How we check claims

Each page in the Evidence Library comes from the same pipeline that runs inside the EvidenceLens app. The process starts with a single claim, or with each factual claim extracted from a video, and restates it as a neutral, language-independent research statement. Next, semantic search compares that statement with our index of scientific publications, and the closest papers are read. In the last step, a language model weighs how much of the evidence supports the statement against how much contradicts it.

The pipeline turns this into a verdict, a confidence score from 0 to 100, the main reasons, the gaps that remain, and the papers that drove the decision. It writes the text you read on a page. Editors may soften or correct the wording, but they never change which evidence a page cites.

Our index holds more than 100 million scholarly works, and claims are matched to relevant research using publication metadata and abstracts. It includes published research and preprints. When results are retrieved, we filter out known non-research work types: datasets, supplementary materials, paratexts, errata and retraction notices. Standard searches also apply a publication-year filter.

Each page links to its papers by DOI so you can read the originals. We publish each paper's title, authors, year and study type, but never its abstract or full text.

Every page must clear the same publishing rules before it appears. Under those rules, a verdict has to use our fixed vocabulary, rest on enough scientific papers, contain no advice, and never be worded more strongly than the evidence.

Every checked-video page and report is read and approved by an editor. Claim pages reach publication by one of two routes, and the note at the bottom of each page shows which one applied. On the first route, an editor reads the page against the cited papers and approves it. On the second, the page is published automatically, but only once it has cleared the publishing rules, a set of evidence checks (enough papers, high confidence, no sign that the papers lean the other way), and an independent second assessment of the verdict. To pass, that second assessment must reach the same conclusion and find the wording not to be overstated. Instead of going out automatically, pages on topics where a wrong answer could cause harm, along with any page that fails a check, wait for an editor.

Our automated checks reduce the number of errors, though they cannot catch all of them. If you spot something on a page that looks wrong, tell us, and an editor will review it.

The review date is printed on every page. Whenever the underlying pipeline improves, the pages are regenerated, and each update goes through the same review again before it is published.

Every verdict rates how well published research supports a single, specific statement. When a statement is marked “Not confirmed”, the evidence does not support it as worded. This verdict does not accuse the person who made the statement of lying, and it does not exclude future findings. A “Not enough data” verdict should be read literally: the question has not been studied well enough to reach a conclusion.

Our verdicts on checked-video pages apply to the claims, not to the creators. We quote no more than one sentence per claim, just enough to help you find the moment in the original video, and we link to the video so you can watch it in context.

Nothing you read here counts as medical, legal or financial advice. Research findings come from specific populations and conditions, which may not match yours. For advice on your own situation, speak to a professional.

If a page misstates a paper, misquotes a video, or checks a claim that should not be here, email support@evidence-lens.com and include the page URL. We recheck the sources before correcting or removing the page, and any material correction is noted on the page itself.