Editing someone else’s draft carries obligations that editing your own does not

The two jobs look identical and are not
On your own draft, you are the authority on intent. When a suggested rewrite changes the emphasis of a sentence, you know immediately whether the new emphasis is what you meant, and you accept or reject on the basis of something only you have access to.
On someone else’s draft, that authority is missing. A rewritten sentence may be clearer and may also say a subtly different thing, and you have no way to tell which, because the distinction lives in what the writer intended rather than in the words on the page.
This is why the same tool that makes you faster on your own work makes you dangerous on another person’s. The mechanical corrections are safe. The improvements are where the risk sits, and improvements are what these tools produce most enthusiastically.
Sort the changes by whether they alter meaning
Some categories are effectively risk-free: spelling, agreement, punctuation, consistent formatting of names and numbers, a genuinely duplicated word. Applying those silently is ordinary editorial hygiene and nobody expects to be consulted about a missing full stop.
A second category changes rhythm and register without changing content — splitting a long sentence, replacing an unusual word with a common one, reordering a list. These are matters of voice, and a writer with a distinctive one will notice all of them. They belong in a suggestion, not in an application.
The third category changes what is claimed. A hedge removed, a qualifier tightened, an example generalised, a maybe promoted to a probably. These are substantive edits and they should be flagged individually, because a writer reading a clean rewritten paragraph will very often not notice that the argument moved.
Ask for annotated suggestions, not a corrected version
The single most useful instruction when editing another person’s work is to return the original unchanged with proposed edits listed separately, each one paired with a reason. A returned clean copy hides the size of the intervention behind a fluent surface, and comparing it against the original is far more tedious than reading a list.
Requiring a reason for each change also filters the list. Suggestions that exist only because a rewrite was requested tend not to survive the requirement to justify themselves, and what remains is closer to what an editor would actually have raised.
It is worth explicitly excluding categories you do not want touched. Leave the vocabulary alone, do not shorten sentences, do not change the order of the sections. Without exclusions you will get a full pass, and the writer will receive back something that reads like a different person wrote it.
Voice is not decoration and it is what gets flattened first
The habits that make a writer recognisable are usually the ones an automated pass identifies as irregularities: a long digressive sentence, an unfashionable word, a paragraph that starts with a conjunction, a joke that does not signal itself. Smoothing those produces prose that is technically cleaner and noticeably less alive.
For internal documents that may be an acceptable trade. For anything published under a person’s name it usually is not, and the writer is entitled to be the one who decides. Handing back a version stripped of everything distinctive, with the message that it reads more professionally now, is a specific way to insult a colleague.
A reasonable default is to restrict the pass to correctness and clarity, and to raise anything stylistic as a question. If the writing is unclear in a way that matters, say where the confusion begins and let the writer solve it, which is what a good human editor mostly does anyway.
Disclosure is the part that is not technical
Norms here are genuinely unsettled and vary by workplace, publication and country, so there is no universal rule to state. What is fairly clear is that people care more than editors expect about whether their words were passed through a tool, particularly when the work is creative or personal.
The safe practice is to say what you did. A line noting that a draft was run for grammar and consistency, and that everything else is a suggestion, costs nothing and prevents the discovery-after-the-fact that damages working relationships. It also gives the writer the chance to object in advance.
There are contexts where you should not do it at all: confidential material that must not leave a system, work under an agreement that forbids it, a student’s writing where the labour is the point of the exercise. Those constraints come from the situation rather than from the text, which is why no instruction you write can identify them for you.
Common questions
Which edits are safe to apply without asking?
Objective corrections — spelling, agreement, punctuation, consistent formatting. Anything touching word choice, sentence length or emphasis belongs in a suggestion, because those are the writer’s decisions even when a change would be defensible in isolation.
Why ask for a list of changes instead of a clean rewrite?
A clean rewrite conceals how much moved. Reading a fluent revised paragraph, most people do not notice that a qualifier disappeared or an example was generalised. A list forces each change to be seen and, if you require reasons, filters out the ones that exist only because a change was requested.
Should I tell the writer a tool was involved?
Norms differ and are still forming, but disclosure costs little and prevents a discovery that damages trust. In some settings — confidential material, agreements that prohibit it, work where the writing itself is being assessed — you should not run the pass at all.
Reporter, Prompt After Prompt
Devika has been reporting on prompt craft, writing with ai, images & audio since long before it was fashionable and reads the small print so you do not have to.