Trust & methodology

How ProofMode works

Most fact-checking tools are black boxes. We think that defeats the purpose. Here's exactly how ProofMode works, what each verdict means, and where our limitations are. Full transparency — because a fact-checker you can't verify is just another source of trust you have to take on faith.

The verification pipeline

InputExtract claimsFind evidenceClassifyReport
1

Input processing

You submit text, a URL, or a PDF. For URLs, we fetch the page and extract readable content. For PDFs, we parse the document. The text is then passed to claim extraction.

2

Claim extraction

An AI model reads your text and identifies every distinct, verifiable factual statement. Opinions, predictions, and subjective statements are filtered out. Each extracted claim is validated against the original text to prevent fabrication.

3

Evidence retrieval

For each claim, we generate optimized search queries and retrieve evidence from multiple web sources. Each claim gets up to 5 independent sources from different angles — not just one search.

4

Evidence-grounded classification

The classifier sees ONLY the retrieved evidence — never its own internal knowledge. It must cite specific sources in its reasoning. If no relevant evidence exists, the claim is marked "Insufficient Evidence" with 0% confidence. We never guess.

5

Trust Score & report

Results are compiled into a shareable report. The Trust Score is calculated only from verifiable claims — unverifiable claims are excluded so they don't inflate or deflate your score. Every claim shows its verdict, confidence, reasoning, and linked sources.

What the verdicts mean

Supported

Retrieved evidence directly and explicitly confirms this claim. The classifier cites which sources support it in its reasoning. Confidence reflects evidence strength — a single source caps at 70%.

Conflicting Sources

Multiple sources were found, but they genuinely disagree. Some evidence supports the claim while other evidence contradicts it. The reasoning explains which sources say what.

Refuted

Retrieved evidence directly and explicitly contradicts this claim. The reasoning cites which sources disprove it. This doesn't mean the claim is intentionally false — only that available evidence contradicts it.

Insufficient Evidence

We couldn't find enough relevant evidence to verify or refute this claim. Rather than guess, we flag it honestly. This is the default — a claim is "insufficient" until evidence proves otherwise.

About confidence scores

Confidence reflects how strongly the retrieved evidence supports the verdict — not an abstract AI "feeling." The scale is calibrated to evidence quality: 0% means no evidence, 10-30% means weak or indirect, 40-60% means moderate single-source, 70-80% means strong multi-source, and 90-100% means overwhelming agreement from multiple independent sources.

Importantly, confidence is capped at 70% when only a single source is available. We never report high confidence from thin evidence.

Our approach to transparency

Every verdict includes the reasoning behind it, citing specific sources by number. You can click through to every source, read the snippets, and decide for yourself whether you agree with our assessment.

ProofMode is a tool to help you think critically — not a replacement for it. We show our work so you can verify the verifier.

Known limitations

ProofMode is a powerful analytical tool, but it has limitations you should be aware of:

  • Claims are verified against web sources — niche, paywalled, or very recent claims may lack evidence and will be flagged as "Insufficient"
  • The AI classifier bases verdicts only on retrieved evidence, but can still misinterpret snippets, especially for nuanced or domain-specific claims
  • Search snippets shown in reports are summaries from the search engine, not direct quotes from the source page
  • Satire, sarcasm, and figurative language may be misinterpreted as factual claims
  • Evidence quality depends on what is available on the open web at the time of verification
  • ProofMode is a verification aid, not a replacement for professional fact-checking in legal, medical, or financial contexts

Your data

Verification jobs and reports are stored securely in our database. Your submissions are used only to produce your results — never sold or shared with third parties. Verification reports can be shared via public links, but only you can see them in your dashboard.

See it for yourself — paste any text and get a full verification report.

Try It Free