But those scores are only the likelihood of the given token being the next one in a plausible phrase, not that the underlying data is correct.

Yes that is correct but it’s still better than nothing, there are some assumptions to be made such as that the model was not trained on junk data.