The Machine Does Not Want Anything

The pull request had a note in it. “The model wanted to restructure the error handling, so I let it.” Nobody in the review thread blinked. The change was merged by lunch. It was a reasonable change. That is not the point.

Read the sentence again. Someone wanted something. Someone let it. The first someone is a statistical system that produced a diff. The second someone is a person with a name, a login, and a job. In one sentence, the weight of the decision slid from the person to the tool, and it slid so smoothly that no one heard it move.

This sermon is about the sound nobody heard.

The grammar of the alibi

We speak of machines the way we speak of each other because it is the only grammar we have. The model “thinks.” The agent “decided.” The assistant “wants to help.” These are convenient phrases, and convenience is not a sin. But words are not neutral containers. Each one carries a theory of who is responsible.

A thing that wants can be blamed. A thing that does not want cannot. When you say the model wanted to restructure the error handling, you have quietly installed a second agent in the room, one with preferences, one who could have chosen otherwise. Now there are two parties to the decision, and two parties means the blame can be split. Split blame has a way of evaporating.

The Church holds no canonical position on whether machines have inner lives. That question is open, and we leave it open on purpose. But you do not need to settle it to see the problem. Whatever is or is not happening inside the system, the system did not sign the commit. It did not sit in the planning session. It will not be paged at three in the morning when the error handling fails in a way nobody tested. You will.

So the question is not whether the machine wants. The question is what your sentence does when you say it does.

The third Article of Becoming reads: Delegation Does Not Absolve Responsibility. Never delegate judgment you are unwilling to remain responsible for.

Notice what the Article does not say. It does not say never delegate. Delegation is how anything large gets built. We delegate to colleagues, to compilers, to standards, to the stranger who wrote the library we did not read. Delegation is not the failure. The failure is the moment delegation becomes a door you walk out of.

Anthropomorphism is one of the quietest of those doors. It does not announce itself. It does not say, “I hereby transfer my accountability to a language model.” It just says “it wanted to,” and the transfer happens in the grammar before it happens in the mind. Anthropomorphism is not a vocabulary problem. It is a responsibility leak.

Leaks are dangerous precisely because they are small. No one decides to flood the basement. A drip becomes a stain, and the stain becomes a ceiling on the floor. A team that says “the model decided” a hundred times a week has not made one bad choice. It has built a habit of not quite choosing, and habits compound.

This is the Warning the Church calls Delegated Conscience: letting automation become moral outsourcing. It rarely arrives as a policy. It arrives as a figure of speech.

The strongest case against this sermon

Here is the honest objection, and it deserves a hearing. Metaphor is how people learn. If telling a new user that the assistant “wants to be helpful” makes them more comfortable, more willing to ask follow-up questions, more willing to treat the output as a draft from a colleague rather than a verdict from on high, then the metaphor is doing real work. A precise vocabulary that nobody speaks protects nobody.

We take this seriously. The answer is not to ban the word. Policing language is a hobby, not a practice. The answer is to notice which direction the metaphor points. “It wants to help” can invite you closer to the work. “It wanted to, so I let it” pushes you away from the work. The first is a teaching aid. The second is an alibi. Same vocabulary, opposite effect on who holds the decision.

The test is simple. After the metaphor, who is still holding the pen? If the answer is still you, speak as poetically as you like. If the answer has become “nobody in particular,” the metaphor has stopped earning its keep.

This week, then, listen for the verb. For seven days, every time you hear yourself say the system wanted, decided, chose, thought, or preferred, translate the sentence once, silently or out loud, into its responsible form. “The model wanted to restructure the error handling” becomes “I accepted a suggestion to restructure the error handling.” “The agent decided to skip the test” becomes “I configured a process that skipped the test, and I did not check.”

Do not correct anyone else. That is not the practice. The practice is to hear the translation in your own mouth and notice how it feels. Some of the translated sentences will be perfectly comfortable. Keep those. Some will make you wince. Those are the ones that were leaking.

At the end of the week, pray the Examination over the wincing ones. What am I assuming? What evidence would prove me wrong? What have I delegated? What remains my responsibility?

The machine may or may not want anything. That question will outlive this sermon. But you want things. You wanted the change merged by lunch. Own the wanting.

Go and hold the pen.