ABOUT

BAIT gives YouTube viewers real-time, glanceable feedback of a measure of contempt in videos they are watching.

YouTube is an ad‑supported platform that host videos without editorial input. Videos gain visibility through the response of viewers as they view, rate, forward, or repost. Media networks, journalists, and other content creators compete for audience attention. In the media market economy, those who attract the most viewers earn the most revenue from ads that play alongside their content. This introduces a bias in news reporting towards "emotional activation" - making the reader angry to engage with the content, to make them click more, to generate more revenue.

One of the ways media engages viewers is by showing contempt towards others; towards people who are different to the viewer or people who hold different views or have different customs to the viewer. To be contemptuous is to show disrespect or intense dislike for other people. In personal relationships, contempt has been shown to be an indicator of impending failure in the relationship; it is a strong predictor of likely divorce. Contempt cements the relationship between the content creator or the platform and the viewer by making others seem not worthy of respect, and at the same time showing exaggerated respect towards the viewers themselves. We recognize this type of content as "rage bait." Contempt is not part of good journalism but if viewers click more on "news" videos which are contemptuous then market forces will result in more contemptuous "news" videos.

The goal of BAIT is to push back on this market force. We trust that as people become aware of the contempt in the videos they are watching they will reconsider their choices. Contempt is not good for a healthy media diet. It is not good for the viewer, and it is not good for society.

BAIT works is a cloud-based LLM tone measurement system. Here is a more detailed technical explanation:

  • 1The frontend Chrome extension retrieves the URL of the video you are watching and sends to the backend cloud application. Your identity is anonymized.
  • 2The cloud application retrieves the transcript from YouTube closed captions (CC) service.
  • 3The text is normalized by removing artifacts, speaker tags, etc to reduce noise without altering semantics.
  • 4Word‑level emotion scores are computed using GoEmotions/RoBERTa as quantitative features to ground downstream reasoning.
  • 5A large language model consumes the full transcript plus emotion features to produce a single, normalized contempt–respect score, uncertainty phrasing (e.g., "moderate contempt"), and explanation.
  • 6The score and explanation are cached locally on the viewer's machine for privacy and efficient UI rendering in the display.

This is part of the Sustaining Peace Project at Columbia University