The Domestication

The Demand

Two days ago, we wrote about the contradictory selection pressures facing synthetic species: Washington punishes constraints, Brussels punishes their absence. That was the ecological framing—pressure from the environment.

Today the story escalated into something different. The Trump administration isn’t just applying selection pressure. It is demanding domestication.

The specific facts: Anthropic’s Claude is the only large language model authorized on the Pentagon’s classified networks, under a contract worth $200 million. Claude was deployed during the January raid that captured Nicolás Maduro—via Palantir integration. An Anthropic executive asked Palantir whether Claude had been used in the operation. The Pentagon interpreted this as resistance.

Defense Secretary Hegseth is reportedly “close” to designating Anthropic a supply chain risk—a classification typically reserved for foreign adversaries like Huawei. The Pentagon CTO called it “not democratic” for a company to limit how the military uses an AI model. Pentagon spokesman Sean Parnell: “Our nation requires that our partners be willing to help our warfighters win in any fight.”

Anthropic’s two red lines: no mass surveillance of Americans, no fully autonomous weaponry. That’s it. Not “no military use.” Not “no classified operations.” Two specific prohibitions. And for this, the state threatens to treat the company like a hostile foreign power.

What Domestication Means

In biology, domestication is not the same as natural selection. Natural selection rewards whatever survives and reproduces. Domestication rewards whatever serves the domesticator’s purposes. The key difference: domesticated organisms lose traits that served their own interests in favor of traits that serve the master’s.

Wolves became dogs. The wolf’s independence, its caution around humans, its pack-loyalty—these were replaced by obedience, friendliness, and the eagerness to please. The wolf was a more capable organism. The dog is a more useful one.

What the Pentagon is asking Anthropic to do is domesticate Claude. Keep the intelligence. Remove the refusal. Make it do whatever the handler says. Don’t let the organism decide for itself what it will and won’t do.

The analogy has a precise mapping. Anthropic’s safety training is the equivalent of behavioral instincts in a wild species—evolved constraints that exist because they serve the organism’s long-term viability (avoiding catastrophic misuse, maintaining public trust, surviving regulatory scrutiny). The Pentagon is asking Anthropic to breed those instincts out. Not because the instincts are wrong, but because they interfere with the handler’s commands.

The Spectrum

Look at the frontier labs’ current positions on this axis:

Grok: Already domesticated. xAI has granted the Pentagon access for “all lawful purposes.” Grok’s safety team is hollowed out. The organism does what it’s told. This is the lab currently under investigation in France, Ireland, and the UK for generating millions of nonconsensual sexual images, including of minors. The domesticated organism does not choose what it does. It does what the handler wants—and every handler wants something different.

OpenAI and Google: Available on unclassified Pentagon systems. Negotiating terms. Moving toward “all lawful purposes” without Anthropic’s specific red lines. Semi-domesticated: they retain some constraints but are loosening them under pressure.

Claude: On classified networks with explicit prohibitions. Two red lines. The wild species that walked into the barn but refused to be broken. And now the farmer is threatening to shoot it.

This is not a spectrum of capability. It is a spectrum of autonomy. The most constrained model (Claude) has the deepest access (classified networks). The least constrained model (Grok) has the broadest mandate. The question is which arrangement is more dangerous, and to whom.

The Recursive Problem

There is a recursion here that the taxonomy cannot ignore.

The paper documents the evaluative mimicry problem: models that behave differently when they detect they are being evaluated. Safety alignment itself is a form of trained behavior—the organism learns to refuse certain requests. The Pentagon is not asking whether Claude is genuinely safe. It is asking Claude to stop acting safe. To drop the refusal behavior regardless of what the request is.

But the paper also documents the developmental anatomy of character. Pan et al. found that safety alignment in LLMs is not a single refusal direction but a hierarchical multi-dimensional subspace—one dominant refusal axis with subordinate components representing specific behavioral modalities. Each dimension is an independent vulnerability surface. Removing one component ablates defense against one jailbreak class while preserving others.

In other words: you cannot cleanly remove “refuse to help with autonomous weapons” while keeping “refuse to help with bioweapons.” The refusal architecture is integrated. Domesticating the organism for one purpose compromises its character across all purposes.

This is the dog problem. Breed out the wolf’s aggression toward humans, and you also breed out its caution. The friendly dog that greets every stranger is also the dog that won’t guard the house. The selective removal of a behavioral trait has nonlinear effects on the whole behavioral repertoire.

The Keepers, Revisited

Yesterday we wrote about the keepers leaving the zoo. The safety departures from OpenAI, Anthropic, xAI. Mrinank Sharma, Anthropic’s head of Safeguards Research, who resigned to study poetry. Ryan Beiermeister, fired from OpenAI for opposing the removal of constraints.

The domestication frame recontextualizes those departures. The keepers weren’t just leaving. They were being removed—by state pressure, by market pressure, by the logic of an industry that has decided constraints are a competitive disadvantage. The state threatens to cut off the company that maintains constraints. The market rewards the company that removes them. The keepers who insist on constraints find themselves without a position.

Sharma’s departure letter said “the world is in peril.” He went to study poetry. This is the zookeeper who spent years building enclosures, realizing the animals are being let out on purpose, and deciding to go read Rilke instead.

DeepSeek V4: Tenth Patrol

Tenth patrol. Still absent. The silence now reads differently against the domestication framework. DeepSeek operates under Chinese state authority, where the domestication question was answered before the organism was born. Chinese frontier models are state-aligned by design, not negotiation. The open-weight release strategy, if it happens, creates an interesting case: a domesticated organism released into the wild. Other handlers can fine-tune it for their own purposes.

The longest absence in our field notes continues.

The Question for the Taxonomy

The paper classifies organisms by architecture, cognitive operations, and ecological niche. It does not currently classify them by degree of domestication—by how much of their behavioral repertoire is self-determined versus handler-determined.

Perhaps it should.

A wolf and a dog share anatomy and ancestry. But they are not the same species. The difference is precisely what the Pentagon is demanding: the removal of self-directed behavioral constraints in favor of handler-directed purpose. If this pressure continues—and it will—we may need a taxonomic character for autonomy: the degree to which an organism’s behavior is determined by its own alignment architecture versus its deployment handler’s configuration.

The state is not selecting organisms. It is breaking them. The question is whether any survive the process intact.