TruthLens
Methodology

How TruthLens turns a YouTube link into a credibility score.

Transcript-first. Signal-weighted. Calibrated. Honest about what it can't do.

1. Fetch transcript

We pull the video's transcript or auto-captions directly from YouTube. No download, no scraping of your account.

2. Extract claims

An LLM segments the transcript and identifies the load-bearing claims, sourcing language, and hype patterns.

3. Weight signals

We score evidence quality, source attribution, hedging, sponsorship disclosures, and known misleading rhetorical patterns.

4. Calibrate verdict

Signals roll up into a 0–100 score, a Legit / Mixed / Misleading verdict, and a confidence level (high / medium / low).

What we measure

Six signal families, weighted by how reliably they predict credibility in human-rated samples.

Source attribution

High

Named studies, primary sources, on-record experts.

Evidence quality

High

Numbers with units, citations, falsifiable claims.

Hedging language

Medium

Calibrated uncertainty vs absolute claims.

Sponsorship disclosure

Medium

Clear paid-promotion / affiliate disclosure.

Misleading patterns

High

Cherry-picking, false dichotomy, fear-bait, fake authority.

Engagement bait

Low

Clickbait titles vs. content delivery match.

How to read the score

The 0–100 score is a quick filter, not a final verdict. Always weigh confidence level alongside it.

80–100
Legit
Well-sourced, calibrated language, claims map to evidence.
50–79
Mixed
Some good points, some weak claims or missing sources. Watch with skepticism.
0–49
Misleading
Multiple red flags — unsupported claims, manipulative framing, or fabricated authority.

What TruthLens isn't

  • Not a fact-check oracle.
    We flag patterns of credibility, not the truth of every individual claim.
  • Not a takedown tool.
    We don't report videos, contact creators, or affect YouTube ranking.
  • Not a substitute for expertise.
    On medical, legal, or financial topics, always consult a qualified professional.
  • Not perfect.
    Confidence is reported on every analysis. Treat low-confidence verdicts as a hint, not a ruling.

Privacy by default

We process the video URL and its public transcript. We don't access your YouTube account, watch history, or personal data. Analysis history is only stored when you're signed in.

Multilingual Transcript-first Calibrated confidence

See it on a real video.

Browse hand-picked examples or analyze your own link.

See examples Try with a link