Well, (f)actually…

"So what you're saying is…"

Real-time fact-checking for YouTube.

Catches every claim, live, while you watch. Local-first inference option. Cross-vendor consensus on the cloud path.

See it in action

Play the video — the sidebar on the right reveals fact-flag cards as the clip plays. It's the same UI the extension renders on YouTube, driven here by a synced preview timeline.

Live Fact-Check
live paused · t=13:11
Sources ?
Left Right Primary
waiting for first card…

Use the recorded / live toggle in the sidebar header to switch modes. Recorded mode replays a transcript-verified session against actual sources. Live mode fires real Gemini 2.5 Flash calls per transcript segment (expect ~3-9s per card). Drag the Sources slider to change which citation family each flag pulls from.

Cards cover the full 29-minute interview — scrub anywhere and the sidebar stays in sync. Earlier flags pre-load on the dossier; new ones land as the playhead crosses each timestamp.

Fact check this  →

Some of our favorite fact checking resources.

What it does

Pre-roll dossier

One-paragraph briefing on the video before captions arrive.

Confidence-tinted flags

1–5 confidence scale, traffic-light coloring, pulse animation on high-confidence cards.

Citation retrieval

Every non-skipped flag gets an authoritative source attached asynchronously.

Bias-balanced sourcing

Five sourcing profiles — Primary, Centrist, Left, Right, All. Same product, both ends of the spectrum.

Cross-model consensus

Opt-in. High-confidence flags get a second opinion from Meta Llama 4 + xAI Grok in parallel. "Consensus" badge when 2-of-3 vendors agree.

In-video chyron

Bottom-of-player overlay on top flags. Follows fullscreen.

Click-to-jump

Any timestamp in the sidebar re-seeks the video.

Session export

Markdown log of every flag with YouTube deep-links back to the moment.

Bias slider

Same fact-check, your source family

The model's classification stays constant — what a flag is doesn't change based on your politics. The citation attached to each flag does. Move the slider on the demo to see it; the table below shows where each profile pulls from.

Left

Left-leaning publications, progressive think tanks, labor-oriented sources.

EPI · Guardian · Vox · NYT · MSNBC

Centrist

Mainstream non-partisan outlets and centrist think tanks.

Pew · Brookings · Reuters · AP · BBC

Right

Right-leaning publications, free-market think tanks, conservative academic sources.

AEI · Heritage · Quillette · WSJ · NRO

Primary

Peer-reviewed papers, primary government data, original source documents — no media intermediary.

NBER · JSTOR · BLS · ONS · JPSP · transcripts

All

No source-lean preference — whatever search ranks highest.

Wikipedia · top SERP · mixed

Under the hood: the classifier (Gemini 2.5 Flash) decides what gets flagged identically regardless of profile. A separate Gemini call with Google Search grounding fetches the citation — the profile injects a source-preference into that prompt, which biases the search ranking. Same claim, different supporting link.

Local-only mode

Point at LM Studio. Nothing leaves your machine.

No cloud calls, no API costs, no caption data egress. The entire pipeline — caption ingest, claim extraction, fact-check, citation retrieval — runs on your hardware via a local LM Studio endpoint.

Real privacy, not "we promise we won't log it" privacy.

Cloud path

Your Vertex AI project. Your bill. No middleman.

Cloud mode authenticates directly against your Google Cloud Vertex AI project. There's no SaaS proxy between you and the model — the extension talks to Google, you pay Google, and there's nobody in the middle to leak, log, or rate-limit you.

Install

Five minutes. Developer-mode unpacked extension — Chrome Web Store listing comes later.

  1. 1

    Download the latest release

    Grab the zip from the releases page and unzip it anywhere.

  2. 2

    Open chrome://extensions

    Paste it into the URL bar — Chrome blocks linking to chrome:// pages directly.

  3. 3

    Toggle Developer mode

    Top-right corner of the extensions page.

  4. 4

    Load unpacked

    Click Load unpacked and point at the unzipped chrome-extension/ folder.

  5. 5

    Fill in Options

    Cloud mode: paste your Vertex AI project ID + a bearer token from gcloud auth print-access-token. Optional: drop in a free Google AI Studio key to route citations + dossier through a separate quota pool. Local mode: LM Studio endpoint URL.

  6. 6

    Visit a YouTube video

    The sidebar opens automatically. Receipts start arriving as the captions flow.

  7. 7

    Turn on captions (CC) on the video

    Click the CC button in YouTube's player controls before pressing play. The extension reads the live caption track from the page — without CC enabled at least once on the video, the sidebar stays at "waiting for captions…"

Pricing

Free covers everything you need to start fact-checking today. Pro is one click and $3.50/mo — for now it just kills the ads, but we'll layer in more as features ship.

Free
$0 forever

Everything that ships with the extension today.

  • LM Studio local-only mode (zero cloud egress)
  • Cloud Gemini classifier (BYO Vertex AI key)
  • Citation retrieval (Google Search grounded)
  • Cross-vendor consensus (Llama 4 + Grok + Claude)
  • Pre-roll dossier
  • Bias-balanced sourcing profiles
  • In-video chyron, click-to-jump, session export
  • 1–5 confidence flags
Download free ↓
PRO
Pro
$3.50 / month

Everything in Free, plus:

  • No ads
  • More benefits coming as Pro features ship

Stay in the loop

Get the Pro launch announcement plus Thoth Intelligence's weekly notes on AI, fact-checking, and what we ship next.

Published at thotharkives.substack.com

What's next

v0.6.0 (May 2026) added a Google AI Studio fallback — paste a free AI Studio key in Options and citation + dossier route through generativelanguage.googleapis.com instead of Vertex AI. The Download button always serves the latest release.