Can AI write a systematic review search strategy?

A systematic review search has to find nearly all the relevant studies, and it has to be written down so that someone else can repeat it. That makes it a hard test for a language model. It can write a Boolean string in seconds, and the string looks professional, but looking right and finding the studies are different things.

This guide covers what AI does well for searching, the errors to expect, a test that tells you whether a search works, and how to report it.

What a search strategy has to do

It combines the concepts of your question with Boolean operators, subject headings (MeSH in MEDLINE, Emtree in Embase) and free-text words, and it is run in several databases. For intervention reviews, Cochrane's standards expect CENTRAL and MEDLINE to be searched and, where the topic is within its scope, Embase. The search should favour sensitivity, which means it accepts retrieving too much rather than missing a study, and it should be reported in full: PRISMA 2020, item 7, asks for the complete strategy for every database.

What AI does well

Synonyms and spellings: it lists alternative terms, plurals, abbreviations and British and American spellings that you might miss.
A first draft: it can turn a PICO into a first Boolean string with the operators in the right places, which you then refine.
Translation: it can adapt a string from PubMed syntax to Ovid or another interface, which you check line by line.
Explaining: it can say what a field tag or a truncation symbol does.

Let SynthGap run the first pass

Enter an area. SynthGap proposes candidate topics, checks each against existing systematic reviews, and lists the primary studies behind it. Start on the Free plan: no card needed, and every run shows every topic it finds.

What goes wrong

Made-up or wrong subject headings: in a 2023 study of ChatGPT-written queries, incorrect MeSH terms were one of the main problems. Look up every heading in the MeSH browser or in Emtree.
Low recall: the same study found high precision but lower recall than the comparison methods, which is the wrong trade for a systematic review. A rapid review can accept that; a full one usually cannot.
Variability: the same prompt can return a different string on a different day, so keep the one you actually ran.
Database syntax: operators, proximity searching and field codes differ between interfaces, and a translated string can run without an error and still mean something else.
Dependence on the prompt: giving the model a good example search improved its results in that study, which shows how much the prompt matters.

A test that tells you whether it works

Collect a set of studies that you know are relevant, perhaps 10 to 20, from earlier reviews or key papers.
Run the search and check that every one of them is retrieved.
For each one that is missed, find out why: a missing term, a wrong heading, or a database that does not index it. Fix the search and run the test again.
Ask an information specialist to review the strategy. The PRESS guideline (Peer Review of Electronic Search Strategies) is a checklist for that.
Search the reference lists of the key papers and the studies that cite them, to catch what the strings miss.

Searching with an AI assistant instead of a database

Some AI research assistants find papers by meaning rather than by keywords. They are good for exploring a field and for finding candidate papers. They do not give you a search you can repeat, and you cannot see what they left out, so they do not replace searching the databases for a formal review. Use them to widen your list of terms and to find key papers, then run the database search.

Reporting

Report the final strategy for every database exactly as it was run, with the date. If AI helped to draft it, say so, with the tool and its version, and say how you checked it. See how to report AI use in a systematic review.

Sources

The figures and rules on this page come from these sources.

Wang S, Scells H, Koopman B, Zuccon G. Can ChatGPT write a good Boolean query for systematic review literature search? SIGIR 2023.
Page MJ, McKenzie JE, Bossuyt PM, et al. The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. BMJ 2021;372:n71.
McGowan J, Sampson M, Salzwedel DM, et al. PRESS Peer Review of Electronic Search Strategies: 2015 guideline statement. Journal of Clinical Epidemiology 2016;75:40-46.
Rethlefsen ML, Farrell AM, et al. Librarian co-authors correlated with higher quality reported search strategies in general internal medicine systematic reviews. Journal of Clinical Epidemiology 2015;68:617-626.
Cochrane. Methodological Expectations of Cochrane Intervention Reviews (MECIR), standards C24 (searching bibliographic databases).

Questions

Can ChatGPT do the searching itself?

A chatbot with no search tool only writes strings. One connected to a search tool may return results, but you cannot see what it did not return, and the search cannot be repeated exactly. Run the strings yourself in the databases.

Is it acceptable to use an AI-written search in a published review?

Yes, if you check it, test it against known studies, report the final strategy in full and declare that AI helped to draft it. The PRISMA 2020 checklist and many journals expect the complete strategy for each database.

Do I still need a librarian?

An information specialist remains the best check on a search, and a study of published reviews linked librarian co-authors to better reported search strategies. AI can speed up the first draft, but it does not replace the review of the strategy.

Which databases should I search?

It depends on the topic. For intervention reviews Cochrane expects CENTRAL, MEDLINE and, where the topic is within its scope, Embase, plus trial registers. Add subject databases such as CINAHL or PsycINFO where they are relevant, and say which ones you used.

Keep reading

General methodological guidance, not medical advice. SynthGap's suggestions are AI-generated from the published literature, so verify them before registering a protocol.