This is how I learned (approximately) how GPT-2 worked. Start with GPT-2 and it's full of nonsense you don't understand. Follow the references, follow more references. Get to basic neural networks with one hidden layer like you learned about in college. Reverse the chain and build upwards.

It's less about checking whether claims are true but about how they are rationalized. E.g. the claim could be a node "masks should be mandatory" with connections to reason, "masks reduce spread of virus" and papers supporting this policy. These would also need to be connected to nodes showing values, e.g. "reducing hospital admissions is more important than freedom of choice".

In this way it would be more obvious why policies are choosen and provide a way to express current knowledge. As more nodes are added to the graph, the policy prescriptions would change which would prevent mistrust that is based on "health department said do x, now they say don't do x".