• 

  Popup notification for true statements in a video about Area 51

• 

  Popup notification for false statements in a video about Area 51

• 

  A twitter post that is likely a scam

• 

  Reddit post about Area 51 fact-checked

• 

  Page overview for factuality

• 

Mostly False — A Lie Detector for the Internet

Tagline: Real-time fact-checking, everywhere you browse.

🎯 Inspiration

The internet is full of claims — in videos, tweets, and posts — and most go unchecked. We wanted to bring fact-checking to the content instead of asking people to leave the page, copy a claim, and search for the truth. What if every check-worthy claim could be verified in real time, with verdicts and sources right where you're reading or watching?

We built Mostly False to do exactly that: a lie detector that runs in your browser on YouTube, Twitter/X, and LinkedIn, using a pipeline backed by 50,000+ verified fact-checks from PolitiFact, Snopes, and 26 other organizations.

✨ What It Does

• Browser extension (YouTube, Twitter/X, LinkedIn)
  Install once; Mostly False runs in the background. It finds check-worthy claims on the page, runs them through our fact-check pipeline, and surfaces inline verdict badges (True, Mostly True, Mixed, Mostly False, False) with tooltips. Click any badge to open the side panel for the full breakdown — source, explanation, and a Truth Meter score for the whole page.

• Dataset & stats (web app)
  A Next.js landing page and a stats dashboard show the scale of our fact-check dataset (50,000+ claims), verdict distribution, and high-level insights — the same data that powers every check.

• Scam detection (same pipeline)
  We use the same implementation to flag scam-like content: same badges and UI, we just mark those claims so users see them at a glance. Built with the SafetyKit challenge in mind — AI for human safety (misinformation, scams, manipulation).

🛠 How We Built It

Architecture

• Browser extension — WXT, React 19, TypeScript. Content scripts inject into YouTube, Twitter/X, and LinkedIn; we extract text, send check-worthy sentences to the backend via REST, and render inline annotations and a side panel (Truth Meter + claim cards). Cross-browser (Chrome MV3, Firefox MV2).

• Backend — FastAPI (Python 3.12). Single fact-check pipeline used by the extension:

  1. Check-worthiness — Azure OpenAI (gpt-4o-mini) or ClaimBuster API scores sentences so we only check factual claims, not opinions or filler.
  2. Known false patterns — Quick rules for common misinformation (e.g. flat earth, vaccine–autism, Holocaust denial) with citations.
  3. Actian VectorAI DB — 50,000+ fact-check claims embedded (Azure OpenAI) and stored; we do semantic similarity search (gRPC) to find matching fact-checks.
  4. Google Fact Check API — When we have a key, we query it for additional verified results.
  5. LLM fallback — When there’s no strong match, we use Azure OpenAI (gpt-4o-mini) to evaluate the claim and label the result as AI analysis so users know the source.

• Web app — Next.js 16, React 19, Tailwind. Landing page and stats dashboard (dataset size, verdict breakdown, insights). Calls the same backend API for stats; no video upload.

• YouTube transcript API — For YouTube pages we can fetch captions via yt-dlp (run in a thread with asyncio.to_thread on Windows) and run the same fact-check pipeline on that text.

• Data pipeline — We ingest LIAR, MultiFC, and ClaimBuster datasets, normalize verdicts, generate embeddings (Azure OpenAI), and load into Actian VectorAI DB for semantic search.

Tech Stack

😤 Challenges We Ran Into

• Windows + asyncio subprocesses — The event loop in some environments didn’t support create_subprocess_exec (used for yt-dlp for YouTube transcripts), causing NotImplementedError. We fixed it by running yt-dlp via synchronous subprocess.run inside asyncio.to_thread so we don’t depend on the loop supporting subprocesses.

• Python 3.14 vs pydantic-core — pydantic-core (Rust/PyO3) only supported up to 3.13. We recreated the venv with Python 3.12 and documented the version for the backend.

• Single pipeline for the extension — We kept one fact-check pipeline (check-worthiness → vector DB → Google → LLM) and one set of FastAPI routes and schemas so the extension gets consistent verdicts and we only maintain one implementation.

🏆 Accomplishments We're Proud Of

• One pipeline, one extension — The same fact-check logic powers every request from the browser; users get the same verdicts and sources on YouTube, Twitter, and LinkedIn.

• 50,000+ fact-checks, semantic search — Using Actian VectorAI DB for meaning-based retrieval (not just keywords) lets us surface relevant fact-checks even when wording differs.

• Transparent sourcing — Every result is labeled as Verified (matched to a known fact-check) or AI analysis (LLM fallback), so users know where the verdict came from.

• Real-time UX — Inline badges and a side panel on the page. No extra tabs, no copy-paste.

• SafetyKit alignment — We’re targeting misinformation and scam-like content with the same pipeline, putting fact-checking and safety in one place.

📚 What We Learned

• A single backend pipeline for the extension (many small requests) works well when check-worthiness and fact-check logic live in one place and the extension just sends text and gets back results.
• Semantic search over fact-checks is powerful but depends on good normalization of verdicts and labels across datasets (LIAR, MultiFC, ClaimBuster, PolitiFact, etc.).
• On Windows, asyncio subprocess support can vary; offloading blocking subprocess calls to a thread via asyncio.to_thread is a robust cross-platform approach.

🔮 What's Next

• More platforms — Extend content scripts to more sites (e.g. Reddit, news comment sections).
• Scam detection — Expand the “mark as mostly false / scam” path with dedicated signals and clearer in-UI treatment.
• Performance — Batch or cache check-worthiness and embeddings where possible to reduce latency and API cost.
• Mobile / share — Explore a shareable “fact-check this” flow (e.g. paste URL → get a report).

Built With

• actian-vectorai-db
• azure-openai
• docker-compose
• fastapi
• google-fact-check-tools-api
• liar
• next.js
• python
• react
• typescript
• websockets
• wxt