Your citation is real. Does it support your claim?

Writing · 4 min read · By Bibek Howlader, Founder, Labrix

You open the reference. The authors exist. The DOI works. The paper is relevant to your topic.

Then you read the result, and it does not say what your paragraph says.

The paper reports an association; your sentence describes a cause. The experiment measures performance on one dataset; your introduction promises performance in practice. A cautious finding in a small sample has become a general statement about people.

Checking whether a reference exists catches one kind of error. To check the argument, you have to put the sentence beside the evidence.

This problem existed before AI

In a 2009 analysis of a biomedical citation network, Steven Greenberg examined how a claim about beta amyloid and muscle injury in inclusion body myositis accumulated authority through subsequent publications. The analysis identified citation bias, amplification, and invention within that network. It is a case study of one claim, rather than an estimate of how often all researchers miscite sources.

The useful question for a writer is whether a familiar claim still means what its original evidence established.

AI adds another place where that meaning can change. Peters and Chin-Yee's 2025 study examined 4,900 summaries produced by ten language models and found that many broadened conclusions beyond the source texts. Those results concern the models and tasks tested in that study; they do not provide an error rate for today's research tools.

Both cases give us a reason to inspect the handoff from source to sentence.

Watch what happens to one finding

Here is an invented example for checking a draft. It is not a report of a published study.

Suppose a paper follows 180 students at one university. Students who voluntarily use a study app obtain higher examination scores. The researchers describe an association and note that motivation could influence both app use and scores.

Now consider three versions of a sentence citing that paper:

All three sentences could carry the same valid DOI. Only one closely describes the finding we invented.

You do not fix this by adding more references after the first sentence. First decide what the evidence allows you to say.

Audit the sentences that carry your argument

Start with the claims that matter most: the reason your project is needed, the finding your abstract highlights, and the conclusion a reader might act on. This is a way to begin an audit, not a substitute for reviewing the rest of the manuscript.

For each claim, answer five questions in a note:

1. What is the exact sentence? Copy it rather than summarising it from memory.
2. Where is the supporting evidence? Record a page, section, table, figure, or passage you can reopen.
3. What did the source actually examine? Include the study design, population or dataset, comparison, and outcome.
4. What boundary must survive the rewrite? Examples include follow-up length, uncertainty, geography, language, or laboratory conditions.
5. What is your decision? Keep the sentence, narrow it, find more evidence, or remove it.

“Related to the topic” is a useful label for a paper you want to read. It is too weak a reason to attach that paper to a consequential claim.

Separate evidence from your interpretation

A paragraph can contain both a reported finding and your own inference. Make the transition visible.

In our invented example, a defensible paragraph might say:

App use was associated with higher examination scores in this university sample. Because students chose whether to use the app, the study leaves open whether motivation explains some of the association. A randomised comparison could help investigate that possibility.

The first sentence reports the imagined result. The second identifies what the design cannot settle. The third proposes a way to learn more. Readers can inspect each step.

If the source is a review, ask whether you are citing its synthesis or a particular primary study it mentions. For a claim about the primary experiment, opening that experiment lets you check what was measured. Keep the review when its synthesis is the evidence you mean to use.

Keep a small claim record beside the draft

You can use this format in a document, spreadsheet, or research notebook:

Claim: The sentence in the manuscript.
Evidence location: The passage, figure, or table.
Support: Direct / partial / unclear / conflicting.
Scope: The conditions the evidence covers.
Action: Keep / qualify / investigate / remove.

The categories are working notes for the author, not a formal quality score. “Unclear” is a useful result: it tells you where another reading or a conversation with a co-author is needed.

In Labrix, you can bring source papers into your library and use them while drafting. Use that access to reopen the evidence when a sentence changes. A resolved reference and a fluent paragraph still need your judgment about support.

Try this with one paragraph today. Choose its strongest claim and find the exact place where the source earns that wording.

Have you ever opened a paper and discovered that a widely repeated claim was stronger than its evidence? Which part changed: the population, the outcome, or the certainty?