Literature review with AI: a step-by-step method

Step 1: define the question and the criteria

A literature review starts with a question written in one sentence. The question sets the scope of the search and the criteria for including a study. In health research, many teams use the PICO framework: population, intervention, comparison and outcome.

Write the inclusion and exclusion criteria before the search. They cover the period, the languages, the types of study and the populations studied. A study that fails one criterion is excluded, and the reason is recorded.

Check whether a review on the same question already exists. The University of Toronto Libraries list this check among the first steps of a knowledge synthesis. For a systematic review, the same guide recommends writing a protocol and registering it before the search begins.

An AI assistant helps at this stage. It suggests synonyms, related concepts and ways to narrow a question that is too broad. The researcher chooses the question and the criteria.

Step 2: search databases, registers and the web

A thorough search covers several types of source. Subject databases index peer-reviewed articles. Registers list ongoing and completed studies, such as clinical trials. Grey literature covers reports, theses and conference papers. Reference lists and citing papers lead to further studies.

The PRISMA 2020 statement sets the reporting standard for systematic reviews. It asks authors to name every database, register, website and other source they searched, with the date of the last search. It also asks for the full search strategy used in each database, register and website, with its filters and limits. Keeping this record from the first day makes the review reproducible.

AI is useful for building the search. It turns a question into keywords and synonyms. It writes search strings with AND, OR and quotation marks, and adapts them to the syntax of each database.

Run the search in the databases themselves. In a study published in 2024, ChatGPT retrieved fewer than one in seven of the studies that human authors had included in their systematic reviews.

Step 3: screen titles, abstracts and full texts

Screening happens in two rounds. The first round reads titles and abstracts and removes the studies that are clearly off topic. The second round reads the full text of the remaining studies against the criteria. The University of Toronto guide recommends a pilot test before each round.

Record the number of studies at each stage: found, duplicates removed, screened, excluded and included. The PRISMA flow diagram presents these numbers. For each study excluded at the full-text stage, note the reason.

PRISMA 2020 asks authors to state how many people screened each record, whether they worked independently and which automation tools they used. An AI tool can sort abstracts and flag the likely matches. A researcher confirms each inclusion and each exclusion, and the review names the tool that was used.

Step 4: read and extract the data

Extraction means recording the same information for every included study. A grid with one row per study and one column per item is enough. The usual items are the question, the method, the sample, the period, the main results and the limits stated by the authors.

Test the grid on a few studies before filling it in for all of them. Add the page number next to each value. A figure without a page cannot be checked later.

Assess the quality of each study with an appraisal checklist suited to its design. The assessment tells the reader how far each result can be trusted.

AI can fill the grid from the full text of each study. Each cell must then be compared with the paper, page by page. Under PRISMA 2020, the review states how many people collected the data and which automation tools they used.

Step 5: synthesise and cite

A synthesis compares the studies with one another. It groups them by theme, method or result. It shows where the results agree, where they diverge and how the methods might explain the difference. It states the strength of the evidence behind each conclusion.

Each sentence of the review cites the study it relies on. Before publication, run the searches again to add the studies published in the meantime.

AI writes a first draft of the synthesis quickly. The draft needs the same checks as the extraction grid. Each statement must match a study in the grid, and each study must say what the draft attributes to it. Many journals ask authors to disclose their use of AI. Check the policy of the journal you are submitting to.

Where AI gets literature reviews wrong

AI assistants make four kinds of error in a literature review. Researchers have measured each of them.

The first error is the invented reference. In a study published in Scientific Reports in 2023, William Walters and Esther Wilder asked ChatGPT for short literature reviews on 42 topics and checked the 636 citations produced. 55% of the citations from GPT-3.5 and 18% of those from GPT-4 were fabricated.

The second error is the wrong detail in a real reference. In the same study, 43% of the real citations from GPT-3.5 and 24% of those from GPT-4 contained substantive errors.

The third error is the incomplete search. A 2024 study in the Journal of Medical Internet Research compared the references suggested by AI with those of 11 systematic reviews. The hallucination rate was 39.6% for GPT-3.5, 28.6% for GPT-4 and 91.4% for Bard. The GPT models retrieved 11.9% and 13.7% of the relevant papers, and Bard retrieved none. The authors advise against using them as the main or only tool for a systematic review.

The fourth error is the distorted finding. In a 2025 study in Royal Society Open Science, Uwe Peters and Benjamin Chin-Yee compared 4,900 summaries written by ten language models with the original texts. Three of the models tested overgeneralised the results in 26% to 73% of cases. AI summaries were nearly five times more likely than summaries written by experts to contain broad generalisations. A diabetes study reported that a factor “was associated with” a lower risk, and one summary stated that it “significantly lowers” the risk. Asking the models for accuracy made the problem worse.

These studies tested versions that have since been updated. The Texas State University Libraries warn that made-up references can also appear in AI tools built for literature research. Researchers call this error an AI hallucination. The checks below apply to every tool.

How to check every reference

Check each reference before it enters the bibliography.

• Search for the exact title in quotation marks in a database or in Google Scholar.
• Open the DOI link and check that it leads to the same title, authors, journal and year.
• Read the passage the review relies on in the paper itself, and note the page.
• Compare each figure with the tables of the paper, including the units and the population.
• Check the tense and the scope of each claim: association or cause, population, period.
• Look for a correction or a retraction notice on the publisher’s site.
• Remove any reference you cannot find.

A quotation found only in an AI summary stays unverified until it is found in the paper.

A literature review with Verbatim

Verbatim gives you a team of AI agents for the whole literature review, from the first search to the final bibliography. You describe your question in plain words. The agents search, read, extract, check and write, for hours or for weeks. Every fact stays tied to its source, and you make every call.

The Research agent searches the web, search engines, public registers and the databases you name. It reads every study it finds from start to finish, methods and appendices included. The Reading agent takes up to 40,000 pages of PDFs, reports and notes, and answers each question with the exact page. Read more about research and reading.

The agents fill in the extraction grid for you. For each study, they record the question, the method, the sample, the period and the results. Each cell is linked to the excerpt it comes from, and a click opens the study at that page. When two studies contradict each other, Verbatim shows the disagreement and the differences in method.

The Checking agent compares every claim in the draft with its source and shows the exact excerpt. Each claim receives a status: verified, contradicted, outdated or inconclusive. A quotation found only in an AI summary is never counted as verified. Several agents work in parallel on the same project, each with its own role, and they check one another’s work. See how source verification works.

The Monitoring agent follows hundreds of sources day and night, including journals, preprints and the sites you add. It tells you when a new study bears on your question, with the passage that proves it. A corrected or retracted study is flagged in your project.

The agents then write the literature review, the synthesis and the bibliography in your style, learned from your published work. Your documents are hosted and processed in Europe and are never used to train a third-party AI. The page on Verbatim for research shows the full workflow. VerbatimAgent starts at €49.95 a month, VAT included, with a 14-day free trial.

Sources

• The BMJ, The PRISMA 2020 statement: an updated guideline for reporting systematic reviews (Page et al., 2021)
• University of Toronto Libraries, Knowledge syntheses: getting started
• Texas State University Libraries, AI research tools for literature reviews
• Scientific Reports, Fabrication and errors in the bibliographic citations generated by ChatGPT (Walters and Wilder, 2023)
• Journal of Medical Internet Research, Hallucination rates and reference accuracy of ChatGPT and Bard for systematic reviews (Chelli et al., 2024)
• Royal Society Open Science, Generalization bias in large language model summarization of scientific research (Peters and Chin-Yee, 2025)