Most advice about AI and literature reviews is either uncritical enthusiasm or blanket warning. Neither is useful, because a literature review is not one task. It is six or seven, and AI is excellent at some of them and dangerous at others.
The dividing line is not tool quality. It is whether a stage tolerates a confident guess.
The principle worth holding onto
Language models produce plausible output. When you can immediately tell whether output is right, plausibility is a feature — it gets you to a usable draft fast, and you correct what is wrong.
When you cannot immediately tell, plausibility is the problem. A fabricated citation looks exactly like a real one. A confident summary of a paper you have not read looks exactly like an accurate one.
So: use AI where errors are visible. Do the work yourself where they are not.
Stage 1 — Scoping the question
AI helps here. Errors are immediately visible, because you are the domain expert.
Useful things to ask: what related fields might treat this problem differently, what terminology do adjacent disciplines use for the same concept, what are the obvious counterarguments to my framing.
The terminology point is genuinely valuable. Literature searches fail most often because you searched the vocabulary of your own subfield and missed three others studying the same thing under different names.
Stage 2 — Finding papers
Use search tools, not chatbots. This distinction matters more than any other on this page.
A search tool queries an index and returns records that exist. A chatbot generates text that resembles paper titles. The output can look identical. One is retrieval; the other is generation.
Semantic Scholar, Elicit, and Consensus query real corpora. ChatGPT asked to “find papers about X” without search enabled does not. Our tool finder sorts these by what each is built for.
Ask a chatbot for search terms. Ask a database for papers.
Stage 3 — Screening
AI helps, with a caveat. Deciding whether 200 abstracts are relevant is exactly the kind of high-volume pattern matching AI does well, and screening errors are recoverable — a wrongly excluded paper costs you a paper, not a false claim in your manuscript.
The caveat: exclusion is invisible. You will never notice the paper that should have been included and was not. Spot-check a sample of exclusions rather than trusting the filter wholesale, and write down your inclusion criteria before you start rather than after.
Stage 4 — Reading and extraction
Split this one.
Extracting stated facts — sample size, method, reported effect — is mechanical and verifiable. AI does it well and you can check any cell against the paper in seconds.
Judging what a paper means is different. Whether the methodology is sound, whether the conclusion is supported by the data, whether a limitation the authors mention in passing undermines the whole thing — that is the actual intellectual work of a literature review, and delegating it produces a review that reads fine and understands nothing.
A practical test: if you cannot argue with a paper, you have not read it.
Stage 5 — Synthesis
Partial help. AI is good at organising material you have already understood — grouping findings, spotting where two papers disagree, suggesting a structure.
It is unreliable at deciding why two findings conflict, which of two contradictory results is more credible, or what the gap in the literature actually is. Those judgements require knowing the field, and confident wrong answers here are hard to catch because they sound like insight.
Stage 6 — Citations
This is the one to be strict about.
Never accept an AI-supplied citation without checking it. Not because tools are bad, but because the failure is invisible and the cost is severe.
Two findings are worth knowing. When nine AI tools were tested on retracted papers, the best managed 53% accuracy and three research-specific tools scored zero. And when hallucinated citations were traced from preprint to publication, roughly 85% survived peer review — meaning nobody downstream will catch what you miss.
Two to three minutes per reference. Our checklist covers the sequence.
Stage 7 — Writing
Use it for revision, not generation.
Asking AI to tighten a paragraph you wrote is low-risk — you know what you meant, so you can see if the edit distorted it. This is particularly valuable if you are writing in a second language.
Asking it to write a section from your notes is higher-risk, because the output will smooth over exactly the uncertainties you needed to think through. Vague understanding produces confident prose, and you lose the signal that told you where your reasoning was thin.
Note also that some journals and, in the EU, transparency regulation now attach disclosure requirements to substantially AI-generated text. Check your target journal’s policy.
The summary version
| Stage | Verdict |
|---|---|
| Scoping the question | Use it — errors are obvious to you |
| Finding papers | Search tools yes, chatbots no |
| Screening | Use it, spot-check exclusions |
| Extracting stated facts | Use it, verify against source |
| Judging what a paper means | Do it yourself |
| Synthesis | Organising yes, judging no |
| Citations | Verify every one |
| Writing | Revision yes, generation cautiously |
The stages where AI helps most are the ones that were always tedious rather than difficult. The stages where it hurts are the ones that constitute the actual thinking.
That is a reasonable trade — as long as you are clear about which is which.
This is practical guidance rather than reporting. Specific findings cited link to our earlier articles, where the underlying sources are listed. Research-based rather than hands-on — see our Methodology page.