Responsibility drifted into the chat.

The tag that costs the most

Social channels carry meanings we never audit. A fluent reply gets tagged, below awareness, as sincere, as understanding, and as accountable. The first two are merely inaccurate. The third one moves something real.

Between people, accepting a task means accepting exposure. If it goes wrong, someone has a problem. That coupling is so reliable that we stopped noticing it is a separate fact from the words. A model produces the words and none of the coupling. The exposure never transferred, and it is still sitting where it always was.

This is why the profile is called drift rather than delegation. You did not hand responsibility over. You experienced it as handed over, which is enough to stop looking for a catcher.

What would have caught a mistake

The deciding answer is the last question: what was most likely to catch a mistake before it caused a problem. Two answers lead here, and they are different situations.

Nothing specific

The session produced a finished-sounding answer, and the question of what would catch an error was never asked. This is the sharper version, because nothing is scheduled to arrive later either.

Someone noticing after it was used

There is a catcher, but it is downstream: a user, a customer, a colleague, production. Delayed discovery is real, and it is doing a job that should have existed before release. The cost of that arrangement is not the discovery. It is that the discovery is made by the person who was harmed.

Three sessions with no one at the door

A background agent runs overnight and returns a completed report. It reads well. It ships to a client, and the person who would have caught a wrong figure is the client's finance lead, three weeks later.

A support team uses AI-drafted replies. Each draft is skimmed by whoever picks up the ticket. Nobody owns accuracy on refunds specifically, so a wrong policy statement goes out eleven times before a customer quotes it back.

An engineer merges an AI-written change on a Friday. The reasoning in the session was good. There is no test that fails if the reasoning was wrong, and the on-call rotation finds out first.

In all three, the content might have been checked. The gap is that no one is standing at the door.

Two things this is not

It is not a verification problem

You can have excellent evidence and still land here. Checking asks is this right. Ownership asks who finds out if it is wrong, and when. A session can pass the first and fail the second, which is precisely why this profile exists separately from The conversation bought a discount on checking.

It is not solved by "I am responsible"

Personal responsibility is the correct legal and professional answer, and every serious treatment of workplace AI says the same thing: the employee remains responsible for the work product regardless of how it was produced. But responsibility that is not attached to a moment and a mechanism does not catch anything. It only distributes blame afterward.

Write the ownership sentence

The principle: ownership is a position, not an attitude. It has to be occupied by someone or something at a specific point before the output moves.

Concretely, write one sentence before you start the next task of this kind:

Before this reaches [consequence], [person or check] must be able to reject it.

Two things make that sentence work. The owner must exist independently of the conversation, and the owner must have a real ability to say no. A reviewer with no time and no authority is a formality, and a test that nobody runs is a comment.

The strongest version is a check you can name and that runs itself: a failing test, a validation rule, a reconciliation that does not balance. The second strongest is a named human with a defined target to inspect, because a reviewer given "have a look at this" reviews nothing in particular. Give them the decisive claim and the consequence, not the whole artifact.

If the honest answer today is that discovery happens downstream, move one step earlier rather than building a review board. Pick the single output type where late discovery hurts most, and put one gate in front of it. One occupied position beats a policy nobody staffs.

Where a gate is not worth it

Not every output needs an owner. Reversible, private, low-cost work should stay ungated, and adding approvals there trains people to route around them.

Match the gate to how hard the mistake is to undo. If you cannot say who would be harmed and how badly, you are not ready to design the control point, and that question is the actual first step.

If the session also ended with the model acknowledging a failure that never got retested, read The apology closed more than it repaired. next, since an unclosed repair is the most common thing an owner is asked to catch.