How researchers humanize literature reviews without losing the citation graph

title: "How researchers humanize literature reviews without losing the citation graph" description: "Researchers draft long lit-review prose with AI assistance — and need the citation structure to survive humanization intact. Here's the workflow." date: "2026-05-13" tag: "Use case" author: "Inksong"

Literature reviews are, in our experience, the highest-volume AI-assisted writing happening in academic research right now. They're long, they're synthesis-heavy, and the prose between citations is largely formulaic — which is exactly the territory where AI drafting is most useful and most obvious. They're also the highest-stakes section for citation integrity. A misplaced reference, a hallucinated author, a year that drifted by one in the rewrite, and a peer reviewer flags it. The whole submission slows down. So the workflow for humanizing a lit review has to do a specific thing: rewrite the prose, leave everything load-bearing alone.

The setup

A realistic workflow looks like this. The researcher has a reference manager — Zotero, Mendeley, or similar — with the working bibliography for a project already organized into a folder. They've drafted a 3,000-word lit-review section with AI assistance, either by prompting against summaries of each paper or by feeding actual passages and asking for synthesis. The draft compiles. The citations exist. The prose reads competent but generic — the signature patterns of default AI writing are all there: uniform sentence length, dense hedging, transitions that feel like they were generated rather than chosen.

The researcher wants the prose humanized without disturbing the citation graph. That's the brief.

What Inksong protects

Citation anchors stay intact. We recognize the common citation patterns — (Author, 2023), (Author & Other, 2021), (Author et al., 2020, p. 47), numeric [1], [1,3,5], footnote markers — and treat the content inside them as do-not-touch. The surrounding prose can be rewritten freely; the parenthetical itself is preserved character-for-character. This is the single thing that has to work for a lit review humanization to be safe, and it's the thing we test most aggressively.

Section headings are preserved. "2.1 Recent advances in X", "Methodological approaches", "Theoretical framework" — heading lines are detected as structural rather than prose and pass through unchanged.

Technical terms are preserved by the academic domain preset. Domain-specific vocabulary — taxon names, method names, theory names, instrument names — is treated as fixed terminology rather than something to vary. If your draft says "polymerase chain reaction" four times, the rewrite doesn't substitute "PCR" in some places and "the technique" in others to avoid repetition; it leaves the term alone.

Reference lists at the end of the document are explicitly skipped. The formula patterns of a Vancouver or APA reference list — author, year, title, journal, volume, pages — don't need humanizing, and rewriting them is how you introduce errors. We detect the reference-list block and bypass it entirely.

What Inksong rewrites

The prose between citations. Sentence rhythm, paragraph transitions, hedging style, connective tissue. The argument shape stays — the claims and the order they're made in — but the way each claim is phrased changes. A hedged sentence like "It has been suggested that X may influence Y (Author, 2022)" becomes something with more varied phrasing and less default-LLM cadence, while the citation pin stays exactly where it was.

Recommended settings

For lit reviews specifically:

• Tone: academic.
• Domain: academic.
• Humanness: 35. Lower than the default 40, because lit reviews are a more conservative genre than other academic writing — too much variation in sentence length makes the section read wrong, the way it would in an abstract.
• Voice Profile: ideally a prior paper you wrote yourself, 3,000+ words, in the same field. Your own past published prose is the best training material for your future published prose.

If you don't yet have a voice profile set up, the academic tone and domain alone will produce something usable. The voice profile is the upgrade.

The workflow, end to end

1. Draft in your usual environment with AI assistance. Get the section to the point where the argument and citations are correct. Accept all tracked changes so what you upload is the final draft, not a marked-up one.
2. Upload the DOCX to Inksong. DOCX preserves formatting through humanization, including headings, italics, and citation styling.
3. Apply the voice profile (if you have one) plus the settings above. Run humanization.
4. Review the diff carefully. Spot-check five citation anchors at random — confirm each one survived unchanged, including the year, the page number, and any "et al." Spot-check five technical terms in the same way. If anything looks off, that's a bug report we want.
5. Download the rewritten DOCX. From here it goes back into your normal workflow — coauthor review, journal template, whatever comes next.
6. If your institution or target journal requires a third-party AI detection check, run it externally. Our score is internal heuristic — useful for iteration, not a substitute for the detector your reviewers will actually use. The honest version of this is covered in Honest about AI detectors.

What to verify by hand

A short list of things only you can check, no matter how careful the humanization is.

• Quote integrity. If you've quoted a source directly, verify the quote text matches what's in the original paper. Humanization shouldn't touch quoted material, but verify anyway — at peer review, this is what they check first.
• In-text reference accuracy. Not whether the anchor survived (we handle that), but whether the cited claim actually appears in the cited paper. AI drafting sometimes attributes claims to the wrong source; humanization won't fix that, because the citation pin is locked.
• Authorial qualifications. Argument structure like "we believe X because Y" should keep the "because" clause. Watch for places where a careful conditional or causal connector got softened. The rewrite shouldn't drop these, but it's worth a read-through to confirm.

Closer

See Solutions → Researchers for the audience-level framing, and the academic guide for a single-paper walkthrough.