How to check if a citation supports a claim

A citation can point to a real paper and still fail to support the sentence. To check support, compare what the sentence claims with what the cited passage actually says.

A reference lookup answers whether a paper exists. Reading the passage answers whether the paper backs the number, population, or conclusion in your sentence. Those are different checks, and a manuscript needs both.

A working DOI confirms that a reference resolves, not that its findings support your wording.

Reference checks catch fabricated papers

An October 2026 audit examined 500,000 science, engineering, and computing papers. About one in 435 papers from January through March 2026 contained a suspected fabricated reference, according to Columbia University's account of the audit.

The audit screened references with DOIs and flagged titles it could not locate. Experts confirmed fabrication in 85.8% of a sample of 400 flagged references. A failed lookup therefore warrants investigation, rather than an automatic accusation.

Publishers are adding checks for this problem. The British Journal of General Practice announced in October that it would screen accepted papers for fabricated references before publishing author accepted manuscripts.

In Benjamin Skuse's experiment for IEEE Spectrum, STM Integrity Hub flagged 12 hallucinated references in a deliberately fabricated paper. A second checker could not validate 24. Those are two different findings, not interchangeable counts of proven fabrication.

These checks address whether references can be found. A real paper attached to an exaggerated sentence requires a comparison of meaning.

A claim can change while its citation stays the same

Consider an invented example. A study reports an association between sleep duration and exam scores among students at one university. A draft cites that study for the sentence, “Longer sleep improves academic performance in young adults.”

The reference is real within this example, but the sentence makes two additional claims. “Improves” implies that sleep caused the difference. “Young adults” extends the finding beyond the students studied.

A useful support check names those differences. “Citation verified” leaves the researcher without a reason to revise either phrase.

The comparison needs more than matching a topic. Read the sentence and passage for these specific points:

The useful output is the unsupported phrase and the passage that explains the mismatch. A sentence that combines several findings may need to be split before its citations can be checked clearly.

How Agent Bayes compares a claim with its citations

Agent Bayes keeps research claims in a mindmap, a tree of short statements with citations attached to the statements they support. AI Verification checks a selected statement against those attached citations.

The implementation passes the statement's text and the retrieved citation content to a verification model. The model has no search tools in this operation. It judges the evidence supplied with the statement, rather than searching the library for a better source.

The result concerns the passages attached to that statement. If the attached passages do not support a phrase, the question is whether to revise the phrase or attach more appropriate evidence. A source elsewhere in the library cannot silently make the existing citation adequate.

The model returns a support score, a list of discrepancies, and four alternative phrasings. The application adds those results beneath the original statement. It leaves the original wording unchanged and retains its citations on the alternative phrasings.

The score describes alignment with the passages

The model uses a written rubric from 0 to 1. The interface displays the result out of 100. The rubric asks the model to assign these meanings:

These are instructions for a model's judgment. A displayed score of 90 does not mean a 90% probability that the statement is true. The check assesses alignment with supplied text, not study quality or the strength of the wider literature.

When a statement has no attached citations, the prompt instructs the model to return zero and explain that citation support cannot be verified. An uncited claim may be correct, but this operation has no evidence to assess it.

Discrepancies identify what to read again

The prompt asks for specific unsupported claims or phrases. It also instructs the model to report only problems observable in the supplied evidence and to return an empty discrepancy list when the passages fully support the statement.

This matters when a sentence uses connective wording such as “in contrast to the above.” The phrase relates statements within a mindmap. The verifier is instructed to assess the substantive claims without treating that connective wording as missing evidence.

Read the discrepancy before deciding what a score means for your draft. An unsupported population label calls for a different revision from a number attributed to the wrong study.

Alternative wording remains a proposal

The verifier offers formal, concise, accessible, and assertive versions. Its instructions require those versions to stay within citation support while preserving meaning and scope. The original statement remains available beside the results.

For an overstated claim, read any proposed wording against the passage before adopting it. An assertive tone cannot justify removing uncertainty that the source requires. A fluent rewrite is still a sentence whose evidence needs to match.

Checking attached citations differs from searching a library

A manuscript may name a source without attaching the relevant passage. In that situation, the first task is to find evidence. Agent Bayes uses a separate operation, AI Support Check, to search the project's libraries and assess what it finds.

Our paper review example used that library search operation on a two-page draft with seven deliberately planted citation mistakes. The check flagged all seven, including a reversed finding and a claim attributed to the wrong paper.

Those seven results describe AI Support Check’s library search workflow on a known set of mistakes. AI Verification instead compares wording against passages already attached to a claim, using the rubric above.

The distinction helps when choosing the next action. Use a library search to locate evidence for a manuscript claim. Use verification to compare a statement's wording with the passages already attached to it.

Use the result to decide which passage to open

A support score helps prioritize reading. The discrepancy identifies the wording to inspect, and the citation leads to the source. Opening that source lets you check whether the model interpreted the passage correctly.

Pay particular attention to attribution. An author may quote a position in order to reject it. A passage can also omit qualifications explained elsewhere in the paper. Read the surrounding discussion when the excerpt does not settle the claim.

The source passage, rather than the score alone, gives you a reason to change a sentence. For the sleep example, that reason is concrete: the passage reports an association in one student sample, so the sentence should preserve both limits.

In an Agent Bayes mindmap, select a cited claim and press F3 for AI Verification. Open a flagged citation and compare the wording with its passage. Each check holds an estimate of two credits, then charges for actual usage. Edit the claim once the passage supports your revision.