From an AI answer to the paper: an evidence search workflow for doctors

Start with a question that evidence can answer

A broad request such as “tell me about this treatment” can mix mechanisms, trial results and recommendations into one answer. A more useful starting point identifies the population, the comparison and the outcome you want to understand. Add a date range when currency matters, and say whether you want a research overview or a comparison of guidelines.

MedSafeAI brings Evidence Search and citation Strict Mode into this workflow. Use them to work from a question toward identifiable sources, then inspect the link between each important claim and the material cited for it. A source attached to a sentence is something to examine, not the end of the review.

Request a claim-by-claim evidence map

Try this instruction: “Find evidence on [question] in [population]. Separate guidelines, systematic reviews and primary studies. For each important claim, give the source, publication date, population, outcome and main limitation. Distinguish information from an abstract from information checked in full text. Flag claims that lack a supporting source.”

Use citation Strict Mode and keep the question focused enough that you can open the decisive sources yourself. Ten loosely related references are less useful for this task than a small set that directly addresses the question. Do not mistake a request for a complete literature review for proof that the search is exhaustive.

Open the source behind the important claim

First check the article title, authors, year and identifier. Then read the relevant methods and results, using the full text when accessible. PubMed's user guide distinguishes its citation and abstract records from full articles, which are reached through links to publishers or PubMed Central. An abstract-only view should remain clearly identified as such.

Pay particular attention when the answer translates a narrow research finding into a broad statement. Check who was studied, what the comparison actually was and which outcome was measured. A source can be real and relevant without supporting every detail in the sentence beside it.

• Claim: what exactly does the AI answer say?
• Location: which passage, table or figure supports that claim?
• Fit: do the population, setting and outcome match your question?
• Limit: what qualification is lost if you quote only the conclusion?
• Status: are you reading the full paper, an abstract or another author's summary?

Separate the size of a finding from its certainty

A confident tone does not describe the strength of the evidence. When a review reports certainty, preserve that assessment and its explanation in your notes. Cochrane's handbook describes certainty in relation to a particular outcome and considers issues including bias, inconsistency, indirectness and imprecision.

Ask a focused follow-up: “Which limitations come from the study itself, and which come from applying its findings to my question? If sources disagree, describe the disagreement and the supporting material on each side.” This is more informative than asking the model to provide a single confidence score without an explanation.

Keep a note you can reuse

Finish with the exact question, a few supported claims, links to the sources you opened and the unresolved points. Keep publication dates and your reading date visible. You can then use the note for a journal club, teaching preparation or the next round of research. Start with one question in MedSafeAI's Evidence Search and inspect the claim that matters most. AI output supports your reading; clinical decisions still require professional assessment of the relevant evidence and circumstances.