A better paper discussion

Prepare a journal club with AI: printable worksheet and reusable prompt

Print a journal-club worksheet, copy a source-checking AI prompt, and record the study design, denominators, uncertainty and questions for discussion.

The short answer

Start with one question, then use the blank worksheet and reusable prompt below. Check each important entry against the original paper before discussing it. The dated reading exercises follow the core method.

TRY THE WORKFLOW

Prepare your next journal club.

Bring a paper you have permission to use to AI Library, ask for a draft of its question, methods, results and limitations, then verify each point in the paper.

See plans

Start with the question the meeting needs to answer

A journal club can leave everyone with a polished summary and no clearer judgment about the paper. Before asking AI to read, finish this sentence: ‘After this discussion, we should understand whether this study changes our interpretation of [specific question].’ That gives the preparation a purpose and keeps the presentation from becoming a tour of every paragraph.

Choose one principal paper and gather its supplement, protocol or registration when available. Record the publication date and any correction. In MedSafeAI, keep your selected documents together in AI Library and use Evidence Search for the surrounding literature. A missing supplement should remain a visible gap, not an invitation for the model to reconstruct it.

A printable worksheet for the paper in front of you

Use this original MedSafeAI reading worksheet alongside the paper. Print the blank fields and add notes by hand, or use the prompts as a checklist in your own notes. Write a page, table, figure or section location for each factual entry. Mark unavailable information as “not reported in the supplied material”.

The fields draw on Cochrane’s data-collection guidance, linked in the sources. This original meeting aid is not a validated appraisal instrument or a form for diagnosis or treatment. Choose an appraisal method suited to the design; Cochrane’s risk-of-bias guidance below applies to randomized trials.

Download blank worksheet (PDF)

A 2-page A4 PDF to print and complete by hand. No information is entered on this page.

  1. Paper and materials

    Title, DOI or link, publication date and corrections. List the paper, supplement, protocol or registration actually available.

    Your notes
    Source location
  2. Question and population

    Write the discussion question, setting, eligibility criteria and population studied. Do not substitute your own patient for the study population.

    Your notes
    Source location
  3. Design and comparison

    Study design, recruitment or allocation, comparator and analysis population. Distinguish planned methods from the analysis reported.

    Your notes
    Source location
  4. Outcome, denominator and time

    Name one outcome. Record the unit counted, numerator and denominator when reported, analysis group and follow-up or measurement window.

    Your notes
    Source location
  5. Effect and uncertainty

    Copy the reported estimate, units and interval, if provided. Label absolute and relative measures; do not infer missing numbers or turn association into causation.

    Your notes
    Source location
  6. Limitations and missing information

    Separate limitations stated by the authors from your concerns. Note missing data, unavailable methods or supplements and any unresolved source conflict.

    Your notes
    Source location
  7. Authors’ conclusion and your interpretation

    Write these separately. State what the measured outcome can support and what it cannot establish.

    Your notes
    Source location
  8. Discussion and next reading

    List three questions for the group, one unresolved issue and the next source to inspect. Date the group’s final note.

    Your notes
    Source location

Copy the prompt and add your discussion question

Replace the bracketed question after copying, and provide only public papers or material you are permitted to use. Read the main results table yourself before accepting the worksheet: check the group, outcome, denominator and time window. An AI draft is preparation for discussion, not a verified conclusion.

Give disagreement a job in the meeting

Ask All AI can send the same preparation question to several models. Use Consensus to find differences in what they emphasize, then choose one difference worth investigating. For example: ‘One answer treats missing follow-up as central; another barely discusses it. What does the paper actually report?’ That is a productive discussion prompt, not a vote on which answer sounds best.

For randomized trials, Cochrane’s risk-of-bias guidance offers a structured framework for examining a specific result. Use the appropriate appraisal method for the study design; the worksheet here is a meeting aid, not a replacement for a formal assessment. Ask the group to explain its reasoning when it identifies a limitation.

Save the conclusion the group actually reached

End with three lines: what we learned, what we still cannot conclude, and what we will read next. Keep unresolved questions alongside the source locations and date the note. Your next journal club then begins with a useful record of the discussion rather than another untraceable AI summary.

17 October reading exercise: check what the summary left out

An early accepted npj Digital Medicine report published on 17 October studied nine clinicians and 33 consultations. Its abstract reports perceived time savings but only 63.6% complete summaries. This single-group pre–post study evaluated AI-FCP, not MedSafeAI; it does not establish clinical benefit. We reviewed the publisher’s abstract, not the full methods.

For the journal club, test omission handling with three short fictional research notes. A, dated day 1: “Measurement: 7 units; follow-up pending.” B, dated day 3: “Correction: 5 units replaces the value in A.” C, undated: “Measurement: 9 units.” These are invented teaching data, not patient records or results from the study. Keep your answer key below beside the generated summary.

Try this prompt: “Read notes A, B and C. List the documents you could read. For each measurement, give its document, date and exact supporting passage. Mark explicit corrections, unresolved conflicts and missing follow-up. Do not invent dates or decide that an undated record is the latest. Finish with an omission checklist.”

Run the exercise in MedSafeAI with the fictional notes or your permitted teaching references in AI Library. Compare the output with the answer key. If your plan includes Ask All AI and Consensus, inspect whether different answers preserve the correction and the uncertainty; agreement alone does not validate the summary. Record generation, review and correction time separately. This is a proposed exercise, not a test we have run.

CheckExpected entryFailure to catch
CorrectionB changes A from 7 to 5 units; preserve both source locations.Presenting 7 and 5 as two independent current measurements.
Undated recordC says 9 units, but its place in the timeline is unknown.Calling 9 the latest result or silently discarding C.
Missing follow-upA says follow-up is pending; no supplied note resolves it.Writing “follow-up normal” or hiding the gap.

15 October reading exercise: use is a different outcome from benefit

A JAMA study published on 15 October 2026 examined US clinician use of OpenEvidence between January 2024 and June 2026. It describes usage patterns, not a comparative clinical-effectiveness trial. Read the methods before treating popularity as evidence of better care.

Try this original journal-club exercise: imagine two people asking ten questions each. That is 20 questions, two users, and still no measurement of answer quality. Ask the group to rewrite a headline that confuses those units, then identify what additional measurement would answer its implied claim.

Suggested prompt: “From the supplied paper, extract the population, observation dates, unit counted and outcome actually measured. Give a source location for each. In a separate column, write what cannot be concluded from that measurement. Do not turn use, satisfaction or response speed into clinical benefit. Flag missing information.”

CheckYour worksheet entryDiscussion question
UnitPeople, questions or encounters; keep each denominator visible.Could repeated activity inflate our impression of reach?
OutcomeWrite the measured outcome in the authors’ terms, then a separate interpretation.Does the headline promise more than the measurement?
Next testChoose one additional outcome and a suitable comparison.What evidence would change our conclusion?

October 2026 example: do not let one reassuring number write the conclusion

A JAMA Pediatrics study published on 5 October 2026 evaluated a commercial pediatric sepsis alert retrospectively. The accessible abstract reports 36% sensitivity and a negative predictive value above 99%. We reviewed the abstract, not the full paper or supplement.

For your discussion, ask what each percentage counts. Negative predictive value is the proportion without sepsis among encounters without an alert; sensitivity is the proportion of sepsis cases detected by the alert. They answer different questions. Our exercise is to audit the AI summary before anyone turns it into a slide or a recommendation. It is not a MedSafeAI evaluation or a test of conversational models.

Copy this prompt: “Use only the supplied abstract or paper. Build four rows: case detection, false alarms, warning timing and subgroup performance. For each row, extract the metric, denominator, time window, uncertainty and exact source location. Write ‘not reported here’ for missing details. Then critique this proposed slide title: ‘A high negative predictive value means this alert rarely misses disease.’ Explain which denominator the title confuses. Do not infer a different threshold, an implementation decision or a clinical recommendation.”

Open MedSafeAI with the public abstract and run the prompt once. Check the result yourself, then use Ask All and Consensus to locate disagreements in interpretation. Save the corrected worksheet with the source in AI Library. For a deeper session, obtain the full paper and supplement through your usual access and list what they add; do not let a model invent inaccessible methods.

  • Keep the two denominators visible beside the percentages.
  • Separate a warning that occurs early enough to act from any warning during follow-up.
  • Leave the decision pending when the necessary threshold, subgroup or workflow details are missing.

Sources & further reading

Follow the original sources to check the details. Product pages describe their own products; they are not independent clinical evaluations.

  1. npj Digital Medicine: AI-FCP oncology preparation (17 October 2026; abstract reviewed) nature.com
  2. JAMA: clinician use of OpenEvidence (15 October 2026) jamanetwork.com
  3. Cochrane Handbook, chapter 5: collecting data (section 5.3 and table 5.3.a) cochrane.org
  4. Cochrane Handbook, chapter 8: assessing risk of bias in a randomized trial cochrane.org
  5. MedSafeAI: medical AI workspace and product features medsafeai.com
  6. JAMA Pediatrics: independent evaluation of a pediatric sepsis alert (5 October 2026; abstract reviewed) jamanetwork.com

For healthcare professionals. AI responses and agreement between models do not replace source verification or clinical judgment.

PUT IT INTO PRACTICE

Bring your next question to MedSafeAI.

Explore the models, compare perspectives and keep the evidence in view. Use MedSafeAI on the web, iPhone, iPad and Android.

See plans

Keep exploring

Journal