An open-source credibility engine for news, web content & social media — provenance, claim verification, source reputation and manipulation signals, with evidence you can check yourself.
🌐 Try the live demo → aravindb98.github.io/verifact — paste a link, an X post, or a WhatsApp forward. Nothing to install.
⭐ If VeriFact helps you (or someone you know who forwards everything on WhatsApp), please star the repo and fork it — stars and forks are how open-source tools against misinformation get discovered. ⭐
Misinformation moves faster than fact-checkers ever can. Most of us have no practical way to answer a simple question in the moment: "Should I trust this article / post / forward?"
VeriFact answers it the only honest way software can — not with a magic true/false oracle, but with stacked, transparent evidence:
| Signal | Question it answers | Works with zero keys? |
|---|---|---|
| 🏛️ Source reputation | Who published this? Wire service, satire site, known disinfo outlet, 3-week-old lookalike domain? | ✅ |
| ✍️ Writing-style signals | Does it behave like journalism — attribution, bylines, dates — or like viral bait (ALL CAPS, "share before deleted")? | ✅ |
| 🔍 Claim extraction | What are the actual falsifiable claims here? | ✅ (heuristic) · 🔑 LLM mode |
| 📚 Fact-check databases | Has Snopes / PolitiFact / AltNews / AFP already checked this? | 🔑 free Google key |
| 🌐 Independent corroboration | Is anyone reputable and independent reporting the same thing? | ✅ (GDELT) · 🔑 better with Brave/Tavily |
| 🖼️ Media provenance | Does the image carry C2PA Content Credentials? Editing traces? Where did it appear before? | ✅ |
Every report ends in an evidence list with links — VeriFact shows its
work, and refuses to score at all (insufficient_evidence) when the
signals aren't there. No truth oracle. No black box.
| Input | How VeriFact handles it |
|---|---|
| X / Twitter link | ✅ Full post extraction via public embed endpoints (FixTweet → X syndication) — text, author, follower count. No login needed. |
| Instagram / Facebook / LinkedIn link | |
| WhatsApp forward | 📋 Forwards have no public URLs — paste the text (Text tab / CLI). Forward-style virality patterns are exactly what the style analyzers target. |
| Any pasted text | ✅ Full style, claim, fact-check & corroboration analysis. |
On social posts the platform is deliberately down-weighted in scoring — anyone can post on X; the poster and the corroboration are what matter.
1. Web UI — paste a link or a suspicious WhatsApp forward:
pip install git+https://github.com/AravindB98/verifact
verifact serve # → open http://127.0.0.1:80002. CLI — for terminal people and scripts:
verifact analyze https://example-news.com/big-story
verifact analyze "Scientists CONFIRM this one weird trick cures everything!!!"
verifact analyze suspicious_photo.jpg --caption "flood in Chennai yesterday"
verifact analyze <url> --json # full machine-readable report3. Chrome extension — one-click check of the page you're reading:
load extension/ via chrome://extensions → Developer mode → Load
unpacked (talks to your local verifact serve).
Or run everything in Docker:
docker compose up # → http://127.0.0.1:8000╭─────────────────────────── VeriFact v0.1.0 ───────────────────────────╮
│ 🚨 Very low credibility 22.4/100 (confidence 71%) │
│ │
│ 3 red flag(s). Strongest signal — source reputation: │
│ newsbreaking-24.xyz is not a recognized outlet; registered 41 days │
│ ago; judge by corroboration, not brand. │
╰───────────────────────────────────────────────────────────────────────╯
Signal Score Details
Source reputation 15 Very new domain, cheap TLD, no reputation record
Writing-style signals 31 Sensational phrasing (4 hits), no attribution,
pressure to share
Claim extraction 60 3 check-worthy claims extracted
Independent corroboration 30 No independent coverage found
Fact-check databases skip No API key configured (free key available)
VeriFact is zero-key first: it works out of the box. Keys unlock depth —
copy .env.example to .env:
| Variable | Unlocks | Cost |
|---|---|---|
VERIFACT_ANTHROPIC_API_KEY or VERIFACT_OPENAI_API_KEY |
LLM claim extraction & reasoning | paid |
VERIFACT_GOOGLE_FACTCHECK_API_KEY |
Global ClaimReview fact-check lookup | free |
VERIFACT_BRAVE_API_KEY or VERIFACT_TAVILY_API_KEY |
Live corroboration search | free tier |
Skipped analyzers are reported transparently — never silently faked.
Weighted, confidence-modulated average across all signals that produced a score, with two honesty rules baked in:
- ≥2 scoring signals required, otherwise verdict =
insufficient_evidence. - Known-bad-source guard: a high-confidence "this is a flagged disinfo outlet / fact-checked false" signal caps the total at 30 — good grammar can't rescue a known liar.
Read the full design in docs/architecture.md.
-
Social-post deep links (X/Twitter + OpenGraph fallback)— shipped in v0.2.0 -
Public demo website— live on GitHub Pages - Telegram public channels & YouTube transcripts
- AI-image detection ensemble + C2PA signing verification UI
- Multilingual sensational-term packs (Hindi, Tamil, Spanish, PT-BR…)
- Propagation analysis — who amplified this first?
- Browser extension on Chrome Web Store / Firefox Add-ons
- Claim-level corroboration matrix (per-claim search, not per-article)
Want one of these sooner? Fork the repo, grab an issue, and send a PR — see CONTRIBUTING.md.
- VeriFact estimates credibility, not truth. A scrappy blog can be right; a legacy outlet can be wrong.
- Breaking news often has no corroboration yet — VeriFact says "unverifiable right now", which is the correct answer, not a bug.
- AI-generated text is essentially undetectable today; we don't pretend otherwise.
- Source lists are curated and citation-gated, but inherently incomplete — PRs with evidence are the fix.
Contributions of every size are welcome — analyzers, source-list citations, language packs, docs. Start with CONTRIBUTING.md and the good first issues.
If you read this far — that's exactly the kind of person this project needs. ⭐ Star · 🍴 Fork · 🐛 Report an issue
Built with the conviction that the antidote to misinformation isn't censorship — it's making verification effortless.
Apache-2.0 © 2026 Aravind Balaji