Koos ten Bras


A Chrome Extension That Flags AI (Sort Of)

Why

One quiet weekend, I got bored and decided to build something fun. While browsing YouTube, I noticed I kept ending up in a rabbit hole of AI-narrated videos—content that felt strangely synthetic, usually from faceless content farms. It made me wonder: Could I build something that flags these types of videos?

At first, I had ambitious ideas—analyzing audio for synthetic narration, checking for human presence in thumbnails or videos—but quickly realized that was out of scope for a hobby project. Instead, I settled on a simple, imperfect proxy: the upload rhythm of a channel.

It’s not particularly accurate, but it was something I could build, and to my surprise, it actually worked just well enough to feel satisfying.

Chrome Extension in Action

The Chrome Extension in Action

What

I built a Chrome extension that flags potentially AI-generated YouTube content by analyzing how rhythmically and frequently a channel uploads. Think of it as a small prototype trying to detect when something’s just… a bit too efficient.

What it does:

  1. Detects the Channel ID from the currently viewed YouTube video.
  2. Uses the YouTube Data API to fetch the last 10 uploads.
  3. Analyzes the average interval between uploads and flags if multiple uploads occur in a 24-hour window.
  4. Applies a very rough, arbitrary classification of AI-likelihood.
  5. Injects a message on the video page (and logs data to the console).

It’s not rigorous or scientific—but for a one-person side project, it delivers on what I needed: a nudge, not a verdict.

Limitations

  • The YouTube Data API has strict quotas—about 100 requests/day with the default key.
  • Navigation quirks: if you go from a channel page to a video, you’ll need to refresh.
  • It doesn’t actually detect AI—just upload frequency—but it happens to correlate surprisingly well with the kind of content I personally find suspicious or low-effort.

How

I built the extension using Visual Studio Code, Cline, and a rotating cast of LLMs—ChatGPT, Claude Sonnet, and others.

At first, I worked manually, editing files one by one to understand how Chrome extensions are structured (manifests, content scripts, background scripts, etc.). It was a bit clunky, but helped me get the basics down.

Eventually, I started using Cline to refactor and generate code across files, which sped things up considerably. It felt like pair programming with an assistant who never gets tired.

I didn’t track API usage precisely, but I estimate the whole thing cost me around $4 in tokens. Not the most efficient, but totally worth it for the experience.

Conclusion

What started as a tool to flag AI-generated content turned into a fun way to explore AI-assisted coding. I’ve heard people call this kind of process vibe coding—building something based on intuition, curiosity, and just seeing where it takes you.

I think I’ll be doing this a lot more.

GitHub Repo
https://github.com/koostenbras/ai-content-flagger