How to Write a Literature Review With AI Without Losing Your Sources

The real bottleneck isn't writing, it's tracking

Ask most students what's hard about a literature review and they'll say "starting" or "structure." But the actual bottleneck, the one that eats the most hours, is tracking: which of your forty papers found what, which two disagree, which one is the outlier everyone else cites to dismiss. A literature review is an argument built out of other people's evidence, and an argument only holds up if every claim in it can be traced back to the specific paper, and the specific passage, it came from. Lose that trail even once, and you either have to re-read the paper to find where you got a claim, or you leave it unsourced and hope your supervisor doesn't ask.

This is precisely the step where AI is most useful and most dangerous at the same time: useful because summarising and grouping is exactly what it's fast at; dangerous because a model with no memory of exactly which passage it pulled a claim from will confidently blend two papers' findings into one sentence, or attribute a claim to the wrong author, without any signal that it happened.

Structure before you write a single sentence

Decide how you're organising the review before you touch a draft. Chronological ("here's how the field's thinking evolved") only makes sense if the evolution itself is the point. For most dissertations and theses, thematic organisation (grouping papers by the argument or finding they contribute, not by publication date) produces a much stronger review, because it forces you to state what each theme actually claims rather than just listing who said something in what year.

Do this sorting on paper (or in a spreadsheet) before you open an AI tool at all. Read each source, note the one or two things it actually contributes, and cluster the notes into three to six themes. AI is much better used on a structure you've already decided than asked to invent the structure for you. The latter tends to produce generic, textbook-shaped themes that don't reflect what your specific set of papers actually argues.

Use AI to synthesise one theme at a time, grounded in that theme's papers only

The failure mode to avoid: pasting a vague prompt like "summarise the literature on X" into a general chatbot and taking the answer at face value. That's a generation task: the model is drawing on everything it's ever read about X, not the specific papers sitting in your library, and it will blend in claims from papers you haven't read and can't cite.

The fix is narrowing scope until the AI has nothing to draw on except the sources you actually gave it: feed it the handful of papers belonging to one theme, and ask it to draft a paragraph synthesising only what's in those documents: where they agree, where they diverge, and which claim belongs to which paper. That's a retrieval task, and it's the one AI tools built around your own document library are actually designed to do reliably.

A synthesis you can't trace back to a passage isn't a fact yet. It's a lead.

Keep a claim ledger as you go

This is the single habit that prevents source-drift, and it isn't exotic: it's index cards for the 2020s. For every claim that's going to land in your draft, log four things in one running table: the claim in your own words, the paper it came from, the exact passage or quote it's drawn from, and the page. Do this at the point you first encounter the claim, not after you've already written the paragraph and are trying to reconstruct where it came from.

A ledger like this does two things a memory can't: it catches the moment an AI-assisted summary quietly merges two papers' findings (the ledger will show one claim with two conflicting "exact passage" entries, which is the tell), and it means writing the actual review becomes an exercise in turning rows you've already verified into prose, rather than writing first and sourcing second.

The revision pass that actually matters

Before you submit, go through the draft sentence by sentence wherever a citation appears, and re-open the source passage for that exact sentence: not the paper's abstract, the specific paragraph. This catches two things spellcheck never will: a claim that's technically sourced to the right paper but overstates or narrows what that paper actually found, and a claim that drifted to the wrong paper somewhere between your notes and your draft. It's the same check that catches fabricated citations, applied to citations that are real but have quietly come unmoored from what they actually support.

Where a tool built around this actually helps

The claim ledger above is, functionally, an extraction matrix: papers down one side, the thing you're extracting across the top, every cell holding the exact quote it came from. MyOwnArc builds that structure in directly rather than leaving you to maintain it in a separate spreadsheet, and because every synthesis answer is generated only from documents already in your library, the citation on a synthesised paragraph opens the specific passage it was drawn from, so the revision pass above is a click, not a re-read. It also pools every citation across your library into one ranked list, which tends to surface the paper your whole field keeps referencing that you haven't actually read yet, often the gap a supervisor's first round of feedback catches.