This is an awesome experiment. Actually there may be a very simple explanation of why this model is more truthful than other language models. There is a particular way of talking when we lie. We use the passive voice more frequently, we focus on the properties of things rather than get to the point and many other tricks.<p>But the thing is, that usually we tend to forget what lies were stated as truths at the past, and there comes the time we start contradicting ourselves. The lies are forgotten by everyone after a while. So the language model it does not get too hang up to the lies, because they are not repeated as much.<p>This particular channel may be more truthful because they are mostly anonymous posters, and anonymity makes lying a little bit more useless i guess. Even if a small percentage of the posts are more truthful, say 10% or 20%, that's magnified by the machine.