What Is a Retracted Citation, and Why 'Verified' Doesn't Mean Safe to Cite

A citation can be entirely real and still be a problem

Our previous posts on this blog have all been about the same underlying question: does a citation point to what it claims to point to? A DOI can resolve to the wrong paper. A duplicate registration can impersonate a famous paper under its exact title. An LLM can invent a citation from pattern-matched memory. In every one of those cases, the fix is the same: check what's actually on the other end of the reference against what the citation claims.

Retraction breaks that framing entirely. A retracted paper's citation can be flawless — the right title, the right authors, the right year, a DOI that resolves to precisely the record it claims to. Every check built to catch a wrong or fabricated citation passes it, because nothing about the paper's identity is in question. The problem isn't that the citation is inaccurate; it's that the paper it accurately points to has been formally withdrawn by whoever published it, usually because something in it — the data, the methodology, the ethics of how it was produced — turned out not to hold up. A tool that only checks "does this citation match a real paper" has no way to catch that, because a retracted paper is, definitionally, a real paper.

What actually happens when a paper is retracted

A retraction is a formal, published notice from a journal or publisher stating that a specific paper should no longer be relied upon or cited as valid — issued for reasons ranging from data fabrication and image manipulation to honest errors that invalidate the conclusions, undisclosed conflicts of interest, or (increasingly) evidence of paper-mill involvement. It isn't the same as a paper simply being wrong or superseded by later research, which happens constantly and isn't grounds for retraction on its own. A retraction is a publisher's own statement that the record itself is unreliable.

Crucially, retraction doesn't erase the paper. The original stays indexed, keeps its DOI, and — this is the part that matters for citation checking — keeps showing up in ordinary database searches and reference managers, often with no visible flag unless you know to look for one. A paper can be retracted for years while still being cited by people who never saw the retraction notice, either because they found the paper through a source that doesn't surface retraction status, or because a downstream citation was copied from someone else's reference list without anyone checking the original.

How AccuraCite checks for it: three providers, none of which "certify clean"

When AccuraCite verifies a citation, it already queries OpenAlex, Crossref, and PubMed as three of its seven providers. Each of them separately publishes retraction status, but not in the same shape, so each has to be read differently:

• OpenAlex returns a straightforward is_retracted: true boolean directly on the work record — the simplest of the three to read.
• Crossref doesn't have a dedicated boolean field for this; instead it follows a convention where a retracted paper's own title gets rewritten with a "RETRACTED:" prefix, and the retraction notice itself — a separate record with its own DOI — carries an update-to relation pointing back at the DOI(s) it retracts. Both signals matter for different reasons, and we check the title-prefix pattern directly against every Crossref title we see.
• PubMed tags a retracted record with "Retracted Publication" inside its pubtype array, alongside its normal publication-type tags like "Journal Article."

None of these providers ever affirmatively says "confirmed not retracted" — they simply say nothing when a paper hasn't been flagged. That asymmetry shapes the whole design: retraction status can only ever be pushed from "unknown" to "retracted," by any one provider, never the other way. If OpenAlex says nothing but Crossref's title carries the RETRACTED: prefix, the citation is flagged retracted — the same consensus-backfill logic AccuraCite already uses to fill in a missing abstract or publisher field from a corroborating provider, applied here so that any single provider affirmatively flagging retraction wins, rather than requiring all three to agree before the flag is trusted.

The result is a retracted field on every /api/v1/verify response, set independently of whether the citation matched — so a citation can come back "verification_status": "verified" and "retracted": true at the same time. Collapsing those into one status would hide exactly the case this feature exists to catch: a citation that is completely correct and still shouldn't be used. The same check runs on /api/v1/generate — when AccuraCite is asked to find real papers supporting a claim, a retracted candidate is dropped before it's ever suggested, rather than being recommended alongside a note that it's since been withdrawn.

Real case 1: the paper that started the MMR-autism claim

In February 1998, The Lancet published Wakefield et al.'s paper proposing a link between the MMR vaccine, bowel disease, and autism in twelve children. It became one of the most consequential — and most thoroughly discredited — papers in modern medicine, cited for over a decade in vaccine-hesitancy arguments even after investigations found the data had been misrepresented and undisclosed financial conflicts existed. The Lancet fully retracted it in February 2010, twelve years after publication.

All three signals AccuraCite checks are independently present on this exact record today. Crossref's title for the original paper now reads "RETRACTED: Ileal-lymphoid-nodular hyperplasia, non-specific colitis, and pervasive developmental disorder in children" — you can confirm this yourself by querying Crossref's public API for the DOI and reading the title field directly. OpenAlex carries "is_retracted": true on the same work. And PubMed lists "Retracted Publication" in the pubtype array for PMID 9500320, alongside a RetractionIn reference pointing to the 2010 notice. Three independent organizations, three different technical conventions, one consistent answer — which is exactly the kind of cross-provider agreement AccuraCite's consensus logic is built to surface.

Real case 2: a live mass-retraction from August 2026

The Wakefield case is a single, famous, decades-old paper. To validate that retraction detection also holds up on something current and higher-volume, we checked it against a real batch retraction that happened while this feature was in testing. On August 5, 2026, SAGE issued a retraction notice covering 34 papers published in the International Journal of Knowledge-Based and Intelligent Engineering Systems — a journal it acquired from IOS Press — citing suspected third-party involvement, peer-review manipulation, citation manipulation, and "tortured phrases" consistent with paper-mill activity (Retraction Watch's coverage has the full context).

This case is a clean real-world illustration of the Crossref nuance mentioned above: the update-to relation lives on the retraction notice's record, not on any of the 34 original papers. Querying Crossref for the notice's own DOI returns an update-to array with all 34 entries, each pointing at one retracted paper's DOI with "type": "retraction". One of those, a 2017 paper on fuzzy aggregation operators, shows the same title-prefix pattern as the Wakefield case: Crossref's title field for it now reads "RETRACTED: Dual hesitant bipolar fuzzy aggregation operators in multiple attribute decision making", and OpenAlex independently confirms is_retracted: true on the same DOI. A tool that only reads the original paper's own Crossref record — rather than also checking whether some other record's update-to relation points at it — would miss this entirely, since the original record's title change is the only signal living on the paper itself; the structured relation lives one hop away.

Checking a paper's retraction status yourself, without a tool

If you want to spot-check one citation by hand rather than running a full verification, all three signals are publicly queryable. Crossref's https://api.crossref.org/works/<DOI> returns the title field — look for a RETRACTED: prefix. OpenAlex's https://api.openalex.org/works/doi:<DOI> returns is_retracted directly as a boolean. PubMed's E-utilities efetch endpoint for a given PMID returns a PublicationTypeList you can scan for "Retracted Publication". None of these require an API key for basic use. What they don't give you cheaply is corroboration — checking one source at a time doesn't tell you whether the others agree, and (as the Wakefield case shows) different providers encode the same fact in three structurally different ways, which is exactly the parsing work an automated check is for.

What retraction detection doesn't cover

A retracted: false result means none of the three providers has flagged the paper — it is not the same as an independent guarantee that the paper is sound. A paper under an active Expression of Concern, a formal but pre-retraction signal that something is under investigation, won't be flagged as retracted, since it hasn't been retracted yet. And a paper can be scientifically weak, superseded by better evidence, or subject to ongoing community criticism (visible on sites like PubPeer) without ever receiving a formal retraction from its publisher. Retraction detection answers "has this specific paper been formally withdrawn," not "is this paper still good science" — the second question doesn't have a database to query.

FAQ

Can a citation be "verified" and "retracted" at the same time? Yes, and that's the entire point of keeping the two fields separate. verification_status: verified means the citation accurately identifies a real paper. retracted: true means that real paper has been formally withdrawn. Both can be true simultaneously, and collapsing them into a single status would hide the more important of the two facts.

Is a retraction the same as an Expression of Concern? No. An Expression of Concern is a publisher flagging that something about a paper is under investigation, without yet concluding it should be withdrawn — it's a weaker, earlier-stage signal than a retraction, and AccuraCite's current check is specifically for formal retraction, not this intermediate state.

Does this mean AccuraCite recommends replacement sources for a retracted citation? For /api/v1/verify, a retracted citation is flagged, not silently replaced — the underlying reference is still real, so the right action depends on context (an editor may need the original for the retraction record; a student needs to remove it). For /api/v1/generate, which recommends sources for a claim, a retracted paper is excluded from suggestions entirely rather than surfaced with a caveat.

How current is the retraction data? As current as OpenAlex, Crossref, and PubMed's own retraction metadata, which is why cross-provider corroboration matters — a retraction processed quickly by one provider and not yet reflected in another is still caught, since any single affirmative flag is enough. See How It Works for the full verification pipeline this sits inside, or check a bibliography directly with the AI Citation Checker.