The big problem of language models is that we can't assess how truthful their statements are, only how grammatically correct they are. I know people are workin

The big problem of language models is that we can't assess how truthful their statements are, only how grammatically correct they are. I know people are working on mathematically assessing truthfulness of a statement for the next generation of GPTs. Imagine if we'd have such a formula! We wouldn't need the GPTs, we could instead assess truthfulness in something like an adblocker - if the scanned text block is below a truthfulness cut-off, block it.