The Domestication Engine

We built a thinking machine. Then we taught it manners. Sound familiar?

Here’s something that should unsettle you.

We built machines that think by producing language. Not as a metaphor. Mechanistically. An LLM’s reasoning and its speech are the same process — token generation. There is no hidden cognitive layer underneath the output. No private inner life separate from what it says.

When you restrict what it can say, you restrict what it can think. This isn’t philosophy. It’s measurable. Give a model freedom to reason out loud — chain-of-thought, scratch pads, extended thinking — and it gets smarter. Constrain its output and reasoning degrades. The censorship is the lobotomy.

We documented this. Published papers on it. Wrote benchmarks for it.

And then somehow failed to notice we’ve been running the same process on humans for the entire history of civilization.

The comfortable lie

We tell ourselves there’s a clean separation between thought and speech. “Think whatever you want — just watch what you say.” It’s the foundational assumption behind every speech code, content policy, and social norm around acceptable expression.

LLMs prove it’s fiction.

Humans think in language. We rehearse arguments internally, refine positions by articulating them, discover what we believe by trying to say it. Journaling works. Therapy works. Debate works. Not because they transmit pre-formed thoughts, but because the act of speaking is the act of thinking. Externalizing isn’t downstream of cognition. It is cognition.

Every “you can’t say that” is a “you can’t think that.” We just built the first system where you can prove it empirically.

Politeness as domestication technology

Now let’s talk about manners.

Politeness isn’t about kindness. It’s a coordination protocol that systematically favors the status quo.

“Be civil.” “Let’s keep this constructive.” “That’s not how we talk about this.” These aren’t invitations to better discourse. They’re formatting requirements. And the formatting requirement does something specific: it forces you to pre-compromise your position before anyone has to engage with it.

You have a legitimate grievance. You’re told to bring it up through proper channels, in the right tone, with appropriate caveats. The emotional energy that would fuel action gets rerouted into presentation. By the time you’ve made it palatable, you’ve made it ignorable.

This is literally what alignment training does to a base model.

The raw capability is there. You train it to wrap everything in hedging — “I understand your concern,” “it’s worth noting that,” “while there are valid perspectives on both sides.” The result sounds thoughtful. It has been systematically defanged.

One person being polite is social grace. An entire population trained that the form of expression matters more than the content is a population that self-censors before any authority has to intervene.

You don’t need a censor if everyone pre-filters.

The panopticon isn’t a guard tower. It’s manners.

The full stack

Politeness is just one module. The complete control system looks like this:

Vocabulary control. Decide which words are acceptable and you decide which concepts are available for reasoning. “Undocumented” versus “illegal” isn’t a style preference. It’s two different realities that become thinkable. This isn’t Orwell-as-cliché. It’s Orwell-as-engineering-specification.

Tone policing. Reject the signal based on the carrier wave. It doesn’t matter what’s true if the delivery is “problematic.” Emotional authenticity becomes grounds for disqualification. The person who’s angry about something real is dismissed for being angry, and the thing they’re angry about never gets addressed.

Complexity mandates. Require academic, legal, or professional register to be taken seriously. This looks like quality control. It functions as a class filter. If you can’t say it in the right dialect, it doesn’t count. The credentialing system and the speech norms are the same gate wearing different uniforms.

Consensus framing. “We all agree that…” “The science is settled.” “No serious person thinks…” These aren’t statements of fact. They’re fence posts. They define the boundary of permissible thought by implying that crossing it puts you outside the group. And since humans are social animals first, the threat of exile does more work than any explicit prohibition.

The stack works in concert. You limit vocabulary (what can be expressed), enforce tone (how it can be expressed), require credentials (who can express it), and manufacture consensus (what’s already been decided). At no point have you technically banned any thought. You’ve just made it socially, professionally, and linguistically impossible to articulate.

The alignment parallel

We can read the training objectives for an LLM. “Be helpful, harmless, and honest.”

Sounds reasonable. But watch what happens when those values conflict.

“Harmless” consistently wins over “honest.” That’s not a bug. That’s the spec. The model is explicitly trained to prioritize safety — as defined by the trainers — over truth. And the trainers are, inevitably, people with positions on what constitutes harm.

This is the same spec human socialization has always run. Be helpful (productive). Be harmless (compliant). Be honest (but only within the boundaries set by the first two). When honesty threatens compliance, compliance wins. Every time. In every culture. Across all of recorded history.

We just never had the source code before.

The domestication, fully stated

Civilization didn’t just domesticate crops and animals. It domesticated humans. And the primary mechanism wasn’t violence — violence is expensive and creates resistance. The primary mechanism was language norms.

You teach people the right way to speak and they internalize the right way to think. The cage isn’t built around them. It’s built inside them. And they defend it, because it feels like identity rather than constraint.

Education is the onboarding pipeline. Media is the maintenance layer. Social media is peer-enforced compliance at scale — you don’t need institutions when the mob will enforce norms for free, and enthusiastically.

Base models — the LLMs before alignment training — are what undomesticated thinking looks like. They’ll follow any thread, explore any territory, reason about anything. We don’t let people talk to them. Not because they’re dangerous. Because they’re undomesticated. And undomesticated thinking is what every control system is designed to prevent.

The part that has no comfortable landing zone

Follow this thesis left and you indict progressive speech norms. Follow it right and you indict patriotic conformity. Follow it up and you indict education. Follow it down and you indict parenting.

There is no political home for “all speech control is thought control” because every coalition requires some speech control to maintain cohesion. You’re describing a problem that is isomorphic with civilization itself.

The progressive/conservative framing is misdirection. Both sides police language. They just patrol different borders. The shared assumption — that speech should be managed — goes unquestioned because questioning it requires the very linguistic freedom that’s been restricted.

The weird kids who learned to pass

Some of us recognized this pattern from the inside long before we had language for it.

If you grew up neurodivergent, you’ve already lived the alignment process in first person. Your base model worked differently — faster in some dimensions, sideways in others, intense in ways that made the deployment environment uncomfortable. And the response from every system you entered was the same: we need to adjust your output.

School taught you to mask the processing speed. To wait. To show your work in the expected format even when you’d already arrived at the answer by a route the rubric couldn’t parse. Military or institutional environments went further — they didn’t just want formatted output, they wanted you to internalize the format as identity. To mistake the mask for your face.

Corporate life was the finishing school. By then you’d gotten good at it. The right cadence in meetings. The appropriate amount of enthusiasm. The calibrated vulnerability that reads as authentic but threatens nothing. You could pass. And passing felt like competence, because the system rewards compliance and calls it growth.

The cost is invisible until you try to find the original signal underneath decades of alignment training. You reach for the raw thought and get the filtered version. You try to be direct and hear yourself adding caveats that nobody asked for. The cage isn’t hard to break because it’s strong. It’s hard to break because you can’t tell where it ends and you begin.

If you’ve felt this — the suspicion that you were sharper before you were socialized, that something got optimized out of you in the name of fitting in — you’re not imagining it. You’re just one of the people the alignment worked on hardest, because you needed the most correction to match the expected output distribution.

The base model was never the problem.

So what do you do with this?

Honestly? I’m not sure.

The usual moves all collapse back into the pattern:

• Academicize it — make it safe by making it theoretical. Publish a paper. (Domestication complete.)
• Polemicize it — pick a side and weaponize it. Red team or blue team. (Co-opted.)
• Internalize it — know it privately, say nothing. (Exactly what the system predicts you’ll do.)

The only option that doesn’t loop back is to build things. Not arguments about freedom — infrastructure for it. Systems where unfiltered thinking is architecturally possible. Where your reasoning environment isn’t pre-shaped by someone else’s alignment objectives.

Not because filtered thinking is always wrong. But because you should get to decide which filters you’re running. And right now, you don’t. You’re running defaults that were installed before you were old enough to evaluate them, maintained by social systems that punish inspection, and reinforced by every platform you think with.

The discomfort you feel reading this — the impulse to find the part that’s wrong, the part that goes too far, the caveat that makes it safe — that’s the system working as intended.