The AI Safety case for UpTrust: AI "Facts" 40% from Reddit, 24% from YouTube, 20% from FB

I knew this to be true but nice to see the numbers:

This good to remember when you get info from LLMs.

But also, in a non-UpTrust world it gets worse: "User Generated Content" on these sites is becoming increasingly AI generated (our startup accelerator is literally teaching all the startups how to do that to keep up with the competition). So as time goes on, these 'sources' will get more or more distorted and self-referential.

I know these models are gluttons for data, any and all. But wouldn't it be nice to identify the most trusted folks that the most trusted people in any domain trust, especially the ones that are trusted across political divides? And then train the models on those people, or use those people to fine tune the models?