Alive Internet Theory

NOTE: All text here was written by a human.

Inspiration

This project was heavily inspired by the GPTZero mission, to scan and identify AI slop in the upcoming era.

Scrolling down YouTube, half of the videos use a generic avatar, text-to-speech audio, and no creativity. Frequent AI-generated sloppy uploading has become the new norm, clouding our feed from the quality YouTubers we used to hold dear. We made this extension to keep people informed of the information they consume on YouTube.

What it does

A Chrome extension that overlays on a YT video/short and evaluates whether it is likely AI, mixed, or likely human, based on various automated scoring factors. It displays the evaluation as "Likely human", "Possibly AI", or "AI slop," and lists the factors it took into consideration.

The factors it takes into consideration:

• Scan video transcript for AI-generated text (using GPTZero API)
• Filler words/stutters (perfect speech = red flag)
• Upload cadence (very frequent uploads of long videos are suspicious)
• Account age (long-standing channels are not as suspicious)
• Channel history (whether the author's other videos were marked "AI slop" from the criteria above)

Our extension also shows AI warning tags on the YT home page and "Up next" suggestions, as well as providing an option to hide videos evaluated as AI above a certain likelihood. Users also have the option of automatically parsing any video on the page and show tags for every video on the current feed, allowing them to curate their watching experience.

How we built it

For each scoring criteria:

• Scan video transcript: Pull YT subtitles and metadata using yt-dlp, or if not available, uses local speech-to-text. Feeds it into GPTZero's document classifier and use its probabilities. Stronger confidence = more it affects the evaluation.
• Filler words and stutters: Analyse the transcript locally for common stutters ("um", "uh", "like",...), filler words ("you know", "i mean"...) and repeating a word twice in a row. The presence of these indicates higher likelihood of human video.
• **ElevenLabs AI detection:" make a request the ElevenLabs AI classifier to check if the audio was generated with ElevenLabs.
• Upload cadence, Account age: Pulls metadata from yt-dlp to evaluate.
• Channel history: Queries other videos by same author and finds existing ratings.

Performance: Running the whole evaluation each time a user opens a video would be super inefficient. Since videos don't change, we evaluate it the first time a video is opened, and save it into a database for future access from any user.

Challenges we ran into

How to fact check? We have a feature to display if informational videos are giving misinformation. This was quite a challenge as GPT's halluncination detection worked mainly on written prose, and did not infer super accurately on spoken transcripts. A LLM call seemed obvious, but we only want to flag misinformation on videos that were meant to look like information. Therefore, we had to use LLMs to break it down into the main points, determine what kind of video it is, and then research the verity individually if the video was meant to be informative.

Accomplishments that we're proud of

• Being able to bring the "anti-AI slop" mission to a new space - videos.
• Being able to complete the transcript flow typically in under 30 seconds.

What we learned

• Detecting AI in video is complex, and there is always room to improve.
• The transcript is most but not all in identifying AI videos. For example, if the video is Ai generated but the audio is just a song, the transcript won't tell it apart. There are some cases where our algorithm won't fit, and that's where the community feedback will help.

What's next for Alive Internet Theory

• Calibrate the score against ground truth. Every deduction is a hand-picked weight, so our next step would be find a labeled set of known-human and known-AI videos to tune against.

• Fix source reachability. We measured 50% of sources usable on a news video but 13% on an oncology one, because that evidence sits behind DOI links and publisher blocks while government and NGO sources fetch cleanly. Adding DOI resolution and ranking toward reachable authorities directly raises how many claims we can actually verify.

Built With

• claude
• gptzero