Skip to content

Command Palette

Search for a command to run...

BioTech

Research and literature with AI

AI can help you ask better questions and organise what you read. It cannot be trusted to find, quote or cite papers on its own.

Last reviewed . AI products change quickly; check the linked sources before relying on details.

← Back to the BioTech hub

Where AI helps in a literature workflow

  • Question refinement. Turn a rough idea into a PICO question (Population, Intervention, Comparator, Outcome) and list search terms and synonyms.
  • Discovery planning. Decide which databases and registries to search and how to phrase searches. The search itself runs in the database.
  • Summarisation. Summarise a paper you have supplied, with quoted support.
  • Evidence extraction. Pull design, population and outcomes into a table, marking anything not stated.
  • Organising a review. Group papers by theme, intervention or outcome.
  • Comparing findings and spotting gaps. Work from your verified table only.
  • Research briefs. Draft a short brief from verified evidence, with every statement linked to a source.
Production warning

Language models can generate references that look real but do not exist, or attach real citations to claims the paper does not make. Never paste a citation into a document until you have opened the source and confirmed it.

Know your source types

Tell the AI which source type each document is, and keep the label in your notes. A preprint, a guideline and a blog post carry very different weight.

Source typeExamplesHow to treat it
Primary researchOriginal studies, trial reportsStrongest for what was done and found; check design, size and bias
Systematic reviews and meta-analysesStructured syntheses of primary studiesUseful overview; check methods, date and included studies
Guidelines and regulator documentsSociety guidelines, agency guidanceAuthoritative for their scope; check version and date
PreprintsbioRxiv, medRxiv and similarNot peer reviewed; label clearly and look for the published version
Secondary summariesNews, blogs, AI summaries, slide decksLead to sources only; never cite as evidence

A source-verification workflow

The evidence chainA claim from an AI tool is unverified until each link is checked: the source is found, its identifier is matched, the claim is checked against the paper, and a checker initials the evidence table.AI claima statementor citationSource foundthe paperreally existsIdentifiermatcheddetails agreeClaim checkedpaper actuallysays thisInitialledchecker signsoff in table
A claim without a checked source stays a lead, not evidence.
  1. Find the original source yourself using a database or publisher site, not the AI's link.
  2. Confirm the paper exists: match title, authors, journal, year and identifier [PMID] or [DOI].
  3. Read the relevant section (methods, results tables), not only the abstract.
  4. Check each AI claim against the text and tables, including numbers, units and population.
  5. Check for retractions, corrections or a published version if it was a preprint.
  6. Record source, date accessed and who verified it.
  7. Mark anything unverified as unverified in your document.

Citation verification is the step people skip. Build it in as a required column in your evidence table: source found, identifier matched, claim checked, checker initials.

Limits to keep in mind

  • Models may not know recent publications and do not reliably know what they do not know.
  • Summaries can drop caveats, subgroup results and safety findings.
  • AI does not judge study quality or bias reliably. Appraisal remains a human skill.
  • Licences: check publisher terms before pasting full text into a service, and do not paste confidential or unpublished work into unapproved tools.
  • Search and screening for a formal systematic review need documented, reproducible methods. Do not use ad hoc AI output as that method.

Prompts for this module

Each prompt lists inputs, checks, data warnings and a reviewer. More are in the BioTech Prompt Vault.

Refine a research question (PICO)

Turn a vague question into a structured, searchable one and list what is missing.

Research analyst
Use it for
Scoping a literature search before you open PubMed or another database.
Who it is for
Research analysts, founders, medical writers
You will need
  • Your rough question
  • Condition or topic area
  • Any known constraints (population, time window, study type)
What you get back
A PICO breakdown, flagged assumptions, alternative framings and candidate search terms.
Human reviewer
Senior researcher or clinical scientist
Prompt
I am a [ROLE] preparing a literature search. My rough question is: [QUESTION].
Rewrite it as a PICO question (Population, Intervention, Comparator, Outcome). List any assumptions you made, any parts of the question that are ambiguous, and 3 alternative framings. Suggest search concepts and synonyms I could use in a database search. Do not cite any papers; this step is only about structuring the question.

Check the output, safety and limits

Verify

  • Check each assumption with a subject expert
  • Confirm search terms against a controlled vocabulary such as MeSH in the database itself

Safety and privacy

  • Use general descriptions of the topic. Do not include unpublished programme details.

Limitations

  • The model does not know which framing fits your regulatory or commercial context.

Learn more: Research and literature

Extract evidence from a paper you supply

Pull study design, population, outcomes and limitations into a table from text you paste.

Research analyst
Use it for
Building an evidence table from papers you have already found and are licensed to use.
Who it is for
Research analysts, medical writers
You will need
  • Paper text or abstract you are permitted to use
  • Fields to extract
  • Citation identifier [PMID] or [DOI]
What you get back
A table with a supporting quote per cell and NOT STATED where information is missing.
Human reviewer
Second analyst or subject-matter scientist
Prompt
Below is text from a published paper [PMID or DOI]. Extract only what is explicitly stated into a table with these columns: study design, population and sample size, intervention, comparator, primary outcome, key result as stated, stated limitations. If a field is not stated, write NOT STATED. Quote the exact sentence supporting each cell. Do not infer, calculate or add outside knowledge.

[PASTE TEXT]

Check the output, safety and limits

Verify

  • Open the paper and confirm each quote exists
  • Check numbers against tables and figures, not only the text
  • Check the sample size and outcome definitions yourself

Safety and privacy

  • Respect publisher licences before pasting full text into a service.
  • Do not paste confidential trial data, identifiable patient information or unpublished proprietary research into a service your organisation has not approved.

Limitations

  • Models can mis-read tables, units and subgroups.
  • Abstracts omit limitations that matter.

Learn more: Research and literature

Compare findings across studies

Line up results from several studies you have already extracted, and flag where they disagree.

Research analyst
Use it for
Spotting consistency, conflicts and evidence gaps in a set of verified sources.
Who it is for
Research analysts, medical writers, founders
You will need
  • Verified evidence table (from your own extraction)
  • The question being answered
What you get back
A comparison narrative, a list of disagreements and a list of gaps, with interpretation labelled.
Human reviewer
Senior researcher or biostatistician
Prompt
Using only the evidence table below, compare the studies for the question: [QUESTION]. Describe where results agree, where they differ, and possible reasons visible in the table (population, design, dose, follow-up). List evidence gaps. Do not add studies or facts that are not in the table. Label anything that is your interpretation as INTERPRETATION.

[PASTE VERIFIED TABLE]

Check the output, safety and limits

Verify

  • Cross-check each claim against the original table row
  • Have a scientist review the explanations for disagreement

Safety and privacy

  • Do not paste confidential trial data, identifiable patient information or unpublished proprietary research into a service your organisation has not approved.

Limitations

  • It cannot weigh study quality or bias reliably.
  • Statistical pooling is outside its safe use.

Learn more: Research and literature

Draft a plain-language summary of a published paper

Produce a first-draft lay summary of a paper you supply, flagged for expert review.

Medical writer
Use it for
Internal explainers or first drafts of lay summaries from already-published work.
Who it is for
Medical writers, communications staff
You will need
  • Published abstract or text
  • Target reading level
  • Citation [PMID] or [DOI]
What you get back
A short lay summary plus a list of points needing clinical review.
Human reviewer
Medical reviewer and your regulatory or compliance reviewer
Prompt
Write a plain-language summary (reading level: [LEVEL]) of the paper text below. Use only what the text states. Do not make claims about benefit, safety or effectiveness beyond what the authors report, and keep the authors' hedging words. End with a list of statements a clinician should check. Include the citation as [PMID/DOI].

[PASTE TEXT]

Check the output, safety and limits

Verify

  • Compare every claim with the source
  • Check that no claim of efficacy or safety is stronger than the source
  • Run your normal medical, legal and regulatory review

Safety and privacy

  • Anything that could reach patients or the public may be promotional or regulated content. Follow your company's review process.

Limitations

  • Simplification can change meaning.
  • Not suitable as approved patient material.

Learn more: Research and literature · Clinical and regulatory awareness

Frequently asked questions

Can AI find and cite scientific papers for me?

It cannot be trusted to find, quote or cite papers on its own. Language models can generate references that look real but do not exist, or attach real citations to claims the paper does not make. Find the original source yourself in a database, and confirm the title, authors, journal, year and identifier before using a citation.

How do I verify AI-generated literature summaries?

Follow the source-verification workflow. Find the original source yourself, confirm the paper exists, read the methods and results rather than only the abstract, and check each AI claim against the text and tables. Check for retractions or a published version, record who verified it, and mark anything unverified as unverified.

What is a PICO question and can AI help write one?

PICO stands for Population, Intervention, Comparator and Outcome. AI can help turn a rough idea into a PICO question and list search terms and synonyms. The search itself still runs in the database, and you confirm the terms there.

Can I cite a preprint as evidence?

Treat it carefully. Preprints on servers such as bioRxiv and medRxiv are not peer reviewed, so label them clearly and look for the published version. Secondary summaries such as news, blogs and AI summaries should lead you to sources only and are never cited as evidence.

Can AI do a systematic review?

Not on its own. Search and screening for a formal systematic review need documented, reproducible methods, and ad hoc AI output should not be used as that method. AI also does not judge study quality or bias reliably, so appraisal remains a human skill.

Where to go next