AI for medical research should start from the clinical literature and show what kind of evidence each answer rests on. AI for Research searches PubMed and Europe PMC, labels every paper with its study type, leaves retracted papers out by default, and writes a cited answer with a consensus line. It is for clinical researchers and health scientists, not for patient care.
Three jobs you can do here
1. Ask a clinical research question against medical sources
Type a question into the Ask box, open Sources and choose Medical (PubMed and Europe PMC), and add Preprints for medRxiv if you want the newest work. Set your role to Clinician or Researcher. A Summary answer (1 credit) reads about 8 papers; a Deep report (15 credits) screens about 250 and reads 12 in full, with figures.
2. Keep only the study designs you trust
Open Filters and pick study types: meta-analysis, systematic review, randomized controlled trial, clinical trial, cohort, case-control, cross-sectional, case report and others. Add a year range, a journal quality tier, a minimum citation count or open access only. Leave out retracted papers is on by default, and preprints are flagged on every result.
3. See how far the evidence agrees
Each cited study gets a label for how it bears on your question, and the answer counts them in a consensus line: "3 of 5 studies that test the question support it, 1 is mixed." Use Claim check (2 credits) on a paragraph of your own draft to see whether published studies support each claim.
How it works, step by step
- Frame the question. Use PICO or PICOT; the Thesis Suite's Research question module (3 credits) can help.
- Ask with the Medical source, Summary depth, and study type set to meta-analysis and systematic review first.
- Read the Cited tab. Each source shows its study type, journal tier, citation count, and preprint or retracted flags. Open the full text where Unpaywall finds one; otherwise the answer is based on the abstract.
- Widen if needed. Remove the study type filter and run a Deep report to see trials and cohorts too.
- Check citations. Run the citation check (free) on your draft's reference list to flag retractions, corrections and expressions of concern.
- For a formal review, move to the Systematic Review workspace: PICO protocol, logged PubMed and Europe PMC searches, two-reviewer screening with Cohen's kappa, extraction with risk-of-bias fields, and a PRISMA 2020 diagram.
Example (invented): asked "Does early mobilization after hip fracture surgery reduce 30-day mortality?" with Medical and study types meta-analysis and randomized controlled trial, a Summary answer cited five studies, with the consensus line "3 of 5 studies that test the question support it, 1 is mixed, 1 does not." The table of cited studies showed one meta-analysis and four trials; two trials were under 100 patients each, which the line alone does not show.
The workspaces you will use
| Workspace | What you do with it | Credits |
|---|---|---|
| Deep Research | Cited answers from PubMed and Europe PMC | Summary 1, Deep 15, Detailed 45 (paid) |
| Find Papers | Search with study type and retraction filters | 0 |
| Citation checker | Claim check; retraction check on references | 2; 0 |
| Systematic Review | Protocol, screening, extraction, PRISMA | AI screening 1 per 25 records |
| Data Lab | Statistics on de-identified data | 0; AI write-up 2 |
| Proposal Studio | Ethics applications (IRB, UK REC, HREC) | Section 3, whole 15 |
Which plan fits, and a month worked out
The free plan (45 credits) covers occasional questions. A clinical researcher reading every week fits Plus: 750 credits for $12 a month or $96 a year, Detailed reports and unlimited PDFs. A systematic review on Free is limited to one review of 50 records, so reviews need a paid plan.
Example month (invented, on Plus): a registrar writing up an audit and starting a scoping review.
| Job | How many | Credits |
|---|---|---|
| Summary answers, Medical source | 20 | 20 |
| Deep reports | 3 | 45 |
| Claim checks on the discussion | 8 | 16 |
| Questions to PDFs | 20 | 20 |
| AI screening suggestions, 1,500 records | 60 batches of 25 | 60 |
| Protocol draft | 1 | 3 |
| Searches, filters, citation check, PRISMA diagram | any | 0 |
| Total | 164 of 750 |
AI-use rules that matter to you
- Results need a professional's review. Nothing here is medical advice. Treat every answer as a starting point for reading the studies, not a conclusion for practice.
- No identifiable patient data. Follow your institution's data protection rules and your ethics approval.
- Journals. ICMJE asks journals to require disclosure of AI use at submission, says chatbots cannot be authors, holds humans responsible for all content, and says AI output cannot be cited as a primary source [1]. Cite the studies, not the AI answer.
- Reviews. Say in your methods how AI screening suggestions were used; in this workspace reviewers decide every record, and the AI only suggests.
What it does not do
- It gives no diagnosis, treatment or dosing advice and is not a medical device.
- It does not search Embase, CINAHL or the Cochrane Library; import their exports for a review. ClinicalTrials.gov and openFDA are in Official Sources, but registries other than ClinicalTrials.gov are not searched.
- No meta-analysis or forest plot yet.
- Study type labels and the consensus line can be wrong. They count studies, not sample size or risk of bias.
Who else it is for
Supervising students? See AI for professors. Writing a doctorate? See AI tools for PhD students.
For the sources and checklists clinical research uses, see AI for clinical research: ClinicalTrials.gov and openFDA records, clinical research checklists such as the site initiation visit, and the reporting guideline checker for CONSORT, STROBE, SAGER and SAMPL.
Learn more
Sources
- International Committee of Medical Journal Editors. Recommendations: AI use by authors. https://www.icmje.org/recommendations/browse/artificial-intelligence/ai-use-by-authors.html
Last updated 2026-11-13