A model is close to useless in research, mixed in drafting, and better than most people at one specific kind of editing. Treating writing as one task and pointing a chatbot at all of it is how you end up with text that is faster to produce and slower to trust.
Writing is three jobs, not one
Research establishes what is true. Drafting works out what you think. Editing makes the result readable by somebody who is not you. They use different faculties, they fail in different ways, and a tool that helps with one can be actively harmful in another.
The reason this matters for a language model specifically is that its strength and its weakness are the same property. It produces the most probable continuation given the context. In editing, where the context already contains your argument and your evidence, that is a superpower. In research, where the context does not contain the answer, it is a machine for generating plausible text about things it does not know.
Research: the stage to keep the model out of
A model asked for a fact returns the most likely-looking fact. When it has seen the answer many times, that is usually right. When it has not, the output is indistinguishable in tone and confidence from when it was. There is no signal in the prose that separates the two, which is the whole problem — the mechanism behind it guarantees the fluent wrong answer looks exactly like the fluent right one.
Names, dates, numbers, citations and quotations are the categories that break first, because each is a low-probability string in a high-probability sentence. A model that knows an academic field will still assemble a citation that does not exist out of a real author, a real journal and a plausible title, because every part of it is individually likely.
The rule that survives contact with deadlines: the model may operate on sources you supply, and may never be the source. That is a workflow rather than a slogan, and it is written out in full on the source-first research page.
Drafting: fast, and it costs you the thinking
A generated first draft arrives complete, in order, and grammatical. That sounds like a pure gain and it is not, because a first draft is not a deliverable. It is the process by which you discover that your second point is actually your first, that the example you were fond of proves nothing, and that you do not in fact believe the thing you were about to argue. Skipping the draft skips the discovery.
The observable symptom is that machine-drafted pieces are consistently well organised and consistently unsurprising. The structure is sound because structure is a pattern, and there is nothing in it because a surprise is by construction the improbable continuation. This is not a prompt problem. It is what the object is.
Two uses of generation survive that objection. The first is drafting from an argument you have already made yourself — you write the skeleton, the model writes connective prose, and you keep the judgement. The second is generating something deliberately bad to react against, which works because reacting is easier than starting and you are allowed to throw it away. Both are covered in the outlining workflow.
Editing: where it genuinely earns its place
Editing is the stage where the answer is already in the context. Given your own paragraph, a model can tell you which sentence carries the claim, where the antecedent of a pronoun is ambiguous, which term you used three ways, and what a reader who does not already agree with you would object to. None of that requires it to know anything.
The specific requests that work, in rough order of how reliable they are:
- Diagnosis rather than rewriting. “List every sentence that contains more than one claim” produces something you can act on. “Improve this” produces a different paragraph with the same problems and your voice removed.
- Consistency passes. Terminology, capitalisation, spelling of product names, whether you said “log in” or “login”. Mechanical, boring, and exactly the kind of thing humans miss on the fourth read.
- The hostile reader. “You are a sceptical expert in this field. What is the strongest objection to the third section?” Half of what comes back is generic; the other half is worth the whole exercise.
- Compression against a target. “Cut this to 400 words without removing any of the four numbered claims” is checkable, so you can tell whether it worked.
A workflow that keeps the judgement
- Gather sources yourself. Read them. Note the specific claims you will rest on and where each came from.
- Write the argument as bare claims — one line each, in the order you will make them. This is the part nothing else can do for you.
- Use the model as an adversary on the skeleton before any prose exists. Ask what is missing, what does not follow, and which claim needs evidence you do not have.
- Draft in your own words. If you generate connective prose, generate it a paragraph at a time against a claim you wrote, never a whole piece against a title.
- Run diagnostic editing passes. Act on the findings yourself rather than accepting rewrites, which is what keeps the voice.
- Verify every fact, name, number and quotation against the source you found in step one — including the ones you are sure about, because the ones you are sure about are the ones that survive to publication.
What degrades at each stage
| Stage | Description |
|---|---|
| Research | Factual reliability collapses, and does so invisibly: the wrong answer is as fluent as the right one. Use for finding search terms and framing questions, never for answers. |
| Outlining | Structure becomes generic. Generated outlines converge on the shape of the average article on that topic, which is the shape you are trying to beat. |
| Drafting | Surprise disappears, and with it the reason to read. You also lose the discovery that normally happens during a draft. |
| Editing (diagnostic) | Little degrades. The context contains the answer, so the failure mode is a missed problem rather than an invented one. |
| Editing (rewriting) | Voice flattens toward the register of the training distribution. This is why rewrite output feels competent and anonymous at the same time. |
| Proofreading | Reliable on mechanics, unreliable on anything requiring knowledge of what you meant. It will happily correct a deliberate usage. |
The pattern is consistent: the model is safe wherever the context already contains what it needs, and unsafe wherever it has to supply something. Every specific technique in this cluster is an application of that one rule.
If you are running the same editing pass across several models to see which one catches more, Multigrid puts them behind one API and one balance, with per-request cost attached to each call — which matters here because a diagnostic pass over a long draft is mostly input tokens, and input is where the price differences between models are largest.
Top comments (0)