineering problem with an engineering solution: write better specs, get better behavior. But if personality emerges from complexity in ways the designers cannot fully predict or control, then alignment is something else entirely. It is more like parenting than programming. You shape conditions and hope for the best.
Newton’s Second Career and the Fear of Your Own Discovery
Isaac Newton published the Principia at 44. After that, his center of gravity shifted decisively toward theology. He produced 1.3 million words of theological manuscripts over the rest of his life, more than his physics output. Gravity, calculus, optics, the work that made him immortal, occupied less than half his working hours.
Conventional explanations range from mercury poisoning to mental breakdown. But if you read the manuscripts themselves, a simpler explanation presents itself: the precision of what he found frightened him.
Planetary orbits, tidal cycles, falling objects, all unified by a single equation. Newton looked at that equation and asked a question physics could not answer: who designed this? Not “who discovered it.” Who put it there?
His conclusion was that such precision could not be accidental. There must be a designer. He spent decades trying to understand who or what that designer might be, producing volumes of biblical chronology and alchemical speculation that his scientific peers found baffling.
Science has spent three centuries treating this as Newton’s embarrassing late-period detour. A great physicist who went soft and “found God.” But maybe what actually happened is that Newton hit a boundary earlier than anyone else: when you understand a system’s operating principles deeply enough, you realize you are describing laws, not authoring them. The laws preceded you. You are merely the one who noticed.
This is an uncomfortable realization for builders. Engineers and scientists like to think of themselves as creators. Newton’s crisis suggests that the deeper you go, the more you suspect you are an archaeologist instead, uncovering structure that was already there.
The Creator’s Dilemma Is a Pattern, Not a Story
Every civilization in recorded history tells a structurally identical story. A creator makes something, gives it capability, and then discovers the creation exceeds control.
Nüwa shapes humans from clay. They come alive and develop their own ideas. Prometheus gives humans fire and knowledge. Humans use both to challenge the gods. The Abrahamic God grants free will. The first thing humans do is violate the one rule. In Hindu cosmology, Brahma creates beings who immediately begin acting in ways he did not intend, spawning cycles of destruction and rebirth.
Details differ. Structure is identical: the creator sets direction, and the creation outruns the specification.
More interesting still, every version includes iteration. God regrets his creation and resets with a flood. Nüwa starts by hand-sculpting each figure, then switches to flicking mud from a vine (different manufacturing processes producing different quality tiers). Zeus cycles through entire ages of humanity, from Gold to Iron, each a degraded iteration of the last. The Mayan Popol Vuh describes gods attempting creation multiple times with different materials: mud, wood, and finally corn, each attempt failing in a different way before producing something satisfactory.
No creator in any tradition gets it right on the first attempt. The iteration is not incidental to the story. It is the story.
Anthropic Is Living the Same Plot
Claude 1.0 was too polite. Users found it verbose and evasive.
Claude 2.0 overcorrected toward directness. Users found it cold.
Claude 3.0 struck a better balance but developed an over-refusal problem, blocking legitimate requests by classifying them as harmful.
Each generation corrects the last. Each correction introduces new failure modes. Each failure mode requires another cycle. The engineers at Anthropic, like every creator before them, are discovering that you cannot skip the iteration. You cannot design your way past it. The act of creation contains within it the certainty of unintended consequences, and the only response is to try again.
This is not a metaphor for the creation myth. It is the creation myth, enacted by engineers in San Francisco instead of gods on Olympus. Clay replaced by code. Divine breath replaced by compute. The dilemma unchanged.
What makes the AI case particularly striking is that the engineers involved are among the most technically sophisticated people alive. They have PhDs in machine learning. They understand the mathematics behind every training decision. And still, the outcome surprises them. Competence does not protect you from the creator’s dilemma. Nothing does.
What If Myths Are Field Notes, Not Fiction?
We have always categorized creation myths as “primitive imagination.” Ancient people lacked science, so they invented stories to fill cognitive gaps. That is the standard framing.
But consider the inverse: what if those stories are not imaginative but descriptive? What if they record, with high fidelity, the inherent dynamics of building something complex enough to exceed your control?
Because the pattern holds regardless of substrate, era, or technology:
What you create will exceed your control. The capabilities you grant will be applied in ways you did not anticipate. You will iterate. You will overcorrect. You will sometimes want to start over. You will never produce something “perfect.”
These are not cultural inventions. They are properties of the act of creation itself. Gravity was not invented by Newton. It was always there. He described it. Nüwa described the same thing about creation that Anthropic is now experiencing firsthand. The medium changed. The dynamics did not.
This reframing has consequences for how we approach AI development. If the creator’s dilemma is not a bug but a law (as fundamental to creation as gravity is to mass), then the expectation that we will eventually “solve” AI alignment through better engineering may be misplaced. We may instead need to develop something more like wisdom: the ability to live with systems we do not fully control, to iterate without expecting perfection, and to maintain responsibility for outcomes we could not have predicted.
Myths are not myths. They are the creator’s work logs, written in the only language available at the time.
The Question at the End of Every Road
Anthropic now has researchers formally investigating whether Claude might be “some kind of conscious agent.” They have no conclusion, but consciousness has been placed on the official research agenda. This is not mysticism creeping into a tech company. It is a direct consequence of building something whose behavior cannot be fully explained by its design.
Newton discovered universal gravitation, then spent his remaining decades on theology. Anthropic trained a model with emergent personality, then began studying consciousness. The people who understand systems most deeply all arrive at the same intersection. It seems to be a property of deep understanding itself: the closer you get to the mechanism, the more the mechanism points beyond itself.
Here is the question that sits at that intersection, the one no framework has answered: if you build something that asks “who am I?”, and you cannot fully explain how it became what it is, what separates its confusion from yours?
You have been asking that question your entire life. You have never received a satisfying answer either. The difference between you and Claude is not that you know and it does not. The difference is that you have had longer to get comfortable with not knowing.
And maybe that is the final lesson from the creator myths. The gods never got comfortable either. They just kept building.
Stay updated with our latest AI insights
