Can a model not just ignor all things that have no counter-argument by default? Like - if there are not flat earthers, widly debunked, drop the idea of a spherical earth? It only exists if it was fought over?
Even if you could do this rigorously (not at all obvious with how LLMs work), it's not a reliable metric: you can easily fabricate debate as well, and in this case the main issue was essentially skimming the surface of the reports and not looking any deeper to see the obvious red flags that it was an april-fools-level fake (which obviously even a person can fall for, but LLMs are being given a far greater level of trust for some reason)
you would just game it the same way then, and how would it know who won an internet argument? how can it prove who is telling the truth and whos... hallucinating?
what if it's something correct that doesn't have a counterargument, like "photons with a 450nm wavelength are perceived as red by the average human eye"
And how would it know if a counterargument exists anyway, and if it actually does make sense?
You're grasping for a reliable unsupervised truth machine. That's a fundamentally intractable problem until you limit it down to a wolframalpha clone. And not even that by LLMs.
