Editing a ChatGPT Draft Into a Submittable Manuscript

A draft comes back from ChatGPT in ninety seconds and reads well enough that the obvious move is to start at the first sentence and polish forward. Two hours later the paragraph that took longest gets deleted, because the claim holding it up turned out to have nothing behind it. The order you edit in decides how much of that work you throw away.

How to edit ChatGPT output in a research context runs in four passes, ordered by how expensive a mistake is to fix later: claims and evidence first, structure second, sentence rhythm third, vocabulary last. Each pass assumes the one before it has stopped moving. Run them in any other order and your best attention goes into sentences that will not survive the week.

How to Edit ChatGPT Output in Four Passes

Edits aren't equally destructive. Moving a paragraph requires rewriting all the joins you smoothed out. Adding a sentence in the third revision resets the rhythm of the paragraph you balanced in the second. Deleting an unsourced claim destroys the sentence you spent twenty minutes tuning. The order exists because we want to work from the most destructive change to the least and leave each pass intact.

Pass	What you are fixing	Why it sits here
One: claims and evidence	Whether each factual sentence is true and traceable to a source you have opened	A sentence you are going to delete is not worth polishing first
Two: structure	Whether the section order carries your argument rather than a list of topics	Moving a paragraph after line-editing means editing every join again
Three: sentence rhythm	Uniform sentence length and repeated clause shapes inside a paragraph	Every insertion in passes one and two resets a paragraph's rhythm
Four: vocabulary	Stock verbs and abstractions that carry no information	The cheapest edit and the smallest lever, so it goes last

Pass One: Check Every Claim Against a Source You Have Opened

Read the draft once with no intention of improving the writing. On this pass you are only marking sentences that assert something factual: a number, a mechanism, a finding, a claim about what the field currently thinks. Everything else can wait, and the temptation to fix a clumsy phrase in passing is the thing that quietly turns a twenty-minute pass into an afternoon.

Give each marked sentence one of three marks. First mark: I have read the source and can cite it. Second mark: I know this is true and need to go and find the source. Third mark: the model asserted this and I've no idea where it came from. Sentences with the third mark get deleted rather than rewritten. That's the hardest instruction here to follow, because those sentences are usually the smoothest ones on the page.

A model draft attributes claims to nobody in particular, and the phrasing gives it away. Recent studies suggest, researchers have found, it is widely accepted: each of those is the shape a claim takes when no specific paper sits behind it. Every one is either a real citation you have not added yet or a sentence with nothing underneath, and the way to find out which is to try to name the paper. If you can, name it in the sentence. If you cannot, the sentence goes.

References the model supplied need checking against a database before they go anywhere near your reference list, since a plausible author, year and journal can be assembled without a matching paper existing. An AI citation checker resolves the identifiers for a whole list in one pass, which is faster than searching each title by hand and considerably faster than discovering the problem in a supervisor's margin note. Do this before you cite anything the draft handed you.

Pass Two: Rebuild the Structure Around Your Argument

Structure comes second because moving a paragraph after you have line-edited it means editing all its joins again. On this pass you are deciding what goes where and what gets cut entirely, and you should expect the draft to get shorter. A model produced a complete-looking document from a prompt, which is a different object from a document built to make one argument.

Read only the first sentence of every paragraph, in order, and nothing else. It should read like an argument: this is what we asked, this is how we looked, this is what we found, this is what it means. The one specific way a generated draft will fail this test is by reading like a list of topics rather than a line of reasoning: the model was completing a prompt paragraph by paragraph with no thesis to serve.

The second structural fault is even weighting. A model gives every subtopic roughly the same amount of room, since nothing in the prompt told it which part matters. A real manuscript is lopsided. The finding you actually have, the limitation that actually bit, the method you changed halfway through: those earn more space than the background, and the background usually loses a third of its length on this pass. While you are there, cut the sentence at the end of each section that restates the section. A reader who has just read it gains nothing, and a generated draft produces one almost every time.

Pass Three: Fix the Rhythm Once the Content Stops Moving

Sentence rhythm is worth fixing only when the content has settled, because every insertion in the two passes above resets it. What you are looking for is a run of sentences that share a length and a shape. Four consecutive sentences of almost identical length will sound flat read aloud, and uniform sentence length across a passage is the structural signature that both a careful reader and a statistical classifier pick up.

The repair is mechanical. Read the paragraph out loud, mark any run of two or three sentences that land the same way, then split one of them or fold two together with a comma doing real work. A long sentence has to earn its length by carrying a complication, a condition or a consequence that a short sentence would have to state twice. Worked examples of these moves, with the before and after versions side by side, are in the guide on how to humanize AI text.

Resist the reverse temptation. Filler sentences, which add nothing but exist only to disrupt rhythm, are spotted by a reader within two paragraphs. Variation has to come from the content having different amounts to say at different moments, which is the same reason a real research paragraph is uneven in the first place.

Pass Four: Vocabulary, Last and Smallest

Vocabulary goes last because it is the cheapest edit and the smallest lever, and doing it first delivers the feeling of progress without the substance. A generated draft leans on a small set of high-frequency academic verbs and abstractions, and a reader who marks a hundred papers a term has seen all of them. Facilitate, underscore, pivotal, robust, myriad and showcase are the usual suspects in a research draft.

Replace a stock verb with the one the sentence actually needs, and use the information test to decide: showed, measured, doubled and failed each tell a reader something that facilitate does not. A free AI word cleaner catches most of the stock vocabulary in seconds, which is the right amount of time to spend on it. Overcorrection is the real risk on this pass, since a draft that has been aggressively de-machined reads as a thesaurus exercise, and a supervisor notices that faster than any software does.

What the Four Passes Leave You With

A manuscript you can defend is one where you can point at any sentence and say where it came from. That is the actual deliverable, and every pass above is a step toward it: pass one gives every claim a source, pass two gives the argument a shape, passes three and four make the result readable at the level a journal or a marker expects.

An AI writing score usually falls across these passes without anyone aiming at it, because each pass removes something the score is computed from: unsourced generic claims, even section weighting, uniform sentence length, stock vocabulary. Treating the score itself as the target inverts the work. Turnitin's own guidance is blunt about the limits of the number, stating that its model "may not always be accurate (it may misidentify human-written, AI-generated, and AI-paraphrased text), so it should not be used as the sole basis for adverse actions against a student."

One piece of context on tooling, then back to the manuscript. An AI humanizer can do the third pass across a long document faster than a person can, and it reports an estimated Human Score computed from the text rather than a verdict from any named detector. It cannot do the first pass at all, because it was not in the room when the work happened and has no way to know which of your claims has a source behind it.

Whether the draft's origin needs stating in the manuscript is a separate question with a real answer, and it depends on your institution or target journal rather than on how much of the text survived editing. The guidance on how to disclose AI use in a paper covers the wording most policies expect and where it belongs in the document.

If the immediate problem is a number on a report, lowering an AI score is treated on its own page. Whether the tools built to do that actually work is a fair prior question, and it has its own answer.

The four passes double as a budget. With one evening before a deadline, spend it on the first two and let some sentences stay slightly awkward. An awkward sentence attached to a finding you can defend survives a viva. A beautiful one attached to nothing does not.

Related research: the findings above are examined at scale in The Detectability of Partially AI-Rewritten Academic Documents, a TextPulse Research working paper with open data, code and a citable DOI. Whether prompting alone can make a model write like a person is tested in Do AI Models Speak Human?.