How to check a claim against the research

This guide is for students, researchers, writers and anyone who needs to cite evidence for a statement. It is part of our series on levels of evidence and study types, which explains how to rank the studies you find.

What does it mean for research to support a claim?

A claim is supported by research when studies that match its population, exposure, outcome and wording found a result in the same direction and of a meaningful size. "Research shows" is not one thing. A claim can be fully supported, partly supported (true for a narrower group or a smaller effect), not supported (studies found no effect or the opposite), or untested.

Claims drift as they travel. In a study of health press releases from 20 UK universities, Sumner and colleagues found that 40% contained exaggerated advice, 33% exaggerated causal claims and 36% exaggerated inference to humans from animal research, and that news stories tended to repeat those exaggerations [1] . So the version of a claim you meet is often stronger than the evidence behind it.

The four parts to match

How to check a claim against research in 6 steps

Here is the full method. Each step has a decision rule so you know when to move on.

1. Write the claim down exactly. Separate a compound sentence into single claims. "Green tea boosts metabolism and helps weight loss" is two claims.
2. Find the original source. Follow citations back until you reach the study or dataset that produced the result.
3. Read what the source actually says. Check the abstract, the results table and the limitations, not just the conclusion.
4. Search for studies that disagree. Run the search a second time phrased the other way.
5. Weigh the body of evidence. Rank studies by design, size and consistency, and look for a systematic review.
6. Write the verdict at the strength the evidence allows. Use calibrated wording and cite what you read.

The sections below take each step in turn.

Step 1 and 2: How do you find the source of a claim or statistic?

Trace the claim back until you reach a page that produced the number, not one that repeats it.

A primary source is the study, survey or dataset that generated a finding; a secondary source reports or summarizes it. You want the primary source, plus any systematic review that covers it.

How to find a source:

1. Search the exact wording. Put the statistic in quotation marks with two or three key words: "68% of students" sleep exam.
2. Follow every link back. News article to press release, press release to journal article. Stop at a page that cites nothing; that claim is unsourced.
3. Search a scholarly index directly. If no link exists, search for the topic in an academic search engine, PubMed for health topics, or your library's databases, filtered to the years around when the claim appeared.
4. Get the full text. Use your library's access, the publisher's open-access version, or a legal open copy found through Unpaywall [2].
5. Check the study still stands. Search for a retraction or correction notice. Crossref now carries the Retraction Watch database, so reference tools that use Crossref can flag retracted papers [3] .

A useful habit from media literacy is SIFT: Stop, Investigate the source, Find better coverage, Trace claims to the original context [4]. Steps 1 to 2 here are the "trace" move applied to research.

Step 3: How do you know if a study supports a claim?

Read the study against the four parts in the table above. The abstract's conclusion is the least reliable part for this purpose, because it is the authors' summary.

Read in this order:

1. Methods: who was studied, how many, what design.
2. Results table: the actual numbers, with confidence intervals.
3. Limitations: what the authors say they could not show.
4. Conclusion: only now, to see if it matches 1 to 3.

Worked example

Example (invented study and numbers, for illustration). The claim: "Drinking coffee cuts your risk of type 2 diabetes by 30%."

You trace it to a cohort study of 24,000 adults followed for 10 years. The results table shows a hazard ratio of 0.70 (95% CI 0.58 to 0.85) for 4 or more cups a day compared with none, adjusted for age, BMI and smoking.

Check the four parts:

• Population: adults in one country, aged 40 to 65. The claim says "you". Partial match.
• Exposure: 4+ cups a day. The claim implies any coffee. Mismatch: at 1 to 2 cups the ratio was 0.94, not significant.
• Outcome: diagnosed type 2 diabetes. Match.
• Result: an association in an observational study, not a tested cause. The claim says "cuts", which implies cause. Mismatch.

Verdict: partly supported. A fair version: "In one large cohort, adults who drank four or more cups of coffee a day had about 30% lower risk of type 2 diabetes than non-drinkers; the study shows an association, not cause."

Step 4: How do you look for studies that disagree?

Search on purpose for the opposite result. A search phrased as "coffee reduces diabetes risk" finds papers that say so. Run these as well:

• "coffee" AND "diabetes" AND ("no association" OR "increased risk")
• The same question with the outcome swapped for a related harm.
• Reviews only: add "systematic review" OR "meta-analysis".

Then check citation patterns. Greenberg's analysis of a citation network in biomedicine showed how a claim can gain authority through selective citation of supportive papers while contradicting ones are ignored [5]. If every paper you find cites the same two studies, you may be looking at one finding repeated, not many findings agreeing. Looking at papers that cite the original, a method called forward citation searching, is covered in citation snowballing.

Step 5: How do you judge the weight of evidence?

Rank what you found by design, then adjust for quality, size, consistency and directness. The GRADE approach rates certainty of evidence as high, moderate, low or very low, starting from study design and moving down for risk of bias, inconsistency, indirectness, imprecision and publication bias [6].

Decision rule: if a recent systematic review answers your exact question, start your verdict from it and use individual studies to explain details. If none exists, say so and describe the best available studies. For the full ranking with each study type explained, see the pillar on levels of evidence and study types.

Step 6: How do you write the verdict and cite evidence?

Match your words to the strength of the evidence. Readers should be able to tell from your sentence how certain the finding is.

When you cite evidence, cite the source you read, not the one that quoted it. If you read only a news story, either find and read the study or cite the news story as the source of the claim, not as proof of it.

Health, legal and safety claims: if a decision about care, treatment or risk depends on the answer, the evidence summary is a starting point for a qualified professional's review, not a basis for acting alone.

Checking claims in AI for Research

AI for Research has tools that speed up steps 2 to 5. They show you evidence and labels; the judgment in steps 3, 5 and 6 stays with you.

• Claim check (2 credits). Paste a paragraph. It splits the text into up to 10 claims, searches papers and PubMed for each, and labels each claim supported, partly supported, not supported or no evidence, listing the papers it used. Evidence strength is shown from the study types. Use it to see whether cited studies support a claim and which papers to open first.
• Citation check (0 credits). Paste a reference list. It checks each reference exists (Crossref by DOI, OpenAlex by title) and flags retractions, corrections, expressions of concern, preprints and wrong years or authors. An optional AI check (2 credits) compares each citing sentence with the paper's abstract.
• Find Papers (0 credits). Search for disagreeing studies with filters for study type, year, journal quality tier and citations; retracted papers are left out by default.
• Deep Research (from 1 credit). Ask the claim as a question. The answer cites real sources, and a consensus line such as "2 of 4 studies that test the question support it" is counted from a label on each cited study.
• Chat with PDF. Upload the study and ask "What population and dose did this study use?" Answers cite the page.

What Claim check cannot judge. It does not decide whether a claim is true. It reports what the papers it found say, and its search can miss studies, especially paywalled ones it can read only by abstract. It does not appraise how well each study was run. A "no evidence" label means it found no matching study, not that the claim is false. Open the papers before you rely on a label.

Free accounts get 45 credits a month, enough for about 20 claim checks. See pricing.

FAQ

How do you know if a study supports a claim?

Compare the claim with the study on four points: the population, the exposure or intervention, the outcome, and the size and direction of the result. A study supports the claim only if all four match. If the study was in mice and the claim is about people, or it measured a blood marker and the claim is about disease, it supports a weaker claim, not this one.

How do I find the original source of a statistic?

Search for the exact number in quotation marks with two or three key words, then follow each citation back until you reach a page that produced the number itself: a study, a survey report or an official dataset. Note the year, sample and definition used. If the trail ends at a page that cites nothing, treat the statistic as unsourced.

What is a reliable source for a research claim?

For a claim about what research shows, the most reliable source is a recent systematic review or meta-analysis that answers the same question, followed by well-designed primary studies such as randomized trials. News articles, press releases and blog posts can point you to evidence, but they are not the evidence. Check that the source is not retracted.

What should I do if studies disagree?

List the studies on each side with their design, size and population. Disagreement often shrinks once you notice that the studies asked slightly different questions or that the larger, better-controlled studies agree. If they still conflict, say so in your writing: 'the evidence is mixed' is an accurate finding, not a failure.

Can AI check whether research supports my claim?

AI tools can speed up the search and show which studies look relevant and how each bears on the claim. They cannot judge for you whether a study was well run or whether the evidence is strong enough for your purpose. Use the AI output as a reading list with labels, then open the key papers and make the judgment yourself.

Next steps

• Learn to rank what you find in levels of evidence and study types.
• Search health evidence properly with how to search PubMed.
• Find papers that cite or contradict a study with citation snowballing.

Sources

1. Sumner, P., et al. (2014). The association between exaggeration in health related science news and academic press releases: retrospective observational study. BMJ, 349, g7015. https://doi.org/10.1136/bmj.g7015
2. Unpaywall. About. https://unpaywall.org/
3. Crossref. Crossref acquires Retraction Watch data and opens it for the scientific community (2023). https://www.crossref.org/blog/news-crossref-and-retraction-watch/
4. Caulfield, M. (2019). SIFT (The Four Moves). Hapgood. https://hapgood.us/2019/06/19/sift-the-four-moves/
5. Greenberg, S. A. (2009). How citation distortions create unfounded authority: analysis of a citation network. BMJ, 339, b2680. https://doi.org/10.1136/bmj.b2680
6. Guyatt, G. H., et al. (2008). GRADE: an emerging consensus on rating quality of evidence and strength of recommendations. BMJ, 336(7650), 924-926. https://doi.org/10.1136/bmj.39489.470347.AD