CiteOwl
CiteOwl

How to fix an AI-generated or inherited draft

A draft you did not fully write, whether a chatbot produced it, a group partner handed it over, or your own past self left it half-finished, needs triage before it needs editing. The first ten minutes are not spent fixing anything. They are spent finding out which of three drafts you are holding, because a draft built on invented sources needs a different plan from one that is merely disorganised, and the plan you pick in minute ten decides whether the next six hours are useful.

The instinct is to open the file at the top and start improving sentences. That feels like work and it is the slowest possible route, because every small fix you make early is thrown away by a large decision you make late.

The ten-minute triage

Three checks, ten minutes, no editing. You are not fixing the draft yet, you are working out what it is.

Sample five references

Pick five references at random from the list, not the first five. For each, paste the exact title in quotation marks into Google Scholar, and if there is a DOI, paste it after https://doi.org/ in your browser. A real paper turns up as the top hit and a real DOI loads the article's page; a fabricated one returns nothing and a fabricated DOI returns a "DOI Not Found" error. Two minutes each.

Five is a deliberate number. It is enough to tell a mostly-invented list from a mostly-real one, and not enough to certify anything. The arithmetic: at the 55 percent fabrication rate a peer-reviewed audit measured for GPT-3.5, a sample of five would come back clean about 3 percent of the time; at the 18 percent rate the same audit measured for GPT-4, about 37 percent of the time (Walters and Wilder, Scientific Reports, 2023, doi.org/10.1038/s41598-023-41032-5). So five clean is weak evidence of a clean list, while two fakes in five is strong evidence of a broken one.

Count the sections and read only the headings

Write the headings out in order on a separate page. Four minutes. You are looking for two things: whether the order makes an argument, and whether any heading has nothing under it but a placeholder.

Find the claim

Read the last paragraph of the introduction and the first of the conclusion. Can you state in one sentence what this draft argues? If you cannot, that is the most important thing you have learned in the ten minutes, and it is not a prose problem.

Now you know which draft you have.

What the triage found What you are holding What to do
Two or more fake references in fiveA draft built on sources that do not existKeep the structure, delete the reference list, rebuild from real sources
References real, no argument findableA well-sourced pile of summaryDecide the claim first, then reorder everything around it
References real, argument findable, order wrongA normal messy draftWork the order below, start to finish

Our advice on the first row is stronger than most guides will give you, and here is the reasoning. If a third of a sampled reference list is invented, do not repair the citations one at a time. Delete the list. A draft written around fabricated sources has an argument shaped by findings that were never published, so fixing references individually leaves you defending claims nothing supports, in an order those non-existent findings dictated. Keep the section structure, keep any sentence that states your own reasoning, and rebuild the evidence from sources you find yourself. It sounds like more work and it is less, because the alternative is discovering the same problem one reference at a time over two days.

Where the citation check belongs

Most guides tell you to verify the citations first. That is wrong, and the reason is arithmetic. A proper check on one reference costs two to four minutes: find the paper, confirm the details, open it, read the passage it was cited for. On a draft with thirty references that is an hour and a half, and doing it before you fix the structure means spending a chunk of that time verifying sources for sections you are about to delete.

So the full check goes after the structure pass, on the sections that survived it. What goes first is the five-reference sample, which is a different operation with a different purpose: it costs ten minutes and it tells you which plan to follow. Sample first, verify later.

There is one exception worth knowing. If the draft is due in a few hours and you cannot do everything, verify the citations and skip the polish. An unpolished paragraph costs a mark. A fabricated citation is an academic integrity conversation.

Fixing the skeleton

Take the heading list from the triage and put a one-line note under each saying what that section is supposed to do for the argument. Writing teachers call this a reverse outline, and it is the fastest way to see the shape a draft actually has rather than the one it was meant to have. The problems announce themselves as you write it: two sections arguing the same point, a results section arriving before the reader knows the method, a heading with nothing under it, a paragraph that wandered into the wrong section and stayed.

Then fix the order. Merge the overlapping sections. Cut the section that does not earn its place, even when it is the best-written thing in the file, because well-written and irrelevant is still irrelevant. Mark the missing sections so you know to come back to them. Reorder so each section sets up the next.

A draft assembled over several sittings has one predictable structural tell: the sections sit in the order they were written rather than the order they should be read. Generated drafts have a sharper version of it, where every section runs the same length and covers its topic in the same shape, because nothing in the process knew which point was load-bearing. If your draft's sections are suspiciously even, decide which two matter and let them be twice the size of the rest. Our companion page on outlines and section proportions has measured numbers for what published papers actually spend where.

The citation pass, in three buckets

Now go through every reference in the sections that survived. Two checks each, in this order.

Does the source exist? The title search and the DOI resolution from the triage. A perfectly formatted DOI that leads nowhere is the most reliable single tell there is, because the format is exactly the part a language model gets right.

Does it support the sentence it is attached to? Open it and read the relevant passage. This is the check people skip and it catches the more common failure. In the same audit, among the citations that did point at genuine papers, 43 percent of the GPT-3.5 set and 24 percent of the GPT-4 set still carried substantive errors. A real paper cited for a claim it does not make is not a smaller problem than a fake one, and a marker who opens it reads it as misuse rather than as an honest slip.

While the paper is open, check one more thing that costs five seconds: whether it has been retracted. Since Crossref acquired the Retraction Watch database in September 2023, retraction data is free and public and rides along with the record itself (Crossref on the Retraction Watch database). An inherited draft is exactly where a retracted source survives, because nobody has looked at that reference since the day it was added.

Every reference then lands in one of three buckets.

The full routine, including the signals that should stop you on an entry, is in how to check if a citation is real, and the specific case of a chatbot draft is in how to get ChatGPT to cite real sources.

Filling the gaps, claim by claim

With the structure settled and the citations honest, the holes are visible: the sections you marked missing, and the sentences that assert something and move on. Work claim by claim rather than citation by citation.

For each unsupported sentence, ask what evidence would genuinely back it, find a real paper that provides it, read enough of it to be sure, and attach it. Three things can happen and all three are useful. The claim survives with a source behind it. The claim turns out to have been overstated and needs softening, which is far better learned now than in feedback. Or no source exists for it at all, in which case the claim itself is probably wrong, and the honest fix is to change the sentence rather than prop it up with a reference that nearly fits.

This is the step that advances the paper rather than tidying it, and in an inherited draft it is usually where most of the remaining work lives.

The seams an inherited draft always has

A document assembled from several sources, sittings or tools carries a predictable set of seams. They are quick to fix and they are most of what makes a draft read as stitched together.

Polish, and only now

Only now do you touch the sentences, because only now do you know which ones are staying. Polishing is the satisfying part, which is exactly why it is tempting to do first and wasteful when you do.

Read for prose alone. Tighten bloated phrasing, break the run-ons, cut filler, and make each paragraph hand off to the next so the thing reads as one document rather than a stitch of passages. Read a section aloud, because your ear catches rhythm your eye skims. Proofread for spelling, grammar and consistent formatting last of all, ideally a day later with rested eyes.

What importing the draft actually does

Every step above works by hand, and the order is the same whatever tool you use. Where CiteOwl helps is that you can hand it the draft you already have instead of rebuilding it in a new document. Import a PDF, a Word file or a LaTeX project including a zip, and it reconstructs the structure as real sections you can reorder, brings figures and equations across as elements rather than flattened images, and resolves every reference against CrossRef and OpenAlex, flagging the ones it cannot match. That is the triage's first check and the structure pass, done as the file comes in.

Two honest limits, because the distinction this whole page turns on applies to the tool as well. Resolving a reference proves the paper exists; it does not prove the paper supports the sentence it is attached to, so imported citations arrive without a supporting quote and the second check stays yours. And a reference the resolver cannot match is not automatically fake, it is unmatched, which is a prompt to look rather than a verdict.

From there every change the agent proposes arrives as a reviewable diff, old text beside new, that you accept or reject, so nothing is rewritten silently behind you. If you would rather not import anything, our free citation checker takes a pasted reference list and resolves each entry against CrossRef with no account, which covers the first of the two checks across a whole list in about a minute.

Things worth knowing.

What order should I fix a messy draft in?
Sample, structure, verify, fill, polish. Spend ten minutes sampling the citations first, not to fix them but to find out what kind of draft you are holding, because a draft built on invented sources needs a different plan from one that is merely disorganised. Then fix the skeleton, because moving a section is cheap and rewriting a section you later delete is wasted. Then verify every citation in the sections that survived. Then fill the gaps where claims have no support. Polish the prose last, when you finally know which sentences are staying.
Should I rewrite an AI-generated draft or fix it?
Keep the structure, question the argument, and be ruthless about the reference list. If a sample of five references comes back with two or more fabricated, our advice is to delete the whole reference list rather than repair it, and rebuild from sources you find yourself. The reason is that a draft written around invented sources has an argument shaped by findings that do not exist, so repairing the citations one at a time leaves you defending claims nothing supports. What is worth keeping in that case is the section structure and any sentence stating your own reasoning.
How do I check the citations in a draft someone else wrote?
Two checks per reference, in order, and the second is the one people skip. First, does the source exist: search the exact title in quotation marks in Google Scholar, and paste any DOI after doi.org to see whether it resolves. Second, does the source actually say what the sentence attributes to it: open it and read the relevant passage. The second check matters because real citations go wrong too. In a peer-reviewed audit of ChatGPT bibliographies, among the references that pointed at genuine papers, 43 percent of the GPT-3.5 set and 24 percent of the GPT-4 set still contained substantive errors.
Can I import an existing draft into CiteOwl?
Yes. You can import a draft as PDF, Word or a LaTeX project including a zip. CiteOwl rebuilds the structure as real sections, brings figures and equations across as elements rather than flattened images, and resolves each reference against CrossRef and OpenAlex, flagging the ones it cannot match. Worth knowing what that does not cover: resolving a reference proves the paper exists, not that it supports the sentence it is attached to, so imported citations arrive without a supporting quote and that judgement stays with you.
Read next.

Import the draft you already started

Free to start. No card needed.

Start writing