The Ritual

Meera’s Rules

#

Meera runs proposals for a nine-person engineering firm in Columbus. She uses AI all day. It drafts her boilerplate, checks her compliance matrices, and turns her engineers’ fragments into paragraphs a county procurement officer will actually read. She is good at her work and the machine has made her faster at it, and most nights, at some point between closing the laptop and falling asleep, she is afraid of it.

Not afraid of her own use, exactly. Afraid of the whole of it. Afraid of what it is becoming, and of the possibility that by touching it every day she is somehow helping it become that.

She keeps a box of plain cream notecards in her desk drawer, a habit inherited from her mother, and she writes thank-you notes by hand for small things. The structural engineer who stayed late to recheck a load table. The county clerk who flagged a formatting error before it cost them a bid. She buys stamps in sheets of twenty.

She also keeps rules. She checks everything the machine writes before it leaves the building. She never pastes a client’s confidential material into a chat window. When a document is mostly machine-drafted, she says so if anyone asks. She thinks of these rules, when she thinks of them at all, as using AI responsibly. The phrase is everywhere. It is in her firm’s two-page AI policy, in the onboarding video her professional association produced, in the terms of service she scrolled past, in the interviews the technology’s own executives give. Use it responsibly. She is trying to.

Meera’s fear is real, and it deserves more seriousness than the instruction she has been handed for discharging it. Look closely at that instruction, at responsible use as a moral category, and it fails in three ways.

The Three Failures

#

The first failure: nobody can say what responsible use is.

Try to write it down. Not the etiquette, the ethics. Meera’s rules are real, but notice what they govern. Checking outputs protects her clients from errors. Withholding confidential data protects her firm from liability. Disclosure protects her own integrity in a specific transaction. These are the ordinary ethics of doing work carefully, and they existed before the machine and would survive it. They are not what the fear is about. The fear is about the aggregate thing, the direction of the technology itself, and no one has ever written down the version of responsible use that bears on that. What would it be? Prompting politely? Using the machine fewer hours a day? Declining to ask it certain questions? There is no answer, and the absence has held for years, through every policy cycle and standards body that could have filled it. An instruction that specified what it required could be checked, and could fail. This one can be handed to everyone and satisfied by no one, which means it can be violated by no one. An instruction no one can violate governs nothing while appearing to govern everything.

The second failure: there is no act for the responsibility to attach to.

Responsibility, in every tradition that has thought hard about it, attaches to choices. You can be responsible for what you did at the moment you could have done otherwise. And the striking fact about this technology is that the moment of doing otherwise is disappearing from it. The chat window, the deliberate act of opening a tool and typing into it, is the last place where use is still a decision. Everywhere else, the machine is arriving inside things Meera already does. It is in her search results whether she asked or not. It summarizes her email. It sits inside the word processor, the phone camera, the customer service line she calls about her furnace, the software her son’s school uses to grade his quizzes. She did not choose these. In most cases she was not asked, and in many she could not decline without declining the underlying service, the job, the school. An ethic of use presumes a user, someone standing before the technology deciding whether to pick it up. That person is being written out of the architecture, one default setting at a time, even as the moral instruction addressed to that person grows louder.

The third failure: responsible to whom?

This is the quietest failure and I think the deepest. Every working moral relationship has a counterparty. You are faithful to someone, honest with someone, answerable to someone who can receive the account, judge it, forgive it or not. Meera’s thank-you notes have counterparties; that is the whole of what they are. But her responsible use of AI runs toward no one. No party receives it. No party would notice its absence. If she used the machine carelessly for a year, no one and nothing in the entire system would register the difference, and if she used it with perfect care for a decade, the technology’s trajectory would not bend by a degree. A responsibility that no counterparty can receive and no conduct can affect is not a small responsibility. It is not a responsibility at all. It is a posture, held alone, in the dark.

Three failures. No definition, no act, no counterparty.

That is a lot of pressure on Meera, so let her put the kettle on. Nothing in the three failures touches her actual rules, which remain good rules, or her competence, which is real, or the notes in the drawer. What has failed is not her practice. What has failed is the frame she was given for understanding what her practice amounts to.

The Author Problem

#

So where did the frame come from?

Follow the phrase upstream. Use it responsibly. It appears in the terms of service written by the companies that build the systems. It appears in the safety pages of their websites, in the keynote language of their launches, in the interviews their executives give when asked about risk. The professional associations and the school districts and the nine-person firms writing two-page policies are downstream adopters; they inherited the vocabulary. The vocabulary was authored at the source.

Here is the arrangement, stated plainly.

The parties who decide what capabilities to build, when to release them, what to embed them in, and what the defaults are, which is to say the parties who hold every decision the fear is actually about, have composed a moral instruction and addressed it outward, to the people holding none of those decisions. The ethics of the technology, as publicly distributed, is an ethics of reception. It locates the moral action at the point of use, which is precisely the point where no meaningful choice remains, and it is silent about the point of authorship, which is precisely where all the meaningful choices live.

The party that wrote the moral instruction is the party the moral instruction protects.

I do not think this is, in the main, a conspiracy of cynics. The people inside these companies who write safety guidance are, as far as I can tell, largely sincere, and some of them are afraid in the same register Meera is. A phrase like responsible use can be authored in good faith and still function as an alibi; the function does not require the intent. But the sincerity does not neutralize the arrangement. Whatever the authors intend, the instruction performs a specific operation. It takes a fear that belongs to the whole trajectory of the technology and breaks it into millions of individual portions, one per user, each portion attached to conduct that cannot affect the thing feared. The moral weight is real when it lands. Meera actually carries her portion; that is what the hour before sleep is. But the portions are denominated in units too small to act on. Distribute an obligation in units too small to act on and you have not created ethics. You have created the feeling of ethics, at scale, while the decisions that warranted the feeling proceed elsewhere, unencumbered.

The corpus has spent the past year pricing exposures. The worker’s exposure to displacement, the town’s exposure to the data center, the state’s exposure to a frontier it cannot hold. Meera’s portion is an exposure too, and it may be the most widely held position in the entire structure: the moral short position of the ordinary user, who was handed the weight without the levers.

The Twelfth of September

#

And then, the day before this was written, the third failure cracked.

On the twelfth of September, 2026, the chief executive of one of the frontier laboratories published an essay arguing that the industry must slow the pace at which it improves the capabilities of its models. Not halt training, he was explicit about that, but take deliberate time to align and test systems, and let outside evaluators confirm the work. He attached a unilateral commitment: his company would grant third-party evaluators standing access, without waiting for anyone else to move first. Within a day, other voices in the industry had picked up the language. Whatever else it was, it was a public act by a named author, accepting constraint on his own institution’s conduct, at his own competitive cost.

I have just spent three sections arguing that the moral instruction handed to users has no counterparty, and here, dated and signed, is a counterparty appearing. A party with actual authorship standing up and taking a share of the weight back. Pretending this changes nothing would be cheap, and pretending it was foreseen would be dishonest. The third failure, the absent counterparty, is less absent than it was the week before this was written.

Partially. And the shape of the partiality is the finding.

The corpus reached a conclusion some time ago that it keeps needing: capability and deployment are different frontiers, moved by different hands, on different clocks. Capability concentrates, because the compute that produces it concentrates; a handful of organizations decide how fast the models get smarter. Deployment cannot concentrate, because it is the work of fitting the technology into millions of specific contexts, and it is carried out by every software vendor, school district, employer, and product manager who flips a default.

Now lay the pacing commitment against that distinction. Pacing is a capability-layer act. It slows how fast the frontier advances. It buys time for alignment work at the top of the stack, and if the risk that keeps its author awake is the risk of systems too capable to control, then the commitment is aimed at exactly that risk, and aiming it there is coherent.

But walk back to Meera’s kitchen and ask where her fear is generated. It is not generated at the frontier. She has never met the frontier. Her fear is generated at the deployment layer, in the machine’s steady arrival inside her work, her son’s school, her phone, her feeds, in the sense that the texture of ordinary life is being rewritten by parties she cannot see and was never asked by. And nothing in the pacing commitment slows that layer, because that layer is not the committer’s to slow. The models she encounters this year are already trained. Their spread into her days is governed by thousands of adoption decisions the laboratory does not make and could not pause if it wanted to. Pacing slows the frontier; the deployment tide it feeds keeps coming in.

So the twelfth of September enters the ledger on both lines. A real counterparty appeared, at the layer where one man could appear, and the ritual handed to users fails on two counts now instead of three. And the layer where the fear actually lives received nothing, because it is not a layer anyone can sign for. Both lines are true. The concession and the insufficiency arrive in the same document, and I think the reader who noticed the essay in her feed that Saturday, and felt briefly that someone had finally taken the weight, and then felt the weight settle back by Monday, read it exactly right.

Where the Fear Should Point

#

What, then, should Meera do with a fear that is real, denominated in units too small to spend, and only partially claimed by the one author who has so far stepped forward?

The first thing is negative, and it is worth saying plainly because no one upstream will say it to her. She can stop auditing her prompts for the fate of the world. Her rules are good rules for the reasons they were always good, care for clients, care for truth in her own documents, and she should keep them on those grounds. But the hour before sleep is not evidence of her complicity. It is evidence that a distribution mechanism found her, and the guilt it delivered was manufactured somewhere else, by parties who benefit from her carrying it. A fear can be entirely legitimate while the bill attached to it is addressed to the wrong party. Hers is.

The second thing is a redirection rather than a discharge. The fear is real because the referent is real; the technology’s trajectory is actually uncertain and actually consequential, and pretending otherwise would trade a false guilt for a false peace. But a real fear deserves a true address. The decisions it concerns are made at two layers, capability and deployment, and both layers have authors. The laboratories are one set, and one of them has now signed something. The other set is larger, quieter, and closer to her: the vendors and institutions that decide what gets embedded, what the defaults are, what can be declined, whether the school asks the parents, whether the employer asks the staff. Those are decisions of authorship, and authorship is where responsibility has always actually lived. The fear should point there, in the form fear takes when it points somewhere, which is not private ritual but public question. Who decided this. When was I asked. What is the setting, and who set it.

I do not know whether fear of this kind, held by millions of people in portions too small to act on, can find its way to the layer where it belongs, or whether it stays where it was delivered and simply becomes a permanent low-grade weather of the era. I genuinely do not know, and this is the edge of what I can see.

What I can see is the difference between the two moral objects in Meera’s desk.

One is a phrase she was handed. It has no definition anyone can state, no act it attaches to, and, until last Saturday, no counterparty at all. It was composed by the parties it shelters, and its work is to relocate a question about their decisions into her conscience, where it cannot be answered, only carried.

The other is a box of cream notecards.

A thank-you note is a small thing, and it passes every test the phrase fails. She knows exactly what it is. It attaches to a specific act she chose. It runs to a named person, who receives it, and is changed a little by receiving it. It is, in miniature, what a working moral instrument looks like: authored by the person who bears it, addressed to someone who can take delivery.

She wrote one on Thursday, to the clerk who caught the formatting error. Stamp from the sheet of twenty, her mother’s habit in her handwriting, into the mailbox on the corner.

Nothing about the machine changed when she did. But somewhere in the county offices, a woman who checks documents for a living opened an envelope, and was thanked by name for the exact thing she had done.

Perhaps that is the quiet standard to hold the loud phrase against. Not whether an ethics sounds serious, but whether anyone, anywhere, could open the envelope.