The article seeks to support a version of Tarski's view of natural language truth, what is now often referred to as the "inconsistency view'. Expressed naively, this view claims that natural languages are inconsistent because they support paradoxical reasoning. The view is mislabelled; it wasn't even Tarski's considered position that natural languages are inconsistent. I shall argue that the attribution of inconsistency to natural language is a kind of category error that reflects the fundamental difference between natural and formal languages: the former do not transparently encode semantic relations in their structure whereas the latter do. Still, the paradoxes of natural truth are insoluble, just as Tarski suggested. This is because, I shall suggest, truth, by its very semantic role, is an inherently risky notion in that its natural expression does not come with any necessary indication of exactly what is being claimed to be true; thus, one may accidentally fall into paradox. In a phrase, truth is an opaque metarepresentational notion. If that is so, then there can never be security against paradox, at least if truth is to retain its metarepresentational freedom. This result is expected on the view that natural language is not semantically transparent.