How TruthPrism Works

Summarize Claude AI

Claude reads your article and produces a concise executive summary capturing the main argument and key assertions. It then extracts 3–6 of the most consequential factual claims — calibrated to the article's complexity.

Simple articles (news reports, opinion pieces) → 3–4 claims
Complex articles (investigative, analytical) → 5–6 claims

Claims are selected for being factually verifiable — specific enough to be checked against real-world evidence, not just interpretive or analytical.

Verify Claims Brave Search

All claims are verified simultaneously using real-time Brave Search — an independent web index that's not Google or Bing. For each claim, Claude generates a targeted 5–8 word search query, retrieves current web results, and evaluates the evidence.

Why Brave Search? Unlike Anthropic's built-in search tool, Brave Search is an independent index with its own crawler — reducing SEO spam and providing fresher, less biased results. It's purpose-built for AI applications.

Parallel verification — all claims are searched simultaneously, which is why Step 2 takes only 15–20 seconds even for 5–6 claims.

Evaluate Context Claude AI

Beyond factual accuracy, Claude evaluates how the article frames its information. This produces the Context Score and the Context & Framing section — the part of TruthPrism that goes beyond simple fact-checking.

What Context evaluation looks for:
• Charged terms used as established fact rather than attributed opinion
• Sweeping motivational claims about people's intentions without sourcing
• Missing counterarguments or alternative perspectives
• One-sided framing that's inappropriate for the article's format

Note: Opinion pieces are scored differently — one-sided framing is expected and appropriate for the format.

The Two Scores

An article can score well on one and poorly on the other — they measure different things.

Factual Score

Rates the accuracy of verifiable factual claims — dates, statistics, quotes, attributions, and assertions checked against real-time web sources.

8–10Claims well-supported by credible sources

5–7Mixed — some accurate, some questionable

1–4Significant false or unverified claims

Context Score

Rates whether the article gives you the full picture — fair framing, appropriate attribution of charged language, and acknowledgment of counterarguments.

8–10Balanced, fair framing, no significant omissions

5–7Notable framing issues affecting interpretation

1–4Highly misleading framing or major omissions

Claim Verdicts

Each verified claim receives one of five verdicts based on real-time evidence from Brave Search.

SupportedMultiple credible sources confirm the claim. High confidence the assertion is accurate.

Partially SupportedClaim is roughly accurate but with important caveats, missing context, or minor inaccuracies.

UnverifiedSearch results don't address the claim directly. Treated as neutral — absence of evidence is not evidence of falsity.

ContestedSources actively disagree with the claim or present significant counter-evidence.

FalseCredible sources directly contradict the claim with clear evidence.

A Note on Consistency

TruthPrism's results are highly consistent across runs on the same article. In 10-run testing on identical articles, scores were identical across every run — a significant improvement over earlier versions that used Anthropic's built-in search tool.

This consistency comes from Brave Search's deterministic results and Claude's structured evaluation prompts. Minor variation may still occur in source citations (Brave finds slightly different pages each run) but verdicts and scores are stable.

TruthPrism uses AI which can make mistakes. Results should be used as a starting point for further research — not as a definitive verdict.

See it for yourself

Install the Chrome extension and click on any article. Your first check takes about 30 seconds.

🧩 Install Chrome Extension Get Started Guide