AI Writing Tips

The Four Repairs: A Procedure for Editing AI Drafts

· 5 cited sources

“Improve this” is not an instruction. It is a shrug with a verb attached. You hand the model a draft it wrote, ask it to make the draft better, and it does what it always does: it smooths. The result reads a little cleaner and fails in exactly the same four ways it failed before, because “improve” has no idea which of the four problems you meant. An AI draft does not break in one place. It breaks in four, and the four breaks need four different repairs, applied in order, by a person who knows which one they are doing. Name them and the editing stops being a vibe you chase and becomes a procedure you run: verify the facts, re-voice the prose, restructure the argument, restore the authority. Skip the names and you will keep asking a machine to fix the draft the machine produced, which is like asking the fox to recount the hens.

This is a procedure, not a checklist of tips, and it belongs to the editing pillar because every step is something you do after generation, on a draft that already exists. The order matters and the separation matters. Try to do all four in one reading and you will do none of them well — your eye cannot hunt for a fabricated citation and listen for a flat sentence in the same pass. Each repair has its own failure mode, its own tool, and its own way of being faked. Here they are, one at a time.

Repair 1: Verification — is any of this true?

Start here because this is the repair that, left undone, ends careers. The model’s single most dangerous habit is producing citations, statistics, and quotations that are fluent, specific, formatted correctly, and entirely invented. The rate is not marginal. When Safran and Çalı prompted GPT-4 to write four scientific paragraphs on musculoskeletal rehabilitation, each with ten DOI-bearing references, only 7.5% of the forty references came back fully accurate while 42.5% were completely fabricated and the remaining half were partially wrong (Safran & Çalı, 2025, p. 695). On the worst of the four prompts — rotator cuff repair — six of ten references were fabricated outright and four of ten DOIs simply did not resolve (Safran & Çalı, 2025, p. 698). These are not typos. They are plausible-looking works that do not exist.

The numbers hold across models and get worse in high-stakes domains. Reviewing the field, Misra and Udandarao report citation-hallucination rates ranging from 18% in GPT-4 to over 70% in other frontier models, climbing to 88% in legal contexts, and note the benchmark gap starkly: on CiteME, language models cite scientific claims accurately only 4.2% to 18.5% of the time against 69.7% for human annotators (Misra & Udandarao, 2026). A controlled comparison they cite found ChatGPT-4o hallucinating references at 20.0% while Gemini Advanced reached 76.7% (Misra & Udandarao, 2026). The lesson of verification is that fluency carries no information about truth. The sentence that names a study and the sentence that invents one are written in the same confident register.

There is a tempting shortcut here, and it half-works, which makes it dangerous. You can ask the model to check its own references, and accuracy does improve: in the Safran study a follow-up verification prompt lifted fully-accurate references from 7.5% to 77.5%, a statistically significant gain across every topic tested (Safran & Çalı, 2025, p. 699). But read what that number is. The verification was performed by the same model that fabricated the citations, and it still left roughly a fifth of them wrong. A self-audit that misses one reference in five is not a verification step; it is a confidence boost. Whether a model can check its own sources has a precise answer — better than nothing, nowhere near enough — and the only reliable version of this repair happens outside the model. Paste the reference list into an independent tool like cytado.com’s bibliography checker, which confirms for free whether each entry resolves to a real work, then chase the survivors by DOI to the source and read the actual page. Verification is the repair that cannot be delegated to the thing that needs verifying.

Repair 2: Re-voicing — does it sound like a person?

Once the draft is true, it is still bland. That blandness is the second break, and it is not a matter of taste — it is a measurable property of the text. AI prose converges on a narrow band: sentences of even length, a thin set of recycled connectors, mid-register vocabulary, flattened emotion, and a predictability so high that detectors score it instead of reading it. I have made the full argument elsewhere about why AI prose all sounds the same — the short version is that the sameness has coordinates you can compute, which means it has coordinates you can edit against.

Re-voicing is the pass where you put the variance back on purpose. Break a paragraph of same-length sentences by setting a four-word sentence against a thirty-word one. Cut the however / moreover / additionally scaffolding and let some sentences simply follow others. Take a position where the model hedged, land a specific noun where it reached for a generic one. The key discipline is that this is a separate pass from verification. If you try to re-voice while you are still checking facts, you will start rewriting sentences whose claims you have not confirmed, and a re-voiced fabrication is harder to catch than a plain one — you have now dressed the lie in your own cadence. Verify first so that re-voicing only ever touches sentences you already trust.

A warning that belongs to this repair: re-voicing is not disguising a machine draft to beat a detector. That is a different and worse project, and it inverts the point. You re-voice because a human reader can feel the flat average and will stop reading; you re-voice to make the prose carry a mind, not to launder authorship. The moment the goal becomes “sound less like AI to fool a tool” rather than “sound like me because I thought this,” you have left editing and entered evasion.

Repair 3: Restructuring — is the argument in the right order?

The third break is the one editors most often miss, because the draft can be true and well-voiced and still badly built. Models have a default architecture: an introduction that previews, three balanced body sections of roughly equal weight, a conclusion that restates. It is a shape, not an argument. Real reasoning is lopsided — one point deserves four paragraphs and two deserve a sentence each; the strongest objection belongs early, not buried in a dutiful “limitations” aside near the end. The model produces symmetry because symmetry is the average of all the essays it has seen. Your actual argument is almost never symmetrical.

Restructuring means reading the draft for its load-bearing claim and rebuilding the order around it, which is editorial work no amount of sentence polishing performs. There is a useful analogue in how writing pedagogy handles accuracy under pressure. Studying medical students writing in English as a second language, Aldafas describes an “accuracy-support model” that sequences revision into three recurring checkpoints — concept clarification, terminology verification in context, and precision-focused revision — specifically to make verification “an explicit part of the writing process” rather than a spot-check applied when something looks off (Aldafas, 2026, p. 10). The transferable idea is that effective revision is structured — named stages run in sequence — not a single undifferentiated “make it better.” An AI draft benefits from the same discipline: decide what the piece is arguing, promote the evidence that carries it, demote the filler the model inserted for balance, and cut the preview-and-restate scaffolding that says nothing twice. This is also where the work you saved by generating reappears, which is the whole thesis of the efficiency paradox: drafting got cheaper, so judgment about structure became the expensive part, and it is still yours to supply.

Repair 4: Restoring authority — whose text is this now?

The last repair is the one with no technical tell, and it is the reason the first three are worth doing. A verified, re-voiced, well-structured AI draft can still be a draft that commits to nothing. Models hedge by default — some researchers suggest, it could be argued, this may indicate — because hedging is safe and safety is what they optimize for. Authority is the quality of a text that takes responsibility for its claims: it says what is true, names who is wrong, and stands behind a judgment a reader can hold the author to. Restoring it means going back through the draft and replacing the model’s practiced non-commitment with your actual position, the one you can defend when someone pushes.

This repair has become, in 2026, a formal obligation as well as a craft one, and the two point the same direction. Publishers now draw the line at exactly this boundary. Elsevier’s updated journal policy states plainly that generative AI cannot be listed as an author because it cannot take responsibility for the work, that basic language and grammar help need not be declared, but that substantive AI-generated content must be disclosed and the human authors must confirm they reviewed it and accept responsibility (Elsevier, 2026). CASRAI’s editorial on disclosure makes the logic sharper still: AI is recorded as a tool or method, never a contributor, and disclosure “does not transfer responsibility; it makes the workflow transparent so that responsibility can be located” (CASRAI, 2026). Responsibility can be located — that is the test. When a reader traces a claim back to its author, someone has to be standing there. Restoring authority is the act of making sure that someone is you: that every claim in the finished piece is one you have verified, understood, and chosen, not one the model defaulted into while you were watching the fluency instead of the substance.

Running the procedure

The four repairs are ordered for a reason. Verify first, because re-voicing or restructuring an unverified draft just makes falsehoods more convincing. Re-voice second, because you need true sentences before you make them sound like yours. Restructure third, once the sentences are trustworthy and readable enough to judge as an argument. Restore authority last, as the final read where you ask of every paragraph: do I stand behind this? Each pass is narrow enough to actually finish, and narrow passes are what make the difference between editing and the thing most people do, which is reading the draft once, feeling vaguely dissatisfied, and typing “improve this.”

The procedure also tells you what the model can and cannot do for you. It can help you verify — by listing its own claims for checking — but it cannot be the verifier. It can suggest re-voicings, but the choice of which to keep is the authorship. It can propose a reorganization, but the argument it is reorganizing has to be one you actually hold. “Improve this” hands all four decisions back to the system that made the mess. Naming the four repairs keeps them where they belong, with the one party who can be held responsible for the result. The draft arrives in four kinds of broken. Fix them in four passes, in order, and the piece that comes out the other side is one you can sign.

Sources

Every factual claim above is tied to a source you can open and check, with page numbers wherever the source has them.

  1. 1. Ertuğrul Safran, Adem Çalı, Fabricated or accurate? Ethical concerns and citation hallucination in AI-generated scientific writing on musculoskeletal topics , Anatolian Current Medical Journal, 2025 , pp. 695, 698, 699. 10.38053/acmj.1746227
  2. 2. Nipun Misra, Vikranth Udandarao, Detecting Citation Hallucinations in Large Language Model Outputs (Student Abstract) , Proceedings of the AAAI Conference on Artificial Intelligence, 2026 . 10.1609/aaai.v40i48.42257
  3. 3. Abdulaziz H. Aldafas, Challenges and coping strategies for achieving terminology precision in EMI medical writing among Saudi medical students , Frontiers in Medicine, 2026 , pp. 9, 10. 10.3389/fmed.2026.1781287
  4. 4. Elsevier, Updated generative AI policies for journals: supporting responsible use while protecting trust , Elsevier, 2026 . link
  5. 5. CASRAI, Disclosing generative AI use: what to declare and where , CASRAI, 2026 . link
More on Editing the AI Out
Built by Sitario.com