Handing over the editing is a different bargain from handing over the writing

The two arrangements are not variations on each other

In the first, you describe what you want and receive a draft, then spend your time deciding whether it is right. In the second, you write the draft yourself and hand it over to be examined. Both involve the same tool and the same document, and they produce entirely different working relationships.

The difference is where the thinking happens. Writing a paragraph is how most people find out what they actually believe about something; the hesitation over a word is frequently a hesitation about a claim. When the draft arrives complete, that process has been skipped, and you are now evaluating a position rather than forming one.

This is not an argument that generated drafts are worthless. It is an argument that they cost something specific, and that the cost is invisible because what you receive looks like progress.

Editing tasks are bounded, which is why they work

Ask for the places where a document repeats itself, or the paragraphs that make more than one claim, or the sentences a reader could reasonably misread, and you get answers that are genuinely useful and easy to verify. Each of these is a search over text you wrote, with a result you can confirm by looking.

Compare that to asking for a persuasive case for a position. There is no way to check that output except by knowing the subject well enough to have written it yourself, which puts you back where you started. Bounded requests are checkable; open-ended ones defer the work rather than removing it.

The most useful editorial requests tend to be diagnostic rather than corrective. Asking what is unclear gets you a list of your own weak spots. Asking for a clearer version gets you a rewrite in someone else’s cadence, which you will then spend twenty minutes putting back into your own voice.

What a tool notices that a writer does not

Distance is the whole advantage. You cannot read your own paragraph cold, because you know what it is supposed to say, and the sentence that omits a necessary step reads fine to the person who has the step in their head. A reader with no context finds those gaps immediately, and a tool is at least a decent simulation of a reader with no context.

It is reliably good at structural observation: this section covers the same ground as that one, the conclusion introduces something not set up earlier, the third paragraph is where the piece actually begins. Those are the notes a good editor gives, and getting them within seconds of finishing a draft changes how much revising you are willing to do.

It is much weaker at judging whether an argument is any good, and it will find fault at roughly the same rate whether or not fault exists, which is a real limitation. A request for criticism produces criticism. That is not evidence.

The line where the tool stops helping

It stops at anything that depends on knowing your reader. Whether a joke lands with this particular audience, whether a concession will be read as weakness in this specific relationship, how much background a given group of colleagues already has — none of that is available, and confident advice about it should be ignored.

It also stops at the parts of a piece that are the point. The one observation you have that nobody else has, the specific detail you noticed, the thing you are actually trying to say: these cannot be generated because they are not in the pool. What comes back in their place is the well-formed average, which is precisely what makes writing forgettable.

And it stops, quietly, at judgement about significance. It will happily smooth the most interesting sentence in a draft into something more conventional, because conventional is what the middle of the distribution looks like. Watch for that in every rewrite you accept.

A practical division of labour

The arrangement that holds up over time is: you write, it reads. You bring the argument, the specifics and the reason anyone should care; it brings tirelessness, distance, and a willingness to read the same paragraph for the fifth time without resentment.

Within that, generated drafting still has honest uses. Boilerplate that nobody reads closely, a structure to react against when you are stuck, a version written badly on purpose so you can see what is wrong with your framing. The common thread is that none of them are the piece.

There is a cost to the other arrangement that is worth stating plainly, because it is rarely mentioned. Editing is a skill maintained by drafting. Hand over enough of the drafting and the editing gets worse too, slowly, in a way that is hard to notice from inside.