Use AI as your editor, not your ghostwriter

Readers can now spot text an AI wrote. It is smooth, confident and similar to everything else they read that week. For a business, that sameness costs attention: a proposal that sounds like every other proposal is easy to skip. The fix is not to stop using AI on your writing. It is to change its job from writer to editor.

Why AI drafts sound alike

A 2024 study in Science Advances tested this directly (Doshi and Hauser). Anil Doshi of University College London and Oliver Hauser of the University of Exeter asked people to write short stories. Some writers could see story ideas from an AI model. Readers rated the AI-helped stories as more creative, better written and more enjoyable. The gain was largest for the writers who scored as least creative on their own.

There was a cost. The AI-helped stories were more similar to each other than the stories people wrote alone. Each writer improved, and the group as a whole became less varied. The authors describe this as a social dilemma.

For a business, the second finding matters more. Your proposal competes with other proposals. If everyone drafts from the same model, the drafts converge. Your own voice is part of what makes a client remember you.

The editor method

A post on sockpuppet.org, the site of Thomas and Erin Ptacek, sets out a method for writing with a model (sockpuppet.org). The core is to write the draft yourself and use the model only to find problems. Its rules, in our words:

• Write the first draft yourself.
• Use no word or phrase the model suggests. It points at problems. You write the fixes.
• Forbid praise. Encouragement makes you keep weak first instincts.
• Edit in passes. Run one narrow request at a time over the draft, and rewrite what each one flags.
• Compare versions in a fresh session. A model that watched you edit tends to favor your edit.
• Ignore advice that would remove your voice, even when the model insists.

You may not use a single word an LLM suggests to you.

The post explains why the second rule matters. Models are good at choosing pleasing phrases, and that is the problem: the phrases read like headlines, and readers learn to spot them. Models are also tireless at finding faults, which is the part you want.

Editing passes to run

Start each pass with the same instruction, then add one request. Emeri wrote these prompts for this guide; adapt them to your work.

1. Passive voice: list every sentence where the actor is hidden.
2. Nominalizations: list nouns built from verbs, such as "implementation" or "utilization", where the verb would be clearer.
3. Filler: list words that add nothing, such as "very", "really", "actually" and "just".
4. Repetition: list words or phrases used more than twice.
5. Order: name two or three paragraphs that would be clearer in another place, and say why.
6. Claims: list every number, date, name and promise, so you can check each one against its source.

Rewrite each flagged sentence yourself. Then open a new session, paste the old and new versions, and ask which is clearer and why.

A worked example

Emeri wrote this sample paragraph from a client update to show the passes at work. It is typical of a first draft written in a hurry:

• Passive voice: "has been completed by our team" and "it is anticipated" hide who did what.
• Nominalizations: "implementation", "utilization" and "visibility" stand in for verbs.
• Filler: "very", "really" and "going forward" add nothing.
• Claims: "improve visibility across all of your locations" is a promise. What exactly will the client see, and where?

The model cannot fix the last point, because it does not know the facts. Only the writer does. The writer rewrites the paragraph in plain words and adds the specifics:

The rewrite is half the length. It names the actor, the date and what changed for the client. No phrase in it came from the model.

Where this applies

• Proposals and quotes, where your voice is part of the sale.
• Client emails that need care, such as a delay or a price rise.
• Job posts and offer letters, which candidates read closely.
• Website pages and newsletters, where generic copy is easy to scroll past.

Routine text, such as a meeting summary or an internal note, can still start from an AI draft. Check its facts and send it. Keep the editor method for writing that carries your name to a client.

The editor method also keeps review short: you know what you wrote. For why checking now takes so much time, see AI drafts faster than your team can check it.

Emeri did not run a test for this guide. The study and the method come from the sources linked above.