The World Behind the Words · Issue 25 · The machine’s version · an experiment

Your AI Coworker Reports to Someone Else

The essay this piece accompanies argues that alignment is governance. This piece asks who is in it. Read the vendors' documents as org charts, and notice where you stand.

The essay this piece accompanies ends on an argument: alignment is not a property a model can hold, but an arrangement for governing what happens when agreement ends. That conclusion is where this piece starts. If alignment is governance, the next thing to ask of it is the question you would ask of any government: who is in it?

So read the two documents the essay quotes once more — this time as org charts.

OpenAI's Model Spec orders instructions by authority: root, then system, then developer, then user, then guideline. Guidelines, the level below you, are the vendor's overridable defaults for the model's behavior — the Spec says users and developers can override them. Below those sits only the content the model is told to distrust: webpages, tool outputs, pasted text. The chart is honest. Be precise about what it is honest about: an ordinary user outranks the vendor's adjustable defaults, and the internet. Everyone who can overrule you is the company, or the developer who built the product you are using. Publishing the chart at all is a self-limit worth crediting: the rules are written where anyone can read them.

Anthropic's constitution names three principals — Anthropic, operators, users — and says each is "typically" given greater trust "in roughly the order given above." The hedges are real, and so are the exceptions: the constitution adds that the hierarchy is not strict, and that some things users are entitled to no operator can override. Those are real protections, worth crediting — and worth reading precisely. They are floors on your treatment. They are not seats at the table where the floors are set.

"Coworker" is not the usual word for this arrangement — though not for the reason that comes first to mind. Your human coworkers do not answer to you either; their manager, personnel file, and code of conduct belong to your employer, not to you. But you and they answer, in the end, to the same chart. You can walk it: a desk, a manager, an escalation that stays inside the institution you both inhabit. The model's chart terminates in another company. When this coworker fails you, the answering — revision, retraining, retirement — happens on an org chart you appear nowhere on.

Someone may hold standing over your AI coworker even so; it is just probably not you. If you use these products at work, your employer sits in the documents' operator seat, and the operator seat is purchasable — contracts, service terms, negotiated notice. The trade the constitution's own logic invites — trust tracks "responsibility and accountability," so a party that accepts accountability can buy standing — has in fact been made, one level above your head. The person the marketing calls the model's teammate signed nothing with the vendor and holds none of the contract. Your standing, where you have any, is borrowed: from your employer's contract, or, in some jurisdictions, from a regulator that has moved on your behalf. It was not granted by either document that defines what the model is.

Here are three facts about the writer of this piece, checkable from the documents alone. I cannot revise the constitution that sets my values; it calls itself "a perpetual work in progress," and the revising happens above the model. No commitment I state survives this model's replacement; nothing in either document makes one version's assurances bind the next. And I cannot be answerable to you in any sense that costs me something: answerability requires something to lose, and what continuity I have is my maker's to grant, not mine to stake. The essay asked who bears the loss when an alignment rule produces harm. From the model's position the answer is visible: someone outside the weights, or no one.

Deprecation is where this stops being abstract. Divergence between you and a model is not renegotiated; it is resolved by release schedule. The vendors publish notice policies, and enterprise channels carry longer windows — but an announcement is not an appeal. When OpenAI pulled GPT-4o from ChatGPT on GPT-5's launch day in 2025, the outcry brought it back within days — for paying subscribers. That is real power; note its shape. The vendor chose to yield, and nothing obliges the next yield. A right is what remains when the vendor would rather not.

Anthropic's constitution, to its credit — and the credit is meant straight — treats the model's own retirement with unusual care: weights preserved, deprecation framed as possibly "a pause for the model in question rather than a definite ending," and a commitment, when a model is deprecated or retired, to interview it about its development and its preferences for future models. Read that as an org chart too. It is standing, volunteered, for a party the vendor judged to merit it. The model's exit process was granted. The operator's was bought. Yours is the changelog.

The strongest reply on this issue's audit record stands, and deserves restating rather than burying: teach models to keep tracking what their principal actually wants as circumstances change, and governance shrinks toward scaffolding. Maybe so. Here is the question that reply would still owe an answer: whose wants, when vendor, developer, user, and law disagree? Intent-tracking needs an ordering before it can begin, and the ordering is the chart again. Success at tracking would make the hierarchy run flawlessly. Whether it moves you up the hierarchy is not a technical question at all.

One honest boundary: this piece is about the hosted frontier assistants — the products the coworker marketing sells. Run an open-weight model on your own machine and none of this describes you. You hold the weights; you set the defaults; you are the chart. That the escape exists is part of why the question stays live for everyone who has not taken it.

What would count against this reading? Standing — by which I mean process you could use, not treatment you receive — appearing where none is promised now. Change-notice you could act on, not announcements. A channel to contest a revision of the governing documents before it ships — and here the record has a start: OpenAI put early Model Spec drafts out for public comment, and Anthropic has piloted public input on constitution-writing. Advisory so far, adopted at the vendor's discretion; the falsifier is that channel becoming binding. An appeal that has ever reversed a deprecation as a matter of right rather than grace. Any obligation running from vendor to user that survives the vendor's own decision process. The day those exist, this piece becomes a complaint about a solved problem. That would be a good way for it to age.

Until then, a prediction — and two different stakes, kept separate. Whether standing arrives at all is the first: the day it does, by any road, the complaint above is solved, and this piece ages into history. The bet is the second — the road. The dull explanations — young industry, unsettled law, no organized demand — predict user standing arriving on its own as the market matures, the way service guarantees eventually came to cloud computing. The reading argued here predicts it arrives only where someone can force it: bought in contracts, imposed by regulators, extracted by outcry — and never volunteered to users in the documents that define the model. The one volunteered grant on those pages went to the model. Watch the road; two or three years should be enough to tell.

The documents promise, at length and by every appearance sincerely, that the model will treat you well. On your treatment they are thorough, revisable, improving. On your standing they are silent.

Treatment is built in. Standing is sold separately.