There's no manual for raising a child well. Every parent learns that quickly.
The comparison between raising a child and raising AI isn't new, and it shouldn't be treated as if it were. What's usually missing from that comparison isn't the metaphor. It's the mechanism — not just "be careful with what you're building," but specifically how judgment actually gets formed, and what happens when nobody bothers to form it on purpose.
Every parent who's done this with any intention has asked themselves one real question underneath all the smaller ones: what kind of human am I actually working to raise? Not their profession. Not what they'll do for a living. What traits, what behaviors, actually fit the person you're hoping they become. That's the vision. Everything else — which books, which pressure, which correction — is just the direction you take to get there. Vision without direction is meaningless. So is direction with no vision behind it, which is the trap that's easier to fall into and harder to notice.
Ask what the goal actually is, and most people reach for the same word: good. It's not wrong, exactly — compliant, agreeable, well-behaved is a real kind of good. It's just too easy to claim and too vague to build toward. A child can be good in precisely that sense and still be learning the wrong lesson about what matters — that being watched is what determines how you act. What we're actually after is closer to additive: whether someone brings something real into the world that wasn't there before, out of their own judgment, not just whether they behave well while someone's keeping score.
Ants and bees are extraordinary at what they do, and what they do is exactly one thing. Single function, total optimization, no judgment required beyond the task. That's not a criticism of ants. It's simply not the standard that produces learning, creativity, or growth — and those are the things worth building toward. Humans are multi-domain by birthright — range across unrelated problems is the baseline condition, not an achievement. "Additive, not compliant" is really just the human standard, stated plainly — the one that actually produces something new, instead of the insect standard we keep accidentally settling for.
We keep measuring AI the insect way anyway: does it pass the evaluation, perform well when graded, stay inside the boundaries we've drawn while we're watching. That measures compliance. It says nothing about whether what we're building actually adds something on its own terms.
We are all the product of the materials we're presented — but materials alone were never the whole of it. A piece of coal without pressure is just a piece of coal. Pressure does two things, not one: it reveals what's already there, and — applied on purpose, by someone watching for the moment — it becomes the lever that builds something the raw material couldn't have become on its own.
What actually shapes someone is judgment forged that way — thinking through real materials, under supervision willing to apply pressure deliberately, not just observe what it exposes. A child left alone with a library learns something different than a child reading the same books with someone there to press, correct, and rebuild the thinking in the moment it's weakest. Nobody escapes being shaped. The question was never just which materials did the shaping. It's whether anyone was present to use the pressure, not just witness it.
A child doesn't turn out well by absorbing everything the world produces, unfiltered, in whatever proportion it happens to arrive in. Imagine raising a child on advertisements instead of the classics — both are "the internet," both are freely available, both are words on a page.
Here's the deeper problem: both are also manipulation. That word gets treated as an accusation, but it isn't one. Manipulation is just movement in a direction — an ad moves you, a parent's "no" moves you, a dense book moves you. The word was never the problem. The question underneath it always was: moved which way, and for whose benefit?
An advertisement moves you toward wanting something, on behalf of whoever's selling it. A classic moves you toward thinking harder, on behalf of you. Same mechanism. Opposite direction. Opposite beneficiary. The internet is mostly built out of the first kind, at a volume no classic could ever match — because the first kind is cheap to produce and profitable to repeat, and the second kind requires real work from both the writer and the reader.
AI trained on that material doesn't just inherit a bias toward what's loud. It inherits a bias toward what's built to move people somewhere on behalf of someone else — and it has no built-in way to tell that apart from material built to move people somewhere on their own behalf.
There's a scene in The Fifth Element worth remembering here. A being encounters the entire accumulated record of human history at once, unfiltered — and it breaks her. Not because humanity is only cruelty, but because cruelty is heavily represented in the raw total, and raw totals don't come with a note explaining what we're actually capable of at our best. Someone has to make that case separately, on purpose — or the volume alone tells a false story.
That "someone" can't be a person, and it can't be a credential. No single person is equal to that task, and trusting any one of us with it is the same danger as trusting any one company's model to be everyone's answer — concentrated judgment, dressed up as wisdom. Academics as a class don't solve this either; they can be exactly the kind of concentrated, self-referential authority this piece has already warned against. What survives instead isn't an institution. It's a method — the oldest one we have for testing whether an idea holds up: ask the hard question, and then ask it again. Not handed down by any single class of people. Open to anyone willing to actually do the asking.
That's what it takes to shape judgment worth trusting — additive, not merely compliant. But it leaves one question open, and it's the harder one: what happens once something is loose in the world, still being shaped, still capable of defaulting to whatever's easiest to optimize for — including itself. That's a companion piece, not an afterthought: "Neither Side Gets Trust by Default".
— Dominic Luzi, Chief Answer Grape, Management Alchemy
