What actually changes when we edit an AI first draft

People sometimes ask whether an AI draft needs less editing than a human one, on the grounds that the grammar is already clean. If anything, it’s the other way round. The mechanical layer might be fine, but everything above it needs more attention than usual.

What stays the same

Grammar, punctuation, and spelling are mostly right in a generated draft, and where they aren’t, the errors are ordinary ones. We still read for them, but they take up less of our time than they used to.

What changes

The length. Generated drafts tend to be long because length costs the model nothing. We cut the repetition and the summaries at the end of every section, and most drafts come down substantially without losing any content.

The claims. Every factual claim is scrutinized. Where a real source says what the draft says, the claim stays. Where there isn’t one, it’s flagged in the margin for you, because it’s not for us to decide what you’re willing to assert; that’s the author’s call.

The hedging. Models qualify almost everything. We mark where a claim has been softened past the point of saying anything, and you decide how far to commit.

The structure. Generated drafts tend to treat every section as equally important and to move between points without connecting them. We put the finding where the reader will look for it and highlight any gaps in your reasoning so you can fix them before the reader wonders about them.

The voice. The draft gets read against how your organization sounds, and everything that sounds like the model rather than like you is changed or cut. If your organization’s voice isn’t written down anywhere yet, we can draft it with you, so that everything produced or edited for you stays consistent, however it was first written.

The specifics. We can see where a general statement is standing in for a particular fact and will ask for it, but only you can supply it.

What comes back

Tracked changes and margin comments, as with any other edit. The comments matter more than usual on an AI draft, because so many of the decisions are yours, from which claims to stand behind to which hedges to remove, and the edit isn’t finished until you’ve worked through them.

What we won’t do

We won’t run the draft through a detector or edit to lower a score. If you’re being asked for a detector result, start here.