Before an AI paper summary: check the PubMed record and PDF version

Start with the paper you actually want to read

You have a PDF in your downloads folder and a convincing AI summary. Before turning that summary into a journal-club slide, give the paper a simple identity card: title, first author, journal, year, DOI or PMID, file name and date checked. Copy identifiers from the source; leave a missing field blank.

Search PubMed using the title or citation details, then compare the returned record with the PDF. NLM’s user guide explains that PubMed holds citations and abstracts and links to available full text. Finding a record does not mean the full article was read. Label your input as abstract, article, supplement or notice.

Open the update notice, not just the original PDF

NLM links articles to published correction, retraction and other notices. Check both the PubMed record and the publisher page. A correction to a table may be linked without appearing inside the old PDF. If a Crossmark button is available, inspect its update information; Crossref says the button itself is not a guarantee.

For your reading note, record the notice type, date and affected section. If you see a retraction or an expression of concern, keep that status visible and read the notice before using the finding. If nothing appears, write where and when you checked; do not turn “no notice found” into a quality certificate.

Keep a small source-checking worksheet

Complete a separate worksheet for each document. An abstract and a full paper may belong to the same study but give you different material to inspect. This worksheet is our proposed reading method, not an automated certification feature.

A fictional example: the citation is right, the version is wrong

Imagine that an old PDF reports a sample of 240, while a linked correction changes the analysis population to 204. These numbers are invented for this exercise. The title and DOI still match, so checking only the citation would miss the difference.

Keep both files with explicit labels, such as “paper-original” and “correction-analysis-population.” Ask which statement in your draft needs revision. Then open the relevant passage yourself and update your note. Do not infer that every result changed, or that the correction has no consequence, without reading its scope.

Try the check with one document in MedSafeAI

Use Evidence Search when you need to investigate a reference, and open the original source yourself. For PDFs you are permitted to upload, AI Library lets you ask questions about your reference material. Keep the correction with the paper when it is relevant. These features support your review; this article does not claim that MedSafeAI automatically finds every notice or obtains inaccessible full text.

Copy this prompt: “Use only the supplied paper and any attached notice. First identify which files and sections you can read. Make an identity-and-version table. Then check this claim: [paste one sentence from my reading note]. Show the supporting section, table or page; classify the claim as supported, contradicted or not located. If a notice changes it, explain the specific change. Do not invent a PMID, DOI, missing text or page number. End with what I still need to inspect.”

Check the locations and identifiers in the answer against the documents. A neat table is useful only when its rows can be traced. If all you supplied was an abstract, keep the result labeled “abstract only” and leave methods or results outside it unresolved.

Save the checked note for the next reading session

Keep the claim, your corrected wording, the source location, the version/notice you consulted and the review date. Your first MedSafeAI exercise can be this single checked paragraph. Expand to several papers only after you can follow each important sentence back to its source. Feature access and usage limits depend on your current plan.