Read papers with AI and build checkable evidence

Find the argument before summarizing

Map the question, claim, design, main result and limitations before deciding which details need attention. An abstract rarely suffices to audit sampling, exclusions and measurement. Separate what the authors claim, what the results directly show and what you infer.

Constrain AI to available material

Label materials S1, S2 with versions and locations. Ask AI to state what it can actually read and attach locations and minimal supporting passages to consequential claims. Missing support stays missing. Abstract access does not support full-method appraisal; verify citations in context yourself.

Build an evidence ledger

Use one row per claim: source ID, access scope, claim, location, design and sample, limitation and verification status. A preprint and its journal version may share data and are not independent evidence. Preserve versions and recheck affected claims after updates. Formatting a table does not verify it.

Move from summaries to synthesis

Compare questions, outcome definitions, designs, time windows and populations before explaining agreement. Two positive studies may measure satisfaction and delayed transfer respectively. Ask AI for possible reasons for differences and the evidence for each. Write your synthesis around relationships among studies rather than author-by-author summaries.

Worked example · fictional

From a problem to inspectable work

Fictional S1 abstract: 48 volunteers using a feedback tool scored higher on short answers than non-users. AI summarizes this as improved learning ability.

1. Return to the actual evidence: a score difference between groups in this sample.
2. Allocation, baseline ability, attrition and scoring are not established by this excerpt; mark them for full-text review.
3. Short-answer performance is narrower than general ability, and participation may reflect motivation. Inspect design before making causal claims.

Ledger row: S1 / abstract only / higher mean short-answer scores among users / supplied sentence / volunteer sample, allocation unknown / selection remains plausible; long-term transfer unmeasured / full text needed. Next, inspect allocation in the methods.

Failure case and repair

Pool unlike outcomes from AI summaries.

Measurement and bias are missing, undermining synthesis.

Extract from full text and assess comparability first.