I don't think that the ability to fuck something implies that there's an ability to unfuck something. I don't think there's any necessity to think that there will be an continually increasing ability to distinguish faked from real in past data.

The analogy would be how we can't distinguish fantasy, parable, and propaganda in histories written in the past, even though we have a lot more technology than they had 2500 years ago. We rule out most things that seem fantastic, like dragons or people rising from the dead (maybe the equivalent of finding extreme statistical improbabilities), while also entertaining the possibility that the past could have been radically different than we see it now, and some of the things we ruled out may have been real. We make suppositions based on the political alignments of the historians, and whether they could have had access to the information that they claim to record. We use geological and anthropological records to check credibility, and use confirmations found as sources for priors when judging other, maybe unrelated, things that said by that same scholar.

We've seen what current AI does when confronted by a situation like that, it fills in the gaps with plausible fantasy and hallucinates.

edit: not that there isn't low hanging fruit in detecting AI creations, but that's just a sign of early AI. AIs can't even draw hands.