Literature reviews with AI tools: what they find and what they invent

1Where AI helps

Used carefully and where your course allows it, a chat tool can save time at the start of a review:

• Suggesting search terms and related concepts you had not thought of.
• Giving a rough map of the main debates in a field, as a starting point to check.
• Explaining a method or theory in plainer language before you read the original.
• Summarising a paper you have open, so you can decide whether to read it closely.

All of these treat the tool as a starting point. None of them replaces finding, reading and judging the sources yourself, which is what a literature review is marked on.

2Where it fails

The weaknesses matter most in exactly the parts of a review that carry the marks. USC Libraries' guide to generative AI lists several: tools may not have current information, they hallucinate facts and sources, they struggle with long structured writing, and their training data cannot be checked for bias.

A review also has to synthesise: to say where the sources agree, where they conflict, and what that means. A tool that has not read your sources cannot do that honestly, and it may present a confident summary of a debate that does not exist.

3Invented citations, and how to catch them

The best evidence on invented references comes from a study by William Walters and Esther Wilder, published in Scientific Reports in 2023. They asked two versions of ChatGPT to write short literature reviews on 42 topics and checked all 636 references against databases, publisher sites and library catalogues.

Two further findings matter for students. Most fabricated references named real journals and publishers, so they looked plausible. And errors in real references were mostly in the numbers: years, volumes, issues and pages. The study tested 2023 versions of the tool; newer tools may do better, but the authors' advice stands: check every citation it generates.

4A checking routine

1. Treat every AI-suggested reference as a search term, not a citation.
2. Find the source in your library catalogue or a scholarly database by its exact title.
3. Confirm the authors, year, journal, volume, issue and pages against the published version.
4. Open the DOI, if there is one, and make sure it leads to the same work.
5. Read the source. Only cite what it actually says.

Record each checked source as you go. An annotated bibliography is the simplest way to keep that record.

5What to disclose

Rules differ, so read your course's policy. At UCL, for example, the guidance says students must acknowledge how they used generative AI in assessed work, but do not cite it as a source, because it is not an author and does not give a stable answer. Other universities ask for a short statement or an appendix of prompts. If you used a tool to find or summarise literature, say so in the form your course asks for.

For a longer project, read AI in a dissertation, chapter by chapter. For everything else on AI and writing, start from the AI and writing hub.

6Our method

We wrote this page from the 2023 study by Walters and Wilder in Scientific Reports (we use its published figures only), USC Libraries' guide to generative AI, UCL's assessment guidance and UNC's literature review handout, each read at the source on 9 November 2026 and listed under Sources below. Every rule or figure on the page comes from one of them; advice that is our own judgement is written as advice, and examples written for this page are labelled as invented. When we update the page we re-check each source, and the date at the top shows the last check.

7FAQ