AI Literature Review Without Fake Citations: A 6-Step Workflow for Researchers

In short: Don't ask an AI to write a literature review from memory. Find the papers in a real database, give the model the papers themselves, have it extract and compare, and check every citation against the source before it goes into your draft. Below: six steps with a prompt for each, and the checks that catch what the model gets wrong.

The problem: the reference looks perfect and doesn't exist

You asked a chatbot for five references on your topic. Four look right: a plausible journal, authors who work in the field, a year, even a DOI. You paste the first title into PubMed and get nothing. The second exists, but the paper says something else.

This happens because a language model generates text that looks like a correct answer. It is not looking anything up. A reference is exactly the kind of text it can imitate convincingly: author names, a journal style and a title that sounds like the field. Newer tools that search the web or your uploaded files reduce this, but they do not remove it, so the workflow below does not depend on the model's memory at all.

Step 1: Pin down the question before you search

Vague questions produce vague reviews. Before you open a database, write the question in a structured form. The model is useful here because it asks the questions a reviewer would ask.

I'm a [role] working on [topic]. My draft question is: "[your question]".
Rewrite it as a focused, answerable question using a PICO-style structure where it fits (population or system, intervention or exposure, comparator, outcome).
Then list: (1) three terms I should define precisely, (2) synonyms and alternative spellings I should search for, (3) what is out of scope.
Do not suggest any papers or citations.

Step 2: Search in real databases, with the model building the query

Run the search in PubMed, Embase, Scopus, Web of Science, Semantic Scholar or Google Scholar, whichever your field uses. A model is good at turning your question and synonyms into a Boolean query with the right syntax for the database. It is not a substitute for running the search yourself.

Using the question and the terms below, write a PubMed search string with MeSH terms and free-text synonyms, combined with AND/OR and field tags.
Question: [paste]
Terms and synonyms: [paste]
Give me a broad version and a narrow version, and explain in one line what each one will miss.

Export the results (titles, abstracts, DOIs) into a reference manager such as Zotero, EndNote or Mendeley. From this point on, a paper is "real" only if it is in that library.

Step 3: Screen abstracts against criteria you wrote

Screening is where a model saves the most time, as long as the criteria are yours and the output is a recommendation rather than a decision. Paste a batch of abstracts and ask for a table.

Below are abstracts, each with an ID. Inclusion criteria: [list]. Exclusion criteria: [list].
For each abstract return a table row: ID | Include / Exclude / Unclear | the exact phrase from the abstract that supports your call.
If the abstract does not give enough information, answer "Unclear". Do not guess and do not use knowledge outside the abstract.
[paste abstracts]

Spot-check the "Exclude" rows in particular. A model that wrongly includes a paper costs you a few minutes. One that wrongly excludes it can cost you a finding.

Step 4: Extract from the full text into a table

Upload the full PDFs, or paste the relevant sections, into a tool that works from the documents you provide, such as NotebookLM, Claude with attached files or a ChatGPT project. Ask for structured extraction with a supporting quote for every cell. The quote requirement is what makes errors visible.

For each attached paper, fill one table row with these columns: Study design | Sample or model system | Key method | Main result (with numbers) | Stated limitations.
Next to each cell, add a short verbatim quote from the paper and the page or section it came from.
If the paper does not report something, write "Not reported". Never infer a value. Use only the attached documents.

Step 5: Synthesize from your table, not from the model's memory

Now ask for themes, agreements and contradictions, restricted to the table you just built. Ask the model to flag where studies differ in method, because that is often the real explanation for conflicting results.

Using only the extraction table above, write: (1) the main themes across studies, (2) where findings agree, (3) where they conflict and any difference in method or population that might explain it, (4) gaps no study addresses.
Cite each statement by study ID in brackets. If a statement isn't supported by the table, leave it out.

Step 6: Verify before anything goes into your draft

This step is yours, not the model's. It takes less time than it sounds, because steps 3 to 5 already tied every claim to a paper and a quote.

• Every citation resolves. Click the DOI or find the paper in your reference manager. If you cannot open it, it does not go in.
• Every quote is in the paper. Search the PDF for the quoted phrase. Paraphrases can drift, so check the numbers in particular.
• The claim matches the paper's claim. A model can cite a real paper for something it does not say, or turn "associated with" into "causes".
• Nothing important is missing. Compare against a recent review, the reference lists of your key papers and your own knowledge of the field.

What this workflow does not replace

It speeds up a narrative or scoping review. It is not a systematic review. If you need one, the protocol, dual screening and reporting standards such as PRISMA still apply, and any AI step should be documented in your methods and agreed with your co-authors.

Also check the policy of the journal you are targeting. Policies on AI assistance differ, and many require disclosure of how AI was used. AI tools cannot take authorship responsibility, so you remain accountable for every sentence. Tool features change quickly, so confirm what your version of each tool actually does with uploaded documents, and do not upload unpublished or confidential material to a tool your organization has not approved.

Keep going

The prompts above work better when each one states its context, task, format and check clearly. The template is in How to Write Good AI Prompts for Scientific Work. If you are deciding what is safe to upload, read Can Pharma and Biotech Teams Use ChatGPT and Claude Safely?.