Give Your AI Editor a Smaller Job

Key takeaways

• A polished rewrite can change a claim’s meaning, scope, or certainty.
• An overly agreeable AI editor can leave weak arguments unchallenged.
• Asking for specific problems before replacement sentences keeps editorial decisions with the author.
• Every accepted edit should have a reason you can explain.

A polished sentence can quietly turn a cautious observation into a product promise. Thomas Ptacek’s “How to Write with an LLM” raises a useful question for anyone publishing under their own name: how much of the writing should you hand over?

A smoother sentence can make a bigger claim

“Make this sound more natural.”

It’s a reasonable editing request. But natural language and accurate language are separate things.

Suppose your draft says:

This feature could reduce the time spent on this task.

Now imagine the suggested rewrite:

This feature dramatically improves productivity.

That sentence has acquired quite a promotion.

“Could” has disappeared. A claim about time spent on one task has expanded into a claim about productivity. “Dramatically” adds a judgment that needs evidence of its own.

The edit has changed the strength of the claim. Checking only whether the sentence reads well will miss that.

Editors routinely replace words and rebuild sentences. Using an AI suggestion doesn’t automatically amount to handing over authorship. But when a revision changes what you’re asserting, accepting it requires a fresh decision about whether you can support it.

An agreeable editor can miss the point

One useful term here is sycophancy: the tendency to agree with users or tell them what they want to hear.

That tendency can undermine editing. If the premise is weak and the model simply improves the prose, you may end up with a more persuasive version of an unsupported argument.

Consider a familiar product marketing request:

Make a more convincing case that our service is better than the competition.

The conclusion is already baked in. Before adding stronger adjectives, an editor should ask what “better” means. Faster at which task? Easier for which users? Compared with what?

Give the model room to challenge the argument:

Read this from the perspective of someone who disagrees with the conclusion. Identify claims that lack evidence and objections the draft needs to address.

The resulting criticism still needs scrutiny. A model can invent a problem or misunderstand a sound argument. Treat its objections as questions to investigate, rather than verdicts.

Get the diagnosis before the rewrite

Try one editing session in which you don’t use any replacement sentences suggested by AI.

This is an exercise, not a permanent ban. The point is to practice deciding what the draft needs.

A useful prompt is:

Don’t write replacement sentences. Identify ambiguous wording, claims that need evidence, and gaps between ideas. For each issue, quote the relevant passage and explain why a reader might struggle with it.

Take this hypothetical release note:

This update completely resolves users’ frustrations.

There are several questions hiding inside that sentence. Which frustrations? What changed? What evidence supports “completely”?

A diagnosis gives you work to do. You have to supply the missing information.

If the actual update added retry instructions to a payment error message, write that:

The payment error message now explains how to retry the transaction.

That sentence describes an observable change. Whether it resolves the underlying frustration is a separate claim.

Writing the fix yourself makes you check the distance between what happened and what the draft promises. It also forces a useful choice: which information does the reader actually need?

Every edit needs a reason

You can accept a typo correction without turning it into a philosophical inquiry. Suggestions that remove repetition or identify missing background can be useful, too.

For more substantial edits, pause long enough to ask:

• What was wrong with the original?
• Did the revision change the meaning or degree of certainty?
• Have I checked any facts or judgments it introduced?

If you can answer those questions, you understand the edit. If your only explanation is “it sounds more convincing,” give it another read.

Pay particular attention to words you wouldn’t normally use. An unfamiliar word might be more precise. It might also be doing little beyond making the sentence sound expensive.

Your voice includes the claims you’re willing to make and the places where you leave room for doubt. Those choices deserve as much attention as rhythm and vocabulary.

At your next editing pass, ask AI to identify three places where a reader might get stuck, then write the fixes yourself. The finished prose should reflect decisions you can explain, down to the smallest “could.”