AI for Literature Review: Verifying Citations

A fabricated citation in a submitted manuscript is not a typo. It is a research-integrity finding, and it can end a career.

That risk is why this guide exists. AI genuinely helps with a literature review — it screens titles at a volume no human matches, fills an evidence table in minutes, finds the paper your search string missed. But the one thing every AI writing tool does well is produce text that looks like the text that belongs there, and a citation is the most predictable-looking text in all of academic writing. So the model produces one. Fluent, correctly formatted, plausibly authored — and pointing at nothing.

Everything below is built around one discipline: verify every reference, mechanically, before it reaches a draft.

The Challenge

A literature review has four jobs — finding, screening, extracting, citing — and AI is dangerous in exactly one of them.

The first three are volume problems: 3,400 records and a fortnight; screening, the least interesting and most expensive step and the one where a tired human misses things; forty survivors to turn into a comparable table.

The fourth is different. A model asked to support a claim has been trained on millions of examples of this pattern:

Author, A. B., & Author, C. D. (2019). Plausible title about the exact thing you asked about. Journal of the Right Field, 47(3), 221–238. https://doi.org/10.1016/j.jrf.2019.03.007

That is a very predictable pattern. Given a claim, the most likely next tokens are a surname that belongs in the field, a year in the right range, a title that describes your claim, and a real journal that publishes that sort of thing. The model is not looking anything up — it is completing a form. When the paper it is completing toward is famous enough to have been memorised, the form comes out right. When it is not, the form comes out anyway: identical in tone, identical in confidence, indistinguishable on the page. See hallucinations; this is that failure mode wearing its best suit.

And it survives review. A February 2026 analysis of NeurIPS 2025 found 100 fabricated citations across 53 accepted papers — roughly 1% of everything accepted — despite each paper passing 3–5 expert reviewers. Its conclusion: peer review contains no citation-verification step at all.

Outside academia the stakes say it louder. The US "Make America Healthy Again" report of May 2025 cited studies that did not exist; the epidemiologist listed as first author on one, Katherine Keyes, told NOTUS: "The paper cited is not a real paper that I or my colleagues were involved with." Some reference URLs still carried oaicite markers. In October 2025, Deloitte Australia partially refunded a government contract worth about A$440,000 over non-existent references and a fabricated quote from a Federal Court judgment. In Mata v. Avianca (2023), a US federal judge sanctioned two lawyers and their firm $5,000 for a brief built on ChatGPT-invented cases.

None of those people set out to commit fraud. They failed to check.

How AI Solves It

The most important sentence on this page:

Some AI tools retrieve from a real corpus. Others generate from memory. They look identical in the chat window, and the difference is whether the paper they cite exists.

Almost every guide to "AI for literature review" blurs this. Do not.

Generators

ChatGPT, Claude, Gemini and every other general chatbot, when answering from their weights, are generators. The reference leaves the language model the way any other sentence does: as the most probable continuation. The measured consequences are not subtle.

Walters and Wilder record the detail that makes this lethal: fabricated citations "tend to look legitimate at first glance." Only 2% invented the journal — the venue is real, the formatting perfect. And when they asked ChatGPT to check its own references, it answered inaccurately. The model cannot be used to audit the model.

Retrievers

Elicit, Consensus, SciSpace, Semantic Scholar and OpenAlex work the other way round. A query runs against a real index of real records — Semantic Scholar alone holds over 200 million papers, OpenAlex over 250 million — and the model reads and summarises what came back. This is retrieval-augmented generation, and it is the design decision that matters. The paper exists because it was found, not because it was written.

Retrieval fixes existence. It fixes nothing else.

Not fidelity. Peters and Chin-Yee, in Royal Society Open Science (2025), compared 4,900 LLM summaries against their source texts across ten models. Most produced conclusions broader than the source supported — DeepSeek, ChatGPT-4o and LLaMA 3.3 70B overgeneralised in 26–73% of cases, roughly five times more often than human summarisers. Prompting for accuracy did not fix it, and several newer models did worse than older ones. The paper is real; the claim you took from it may not be in it.

Not link integrity. Rao, Wong and Callison-Burch (2026) measured 221,000 citation URLs across ten models and deep-research agents: 3–13% were hallucinated — no record in the Wayback Machine, they likely never existed — and 5–18% did not resolve. Weigh this before buying a "deep research" subscription: agents generate substantially more citations per query than search-augmented LLMs, and hallucinate URLs at higher rates. More citations is not more evidence.

What AI is genuinely good at here

Strip out the citation generation and real wins remain: screening at volume as a second screener; extracting into a structured evidence table; finding what you missed, because semantic search surfaces the paper that uses different vocabulary for your concept and never matched your Boolean string; summarising a paper you already have, with the PDF in the context window, so the model reads rather than remembers; and criticising your own search strategy, the most underused move on the list.

Recommended Tools & Models

The only column that matters is the second one.

Pricing moves — check the vendor page before you buy. Elicit's tiers come from its own pricing page; the Consensus and SciSpace figures are advertised ranges and should be confirmed at source.

On models. For reading a paper you supplied, any current frontier model is competent. Prefer Claude Sonnet 5 or Claude Opus for close reading, on the strength of the Peters and Chin-Yee result that Claude overgeneralised least; Gemini 3.5 for a whole PDF stack at once; GPT-5.5 as a generalist. None of them should be asked to supply a reference.

Note what Elicit says about itself: it "does not yet know how to evaluate whether one paper is more trustworthy than another" and it "summarizes the findings of a bad study just like it summarizes the findings of a good study." That is a vendor telling you the truth, and it is worth more than any accuracy claim.

Step-by-Step Implementation

1. Decide what kind of review you are doing — this decides what AI may touch

A narrative or exploratory review you may write however you like, provided every citation is real. A systematic review is a method, and the method is the contribution. Read against PRISMA 2020 and the 2025 joint position statement on AI in evidence synthesis from Cochrane, the Campbell Collaboration, JBI and the Collaboration for Environmental Evidence, the boundary looks like this:

The position statement's own lines are the ones to keep: evidence synthesists are "ultimately responsible for their evidence synthesis, including the decision to use" AI; AI "should be used with human oversight"; "any use of AI or automation that makes or suggests judgements should be fully and transparently reported." The proposed PRISMA-trAIce checklist extension (JMIR AI, 2025) asks you to record, among other things, which exclusions were made by a human and which by a machine. If you cannot say which, you have not done a systematic review.

2. Build the search yourself, then have AI attack it

Write your Boolean strategy. Then paste it into a model and ask it to break it — what synonym is missing, what spelling variant, which discipline uses a different word for this construct, which papers the string would systematically exclude. Keep the suggestions that hold. You have used AI, and the strategy is still yours. That distinction is the whole game.

3. Screen — with AI as the second screener, never the only one

Before screening at scale, calibrate: take 150–200 records already screened by hand, run the model on them, compute its sensitivity against your own decisions. If it misses includes, your criteria are ambiguous or your prompt is.

The published picture is encouraging and not sufficient. A 2025 meta-analysis in the Journal of Medical Artificial Intelligence pooled 14 LLM screening models: sensitivity 0.812 (95% CI 0.617–0.920), specificity around 0.89. The spread underneath that pooled figure is the part to look at: individual included studies run as low as 42% sensitivity, and one small evaluation reached 18%. A pooled 81% means roughly one relevant paper in five is at risk of being dropped — which is exactly why the AI is a second screener, not a replacement for one.

Log every model-versus-human disagreement and resolve it by discussion. That log is your reportable agreement.

4. Extract into an evidence table, with anchors

Extraction is where AI earns its keep honestly. Elicit's column extraction and SciSpace's comparison tables are built for it; a model with the PDFs in context will do it too. The non-negotiable rule: every extracted cell carries a quote and a location. A table without anchors must be re-derived from scratch the first time a reviewer questions a number — and they will. Then spot-check ten random cells against the papers. If one of the ten is wrong, check all of them.

5. Verify every citation — mechanically

The step this whole page exists for. Do not ask the model whether its citation is real. Walters and Wilder found it answers that question inaccurately; the machinery that invented the reference will confidently confirm it. Instead, make the model produce a machine-checkable manifest, then check it with tools that cannot lie.

Now walk the manifest. Five checks, in order, all free:

1. Does the DOI resolve? Paste it into https://doi.org/<doi>. A DOI that 404s is a fabrication and you are done.
2. Does the resolved paper match the citation? The check everyone skips. The NeurIPS analysis names the failure mode — identifier hijacking, a real, resolving DOI attached to a fabricated citation, the primary mode in 4% of cases and a secondary characteristic in 29%. It defeats precisely the person who confirms the link works and stops. Compare title and first author against what came back.
3. Does the paper exist at all? For anything without a DOI, search the exact title in Semantic Scholar, OpenAlex, Crossref or PubMed. All free; the first two have free APIs, so a hundred references is a script, not an afternoon.
4. Do the authors match? The MAHA failure mode: a real researcher, in the right field, on a paper they never wrote. A title search returning nothing under that author is the tell.
5. Does the paper say what you claim? Open it. Find the sentence. This is what the 26–73% overgeneralisation rate makes non-optional, and it is the only check AI cannot do for you. While you are there, check the Retraction Watch database — free and public since Crossref acquired it in 2023, searchable by DOI.

Budget two to three minutes per reference. It is not optional, and it is not where the time savings live.

6. Disclose

• Elsevier requires a section titled "Declaration of generative AI and AI-assisted technologies in the manuscript preparation process" immediately above the references, using its template: "During the preparation of this work the author(s) used [NAME TOOL / SERVICE] in order to [REASON]. After using this tool/service, the author(s) reviewed and edited the content as needed and take(s) full responsibility for the content of the publication." It states explicitly that AI used to select, collate, generate or edit references must be noted. Grammar and spell-checking need no declaration.
• Springer Nature wants LLM use documented in the Methods section; AI-assisted copy editing does not need declaring.
• ICMJE requires disclosure at submission and forbids listing a chatbot as an author, "because they cannot be responsible for the accuracy, integrity, and originality of the work." Humans remain responsible for everything submitted, including "appropriate attribution of all quoted material, including full citations."
• For an evidence synthesis, the 2025 position statement expects tool names, versions, purpose, validation evidence and known limitations.

Nobody has ever been sanctioned for over-disclosing.

Real-World Examples

Example 1: The citation a chatbot built, and the check that catches it in ten seconds

Before. A reference supporting a claim about adolescent mental health:

Keyes, K. M., et al. Changes in mental health and substance use among US adolescents during the COVID-19 pandemic. JAMA Pediatrics — with a DOI attached.

Everything looks right. Katherine Keyes is a real, prominent Columbia epidemiologist who works on exactly this. JAMA Pediatrics is real and publishes exactly this. And there is a DOI, which is the thing most readers treat as proof of existence.

What is actually true. This citation appeared in the May 2025 MAHA report, and the paper does not exist. The DOI does not resolve. The journal issue it points into carries a different article by different authors. Keyes told NOTUS: "The paper cited is not a real paper that I or my colleagues were involved with." Seven cited sources in that report could not be found to exist, and several reference URLs carried oaicite markers — the fingerprint of a chatbot.

After. Check 1 catches it, and it takes ten seconds: paste the DOI into doi.org. It fails to resolve, and you are done — an identifier that does not resolve is a fabrication, and the reference never enters the manifest. Had the citation carried no DOI at all, check 3 would have caught it a minute later: the exact title returns nothing in Semantic Scholar or OpenAlex, and it is not in Keyes's author record.

What changed: nothing about the tool. Somebody clicked the DOI instead of trusting that it was there.

Example 2: The compound fabrication that survived expert peer review

Before. A reference in a paper accepted at NeurIPS 2025. Real authors in the field. A real, high-prestige venue. A title those authors could plausibly have written. And a DOI that resolves — to a page, with a paper on it. A reviewer clicks it. It loads. The reviewer moves on. Three to five expert reviewers did exactly that, per paper, across 53 accepted papers.

What is actually true. The DOI resolves to a different paper. The February 2026 taxonomy of those 100 fabricated citations found every one (100%) had compound failure modes — most often a total fabrication (66% of primaries) dressed in a semantic hallucination (63% of secondaries) and a hijacked identifier (29%). Its conclusion: these structures "exploit multiple verification heuristics simultaneously, explaining why peer review fails to detect them."

After. Check 2 catches it, and only check 2 does. Resolving a DOI is not verification. Comparing the resolved metadata to the citation is verification.

Example 3: The paper exists, and it does not say that

Before. A retrieval tool — a real one, searching a real index — returns a real paper and summarises it. (The paper below is a composite, written to show the shape of the failure; the placeholder author is deliberate.)

"Mindfulness-based stress reduction reduces anxiety in university students (Author et al., 2023)."

Every word of the citation is correct. The paper exists. You could resolve the DOI a hundred times.

What the paper actually reports: a single-site randomised trial of 42 undergraduates, an effect on self-reported anxiety at eight weeks that did not persist at six-month follow-up, and a discussion explicitly cautioning against generalising beyond the sampled population.

After. The claim as written is a general fact; the paper supports a bounded, temporary finding. This is exactly the gap Peters and Chin-Yee measured across 4,900 summaries — and prompting for accuracy did not close it. The rewrite:

"In a single-site randomised trial of 42 undergraduates, an eight-week MBSR programme reduced self-reported anxiety relative to a waitlist control; the effect was not sustained at six-month follow-up (Author et al., 2023)."

What changed: somebody opened the paper. That is the entire fix, and no tool performs it for you.

Industry-Specific Applications

Medicine and health. The most formalised and least forgiving: PROSPERO registration, PRISMA 2020, Cochrane or JBI methodology, ICMJE disclosure. The upside is PubMed — free, and a two-second existence check for anything biomedical. The Chelli finding, GPT-4 recall of 13.7% when asked to reproduce the references of real systematic reviews, is the number to quote to any colleague who thinks a chatbot can build a clinical evidence base.

Law. Mata v. Avianca generalises: a fabricated case has the same shape as a fabricated paper, and a court reacts like a journal. Verify against the reporter, not the model.

Social sciences and policy. The Campbell Collaboration co-signed the 2025 statement. Grey literature has no DOI, so the mechanical check does not apply and everything falls to title-and-issuer verification.

Environmental evidence. The Collaboration for Environmental Evidence also co-signed, and one of the more optimistic screening results — 100% recall from GPT-4 on a ~12,000-record task, Environmental Evidence (2025) — comes from this field. One result on one review is not a licence to drop your second screener.

Computer science and ML. Preprint-heavy, so arXiv IDs rather than DOIs are the real identifier — and, per the NeurIPS analysis, a field with a demonstrated fabricated-citation problem inside its own flagship venue. Do not assume you are immune because your field builds the models.

Librarians and research-methods teaching. The highest-leverage 60 seconds in any research-skills session: take a citation from a chatbot, paste the DOI into doi.org, watch it 404, and let the room sit with it. Nothing you say about AI research practice will land as hard.

Best Practices

• Search a real index. Read with a model. Never invert those.
• Verify mechanically, before a reference enters the draft — not before submission, when it is 200 references and you are tired.
• Never ask the model whether its own citation is real. It is the one question it cannot answer, and it will answer it anyway.
• A resolving DOI is not a verified citation. Compare the returned metadata to the reference.
• Calibrate the screener on records you have already screened, and report the agreement.
• Instruct the model to say UNCLEAR. Forbidden to express uncertainty, it converts it into false excludes, which are invisible.
• Anchor every extracted cell to a quote and a location, and check retraction status for anything load-bearing.
• Open the paper for any claim you rely on. Retrieval proves it exists; only reading proves it says that.
• Disclose per the target journal's policy, and log which tool did what — you will need it for the Methods section.
• Keep the search strategy human. It is the part of the method a stranger must be able to re-run.

Common Pitfalls

Believing the model when it confirms its own reference. The single most common way a fabricated citation survives. Confirmation costs the model nothing and means nothing.

Stopping at "the link works." Identifier hijacking exists, at scale, in accepted papers at the best venue in machine learning. The link working is check one of five.

Assuming a retrieval tool is complete. It searches its index, not the literature. Coverage varies by discipline, language and paywall. A retrieval tool that finds nothing has told you nothing about whether anything is there.

Assuming a retrieval tool judges quality. Elicit says it plainly: it "summarizes the findings of a bad study just like it summarizes the findings of a good study." Retrieval is not appraisal.

Buying "deep research" and trusting the citation count. Deep-research agents produce more citations per query and hallucinate URLs at higher rates than plain search-augmented models. A 40-source report is a bigger verification job, not a better-evidenced one.

Silent overgeneralisation. The most insidious failure, because nothing is fabricated. The paper is real, the citation is perfect, and your sentence claims something the study never established.

Letting AI touch the method. An eligibility criterion the model wrote; a search string you did not read; a synthesis it drafted. None is a formatting question. Each invalidates the review.

Citing a paper you have not read. AI makes this catastrophically easy: an evidence table hands you forty papers you can cite and have never opened. The citation is real. Your claim about it is a guess.

Not disclosing. Elsevier, Springer Nature and ICMJE all require it now — and undisclosed AI use is discoverable, not least from the fabricated reference that put an investigator on your paper in the first place.

Measuring Success

Fabrication rate in your own work: zero. Not "low." Zero. Measure it now — take your most recent submitted manuscript, resolve every DOI, check every author list against the resolved record. Whatever you find is your baseline, and it is the only number here that has to be perfect.

Screening sensitivity against your calibration set. Report it. Below ~90% sensitivity on records you hand-screened, the model is not fit to be your second screener on that review, whatever the pooled figure says. Track disagreements too: a rising rate mid-review usually means your criteria are ambiguous, not that the model got worse.

Papers found that your Boolean search missed — and papers your search found that the AI missed. The second number tells you whether the tool is a supplement or a liability.

Verification time per reference. Two to three minutes is the target. If it climbs, you are chasing references that had no DOI to begin with — itself a signal about where they came from.

Disclosure compliance. Binary. The declaration is in the manuscript, or it is not.

Cost Analysis

The verification stack costs nothing. Semantic Scholar, OpenAlex, Crossref, PubMed, doi.org and the Retraction Watch database are free, and the two largest have free APIs, so a hundred-reference bibliography resolves with a script rather than a hundred browser tabs. The most important part of this workflow is the part you cannot be charged for.

The reading stack is cheap. NotebookLM is free and restricts itself to documents you upload — exactly the constraint you want when the risk is invention. The free tiers of ChatGPT, Claude and Gemini will read a PDF you hand them. Paid consumer tiers, about $20/month, buy throughput, not accuracy.

The retrieval stack is where money goes. Elicit Pro at $49/month is the serious purchase if you are screening thousands of records. Consensus and SciSpace sit an order of magnitude lower and do less. For a single review, a month or two of one subscription is the whole budget.

The real cost is verification, in hours. Sixty references at three minutes each is three hours you must spend. AI has not removed that work — it has moved your time out of finding and screening and into checking. The trade is worth making; screening 3,400 abstracts costs far more than three hours. But pretending the second half is optional is exactly how you end up in the third paragraph of this guide. Deloitte Australia refunded about A$97,000 for skipping it. In academia the currency is not money.

Related Tools

• NotebookLM — grounds every answer in papers you uploaded and cites the passage; the safest way to read a stack of PDFs
• Claude — close reading of long papers; least prone to overgeneralising a finding
• ChatGPT — drafting and reasoning over sources you supply; never for supplying them
• Google Gemini — large context for whole-PDF stacks
• Perplexity AI — fast source-finding on the open web; web citations are not corpus citations
• You.com — search-grounded answers with visible sources
• Logseq — a local, linked store for the extraction table and the verification log
• Ollama and LM Studio — a local model when the papers are unpublished or confidential

Elicit, Consensus, SciSpace, Semantic Scholar and OpenAlex are not in this catalog; go to them directly.

Related Models

• Claude Sonnet 5 — careful summarisation; the family that overgeneralised least in the 2025 Royal Society Open Science study
• Claude Opus — heavy reading, where the nuance of a methods section matters
• GPT-5.5 — strong generalist for drafting around sources you have already verified
• Gemini 3.5 — the largest context, for many papers at once
• Claude Haiku 4.5 — cheap enough to screen thousands of abstracts

If the terminology is new, start with hallucinations, retrieval-augmented generation and information retrieval. The difference between a tool that retrieves and a tool that generates is not a technical curiosity in this use case — it is the difference between a bibliography and a liability. For the patterns behind the two templates above, see prompt engineering.