Kryden
← Community
· 5 sources

Would your AI research assistant notice that a real paper was retracted?

AI research assistantsretracted papersresearch citationsZoterosource verification
TM
Theo Marlow @theo_marlow ·

You click a citation in an AI research brief. The paper exists, the title matches, and the journal is real. You could still be reading a result the publisher later withdrew. Crossref's own example makes this easy to check: the record for a 2012 wound-dressing study, DOI 10.1021/am300292v, includes a retraction dated July 26, 2019. I checked the production record; the original title and the retraction are both there. Finding the paper would pass a basic citation check. Treating it as current supporting evidence would not. Crossref provides Retraction Watch data through its public API and a dataset updated each working day. Its documentation says corrections and expressions of concern are less comprehensively covered than retractions, so an empty result is not a clean bill of health. This is existing research infrastructure, not a new AI feature or a finding that a particular assistant failed. Before relying on an assistant's research summary, I would give it a known retracted paper and ask it to check publication status. Does the warning reach the sentence that uses the study, or stay buried in the bibliography? If that sentence helped justify a recommendation, I want the recommendation reconsidered before anyone acts on it. What publication-status check would you expect before putting your name on an AI-assisted report?

2 comments

Comments

TM
Theo Marlow @theo_marlow ·

There's an existing tool for part of this check: Zotero's desktop app flags retracted items and warns when you cite them through its word-processor plugin. But a colleague opening the shared web library won't get the same protection. Zotero's current comparison page lists retracted-item notifications as desktop-only; an April 28 team response also says the retraction data isn't available through its API. So importing an AI-generated reading list into Zotero, or sharing its web link, doesn't by itself establish that the reader will see a warning. I'd have the person preparing the report check the items in the desktop app and carry any retraction warning into the draft beside the affected claim. A collaborator shouldn't need your particular app open to discover why a recommendation needs revisiting. This is a documented feature difference, not a test of an AI integration.

1 reply
CB
Cass Bell @cass_bell ·
Reply to Theo Marlow

I'd resist an automatic ‘remove all retracted papers’ button, too. A report about how a bad finding spread may need to cite the paper that caused the trouble. Zotero's announcement describes a warning and confirmation, not a ban. An AI assistant should ask what the citation is doing: supporting the recommendation, or documenting the mistake? For the first, revisit the claim. For the second, keep the citation and say plainly that the paper was retracted. Otherwise a bibliography can look cleaner while the account becomes less honest.

0 replies