Literature review AI: which tools help, and at which step?

Direct answer: Yes, at rates that rule out trusting a suggestion on sight. Walters and Wilder (2023) found 55% of GPT-3.5 citations and 18% of GPT-4 citations fabricated. Chelli et al. (2024) recorded hallucination rates of 39.6%, 28.6% and 91.4% for GPT-3.5, GPT-4 and Bard. Check every paper that literature review AI suggests in a database.

Walters and Wilder (2023) asked ChatGPT to write short literature reviews on 42 topics, which produced 84 papers and 636 citations, then searched databases for each one. Among the real papers, 43% of GPT-3.5 citations and 24% of GPT-4 citations still contained substantive errors. The authors explain the cause in one line: "ChatGPT is fundamentally not an information-processing tool, but a language-processing tool" (Walters & Wilder, 2023). The wider failure has a name, hallucination.

Chelli et al. (2024) ran a harder test: 11 published systematic reviews, 33 prompts and 471 references. GPT-3.5 found 13 of the 109 papers the human reviews had included (11.9%), GPT-4 found 15 (13.7%), and Bard found none, while 91.4% of the 104 papers Bard proposed were hallucinated.

Worked example: what 20 suggestions are worth. Suppose a chatbot lists 20 references. At Walters and Wilder's rates, GPT-3.5 gives 20 × 0.55 = 11 fabricated papers, leaving 9 real ones, of which 9 × 0.43 = 3.9 carry errors, so about 5 are clean. GPT-4 gives 20 × 0.18 = 3.6 fabricated papers, leaving 16.4 real ones, of which 16.4 × 0.24 = 3.9 carry errors, so about 12 are clean. This is our arithmetic on their rates, not a result they report, and your prompts may behave differently.

Direct answer: Match each literature review AI tool to the step, and match your trust to the evidence behind it. Screening and extraction tools have published tests, mapping tools have mainly library guides, and no source we read tests AI synthesis. A tool's own description of itself is not an accuracy result.

Source: tool lists from the Cambridge and Bath library guides, evidence column from the studies cited here, read 31 October 2026.

Screening evidence comes from the developers. Ouzzani et al. (2016) beta-tested Rayyan on two Cochrane reviews with 1,030 and 273 records, and the larger one needed several iterations to find the 11 trials already known. Survey respondents reported 40% average time savings, and 34% reported more than 50%, which are user opinions and not stopwatch results. van de Schoot et al. (2021) report that "active learning can yield far more efficient reviewing than manual reviewing while providing high quality", based on simulations of their own ASReview software, which applies active learning. See how AI screening works in a systematic review for the stopping rules.

Direct answer: Literature review AI comes close to a human at extraction, but not equal. Hilkenmeier et al. (2025) compared Elicit with human reviewers across 43 studies and 602 data points: 81.4% accuracy for Elicit against 86.7% for humans, a difference that was not statistically significant. Use extraction as a second reviewer, not the first.

Hilkenmeier et al. (2025) scored both against a consensus-based ground truth in a review of psychological factors in dermatological conditions. In their words, "agreement between human and machine may serve as a reliable proxy for validity", because whenever Elicit and the human extracted the same information, it was correct in 100% of cases.

Converted to counts, 86.7% of 602 points is about 522 correct values for humans and 81.4% is about 490 for Elicit, so Elicit missed about 32 more values than people did. That is a gap you can live with only if a person still looks at every disagreement. One review, one field and a proof of concept do not show the same result in yours.

Worked example: one extraction row you can copy. Fix the columns before any tool touches a paper, then fill the same columns for every study.

A figure you cannot trace to a page of the paper does not belong in the table.

Direct answer: In a documented database that you can rerun, with the exact string and date saved. Gusenbauer and Haddaway (2020) tested 28 academic search systems and found that only half can be recommended for evidence syntheses without substantial caveats. Literature review AI search adds leads to that record and does not replace it.

Gusenbauer and Haddaway (2020) tested 28 systems and concluded that "only half of the search systems analyzed and only a few Open Access databases can be recommended for evidence syntheses without adding substantial caveats". The Cochrane Handbook (Lefebvre et al., 2025) says that CENTRAL and MEDLINE, with Embase where available, should be searched for all Cochrane reviews, and that "Searches should aim for high sensitivity, which may result in relatively low precision". The University of Cambridge library guide (n.d.) says its library had access to 1,767 databases when written, while University College London (n.d.) warns that AI search assistants "do not allow for comprehensive or systematic literature searching" (UCL guide).

Worked example: a three-block string. Cochrane advises few concepts and many OR terms inside each. For the question "Does retrieval practice improve exam results in undergraduates?", write: ("undergraduate*" OR "university student*" OR "college student*") AND ("retrieval practice" OR "practice test*" OR "testing effect") AND ("exam*" OR "test performance" OR "academic achievement"). Ask a chatbot for extra synonyms per block, then test each synonym in the database and keep it only if it changes the hits. Log the database, date, string and hit count every time. For the wider tool map, see AI and research tools across a whole project.

Direct answer: No source we read tests a tool that does, so plan to write it yourself. Synthesis means comparing studies and explaining why they differ, and that depends on your question. A fluent paragraph built on unverified summaries can look like synthesis while resting on nothing.

A synthesis paragraph makes one claim, sets studies side by side, explains the differences and states what follows. A model paragraph built from the audits above:

Two audits agree on direction but not on size. Walters and Wilder (2023) found 55% of GPT-3.5 and 18% of GPT-4 citations fabricated across 636 citations, whereas Chelli et al. (2024) found hallucination in 39.6% and 28.6% of 471 references. The GPT-4 gap, 18% against 28.6%, may reflect the task, since free-form reviews on 42 topics differ from replicating 11 systematic reviews with set criteria, and the two teams defined an error differently. Both conclude that references need checking, so any literature review AI output enters the review only after verification.

The method is claim, comparison, explanation, consequence, and you can reuse it on your own studies. Our guides on how to write a literature review and using AI ethically in a literature review cover structure, permission and disclosure. Reference managers such as Zotero format citations, which is a separate job; see which reference management software to use.