How it works

Five signals.
One badge.
Your scroll back.

AI detection isn't a single model. It's a stack — cheap checks first, paid scans only when needed. Here's exactly how each layer works, and where each one falls short.

LAYER 01

Community library

Every video that's been scored before lives in a shared database. When a video loads, we ask the library first. If it's been seen, the score appears instantly with no API call.

~10ms Free Network effect
GET /rest/v1/videos
where video_id =
"x5RXtUn2ZwI"
{ score: 98, source: "deep-scan" }
LAYER 02

Platform AI label

YouTube now requires creators to disclose AI-generated content with an official "Altered or synthetic" label. We read that directly from the page. If the platform already says it's AI, we trust them.

~50ms Free High confidence
DOM.querySelector
'.ytd-clarification-renderer'
match "Altered or synthetic"
score: 95
LAYER 03

Keyword fingerprint

Titles and descriptions are scanned against a database of 200+ keywords across 25 languages — tool names (Sora, Veo, Midjourney, Runway, Kling), self-disclosures ("made with AI", "AI generated"), and aesthetic tells.

~5ms Free 25 languages
text.includes
"sora" ||
"ai生成" ||
"généré par ia"
flag: AI
LAYER 04

Comment signals

The community usually spots AI before the algorithm does. We scan visible comments for strong AI flags ("this is AI", "deepfake", "fake") across multiple languages. Two or more independent calls = high confidence signal.

~20ms Free Crowdsourced
comments.filter(
  /this is ai/i,
  /deepfake/i,
  /假的/i,
  /ai생성/i
)
if strong ≥ 2 → flag
LAYER 05

Deep pixel scan

When the first four layers don't fire, we capture five frames from the video and send them to our forensic AI-detection model. The per-frame scores are combined into a single result.

~2s Paid Multi-frame
POST /scan · AI model
frames: [1, 2, 3, 4, 5]
mean(score) →
{ ai_generated: 0.98 }
{ deepfake: 0.94 }

Less data, by design.

The extension is engineered to collect almost nothing. Here's what does and doesn't get stored.

+ STORED

Video IDs & scores

So the shared library works. No URL parameters, no personal context — just the platform-assigned video ID and its score.

+ STORED

Hashed device ID

A random UUID generated locally on first install. Used only to prevent the same person from flagging the same video twice.

− NOT STORED

Browsing history

The extension only activates on supported video sites (YouTube, TikTok, Instagram, Facebook, X, Reddit). Other sites never see it.

− NOT STORED

Your IP address

For rate limiting we use a one-way salted SHA-256 hash. The original IP is never written to the database.

− NOT STORED

Account info

No login. No email collection. No account to create. The extension works the moment you install it.

− NOT STORED

Captured frames

Frames are sent for scoring and never persisted on our servers. Only the resulting score is kept.

Roadmap, honestly.

Now · Private beta

v2.0 — Chrome extension

Working on YouTube, TikTok, Instagram, Facebook, X, Reddit. Forensic AI detection, community library, free tier.

Next · 1–2 months

Chrome Web Store launch

Public availability. Pro tier with unlimited scans for $2.99/month. Expanded keyword database, improved channel reputation scoring.

After · 3–4 months

Watermark detection

Direct OCR detection of Sora, Veo, Runway, and Midjourney watermarks in captured frames. Free, instant, no API needed.

Later · 6+ months

Mobile share-sheet app

iOS and Android apps that use share-to-scan. By then the community library will cover most viral videos instantly.

Eventually

Firefox & Safari extensions

Browser parity once Chrome and mobile are stable. Safari requires a wrapped native macOS app per Apple's rules.

Try it free.
3 scans a day, no signup.

The extension installs in two clicks and starts working on your next scroll.

Add to Chrome — Free →