AI Literature Review Tools: What They Actually Do, and Where They Invent Citations

The short answer

The AI tools worth using for a literature review are the ones grounded in real sources — either a searchable index of academic papers, or documents you upload yourself. Ungrounded chatbots are the wrong tool for finding literature: they produce references that look convincing and do not exist.

What none of them removes is the reading. They help you decide what to read and organise what you have read. You cite what you have checked yourself.

The distinction that decides everything

Almost every guide on this subject treats “AI” as one thing. That is why students keep arriving at supervision with fabricated references. There is a technical difference, and it determines which errors are even possible:

• Ungrounded language models generate the most probable next text. Ask for literature and you get something with the shape of a citation — author, year, a plausible title, sometimes a DOI. Whether the paper exists is not a category the model operates in. This is not a malfunction; it is the mechanism.
• Grounded tools retrieve first and write second. They search a real index, or answer only from the PDFs you gave them, and they point back at what they found. They cannot hand you a paper that does not exist, because they only discuss retrieved results.

The catch: grounded does not mean accurate. These tools do not invent papers, but they can misrepresent real ones — overstate a finding, drop a limitation, attribute a result to the wrong study. The error moves from “this source does not exist” to “that is not what it says”. The second is harder to notice and just as damaging when someone checks.

Three types of tool

Products change monthly; the categories do not. Place any recommended tool in one of these first:

1. Search tools over a paper index. Elicit is the best-known, and one of the fastest-rising search terms attached to thesis writing anywhere. You ask a question, it searches indexed papers and lays findings out in a comparison table. Strong for screening at volume; weaker for non-English scholarship and anything not in its index.
2. Tools over your own documents. NotebookLM and Anara answer only from files you upload, with pointers back to the passage. Useful when you have thirty PDFs and want to know who says what about one sub-question. They will not invent an external source — and will not find anything you did not give them.
3. General chatbots. Good for clarifying a concept, generating search terms, and explaining a difficult paper you already have. Not for finding literature.

What they are good at, and what they are not

• Good: screening. Reducing eighty results to the fifteen that actually touch your question. This genuinely saves days.
• Good: comparison. Who used which method, which sample, what result — as a table. That is the groundwork for your literature review.
• Good: comprehension. Having a dense paper explained before you read it properly.
• Good: search vocabulary.Finding the field's actual terminology so you can search databases with it.
• Bad: completeness.No tool covers your library's licensed holdings. A systematic review needs the subject databases.
• Bad: non-English scholarship. Noticeably thinner coverage, which matters enormously in some fields.
• Bad: evidence. An AI summary is never a citable source.

A workflow that holds up

1. Sharpen the question — without AI. What exactly do you want to know?
2. Build search vocabulary, chatbot help is fine here.
3. Search the subject databases systematically. This stays the foundation — see our guide to finding academic sources.
4. Cross-check with an index-based tool for anything you missed.
5. Collect the PDFs and load them into a tool that works over your own documents.
6. Screen and organise — who says what, where findings conflict.
7. Read the papers that matter. At minimum: method, results, limitations.
8. Write it yourself, citing only what you read.

Step 7 is the one people skip, and the one that shows. A literature review built from summaries of summaries reads smoothly and collapses at the first specific question — which is precisely the question a supervisor asks.

The thirty-second verification

Per source. It costs almost nothing and prevents the conversation nobody wants:

1. Search the exact title in Google Scholar or your library catalogue. No result means the paper probably does not exist.
2. Open the DOI directly. Fabricated DOIs resolve to nothing.
3. Check author and year. A common error is a real paper with the wrong year or the wrong authors attached.
4. Find the claim in the full text. The critical step with grounded tools: the paper is real, but the sentence attributed to it is not in it.

Is this allowed?

For searching and screening, usually yes — and treated far more permissively than using AI to produce text. But your institution decides that, not a guide.

Three things to settle before you start:

• Do your regulations permit AI tools, and for which steps?
• Must you declare the use? Many universities have added an AI clause to the standard declaration of authorship.
• Does your supervisor expect a note on your process?

Keep a record of which tool you used for what, as you go. It takes two minutes a week, it is what a declaration asks for, and it is the same record that protects you if the authorship of your work is ever questioned — which we wrote up in how to prove you wrote it yourself.