Find the evidence

Turn a question into a search you can run and refine in a real database. AI can suggest terms; it cannot replace the search or your assessment of the papers.

1. Discover papers with Consensus

A named tool / Consensus

Consensus searches research literature and can generate a cited synthesis. Use it to find candidate papers, then check each paper and its relevance yourself. Results can change between searches.

Standalone: open Consensus and sign in if prompted. Paste the question below into its search box and run the search.
Inside ChatGPT: if your account offers the Consensus plugin, open ChatGPT, find Consensus in Plugins, connect it and sign in to Consensus if prompted. In a new chat type @Consensus or select it with the + button, then paste the same question. Consensus setup instructions.
Open one suggested paper. In PubMed or the publisher record, check the title, authors, publication year, population and outcome. Read the original study before relying on an AI summary.

Consensus in ChatGPT works with its free plan according to the provider. Access, quotas and available features can vary by account, location and organisation. You can follow the prepared example without an account; signing up is not required for the workshop.

Copy: find relevant studies

What published studies have evaluated pharmacist-led antibiotic review or antibiotic time-out interventions for adult hospital inpatients? Identify relevant studies, give the title, authors, year and a link or identifier for each, and briefly describe the intervention, study design and reported outcome. Flag whether review or documentation within 72 hours was measured. Do not assume a study is an exact match if it uses a different outcome.

Check: a relevant-looking citation may measure something different from the fictional question. Source links do not guarantee that a summary is accurate or that the search is comprehensive. Consensus discovery complements the structured PubMed search; it does not replace a reproducible database strategy or independent screening.

No account or live access? Examine the real PubMed example and the deliberately altered citation already provided below. Identify the incorrect author and compare the study outcome with the fictional 72-hour review outcome.

2. Build a PubMed search

Consensus helps you discover possible papers. Now turn the same question into a search you can run and document in PubMed. You can ask an AI assistant to suggest the terms, or use the prepared search below.

Use the worked question or replace it with your own.
Copy the prompt and ask your AI tool for a draft strategy.
Test the strategy in PubMed; inspect its search details and results. Check suggested MeSH terms in the MeSH database.
Adjust one concept at a time, record the database, exact search, date and results, and ask a librarian for support when needed.

Copy: build a search

I am a hospital pharmacist developing this question: Among adult hospital antimicrobial treatment episodes, does introducing a pharmacist-led review process change the proportion with a review documented within 72 hours? Help me build a sensitive starting search for PubMed. Identify the main concepts and synonyms; suggest candidate MeSH terms, flagging that I must verify them in the MeSH database. Combine terms using OR within each concept and AND between concepts. Provide a PubMed search string using appropriate field tags, explain each concept, and suggest ways to broaden or narrow it. Do not invent citations. I will test and document the search myself.

Here's one we prepared earlier

This is a starting search for our case, not a completed systematic review. It may miss relevant records and should be refined with a librarian and other databases.

Click Copy search, then open PubMed.
Paste the complete search below, including its brackets and parentheses, into the PubMed search box and select Search. You do not need to type the quotation marks around the entire search.
Inspect the results and Search Details; change one concept at a time and record the database, exact search, date and result count. Check candidate terms in the MeSH database.

PubMed search

("Anti-Bacterial Agents"[MeSH Terms] OR antibiotic*[tiab] OR antimicrobial*[tiab]) AND ("Antimicrobial Stewardship"[MeSH Terms] OR stewardship[tiab] OR "pharmacist-led"[tiab]) AND (hospital*[tiab] OR inpatient*[tiab])

The search is deliberately broad: because "pharmacist-led" is an alternative within the stewardship group, results do not have to mention a pharmacist. To focus the results, test adding a separate pharmacist or pharmacy concept with AND, including suitable synonyms. Inspect the results for alternative terms such as "antibiotic review", and decide whether the hospital concept restricts the search too far.

What does the search syntax mean?

OR combines alternatives within a concept: a record can match any of them. AND connects the three concepts: a record must match at least one term in each group. The parentheses group the alternatives so the search is combined as intended.

The square brackets contain field tags, telling PubMed where to look. [MeSH Terms] searches its controlled subject headings; [tiab] searches the title and abstract fields, including other abstracts and author keywords. MeSH headings often include narrower headings automatically; check the heading in the MeSH database.

* is a wildcard: antibiotic* can match both antibiotic and antibiotics. Quotation marks around a phrase keep those words together. Explicit field tags and wildcards switch off PubMed's automatic term mapping for those terms, so inspect Search Details and test the results.

Some recent records have not yet received MeSH indexing, and some records are never MEDLINE-indexed. Combining headings with text words helps find material a MeSH-only search could miss. See the PubMed field-tag guidance and wildcard guidance.

3. Check a suggested citation

A record you found directly in PubMed exists as a PubMed record. This exercise tests something different: whether details or claims supplied by an AI assistant accurately match that record, and whether the paper actually answers your question. An assistant may mix up authors or outcomes, or give an incomplete or inaccurate citation.

For any AI-suggested citation, find its original record, compare its title, authors and identifier, then inspect the abstract or full paper for its population, intervention and outcomes. A record can be real but still be a poor match for your question.

Try a prepared source check

Imagine an AI assistant suggests the deliberately altered citation below. Open the real PubMed record and compare its first author and outcome with the suggestion and our question.

Paulson et al., “Impact of a Systematic Pharmacist-Initiated Antibiotic Time-Out Intervention for Hospitalized Adults” (PMID 33353452) is a real study. The abstract describes a pre–post study of a pharmacist-initiated antibiotic time-out and documentation of an antibiotic plan by 72 hours. It is relevant to our worked question, but its outcome is not identical to our fictional measure of a documented review; read the full methods and assess applicability before including it.

Prepared error for practice: “Smith et al., Impact of a Systematic Pharmacist-Initiated Antibiotic Time-Out Intervention for Hospitalized Adults, PMID 33353452.” This citation has been deliberately altered for teaching: follow the PMID and check the first author. It is Paulson, not Smith. A real identifier does not validate the rest of a citation.

Copy: screen supplied abstracts

I will paste published abstracts that I found myself. For each one, identify the population, intervention, study design and relevance to this question: Among adult hospital antimicrobial treatment episodes, does introducing a pharmacist-led review process change the proportion with a review documented within 72 hours? Quote the wording in the abstract that supports each judgement, note uncertainty and missing information, and do not infer results that are not reported. Suggest include, exclude or uncertain against these criteria: adult hospital inpatients, pharmacist-led antimicrobial review or time-out, and a reported review, documentation, prescribing or patient outcome. Exclude non-human studies; mark unclear cases uncertain. I will make and record the final screening decision independently.

ABSTRACTS: [PASTE PUBLICLY AVAILABLE ABSTRACTS HERE]

Check: abstract-based judgements can still be wrong or miss context. Verify against the source and review the full text when needed.

Explore other literature tools

Elicit offers research workflows that can search, summarise and compare papers. Its access and free usage are limited; an account is required. Elicit's workflow guide explains its options.

scite shows citation context that may help you explore how other papers refer to a study. Access depends on the available plan or trial. Context and labels can be wrong or incomplete, so read the citing paper rather than treating a label as a verdict. Explore scite's features.

PreviousGet started

NextEvaluate articles