...Because if it meant "asking bots instead of reading",

not only of course we want the real, original material to be available - and much intellectual activity will still involve reading, analyzing and studying it,

but also, whenever actual artificial general intelligences will be available, they will be trained on said texts.

Yes, instead of technical limitations of OCR, we’ll have inscrutable language models making up random shit, and people too lazy to look at the primary source material.

I foresee a day when the expectation for any peer-reviewed article in a major journal will be to provide proof that every citation was validated against a physical copy of the material cited, not just an online version of the citation. Give the grad students something useful to do…get a free flight to the appropriate physical repository, tour around the campus, meet some friends…

You know an OCR with a Statistical Language Model acting as text plausibility checker?

They probably mean the use of GPTs to correct OCR errors

That is what I meant. But you have to integrate the subsystems properly, and I am not aware of an engine that does it.

Indeed, that is part assumption on my part the OCR will get better with a GPT like setup, but, as I said on another comment, I have trouble believing OCR won't be one of the better applications, as "is this an A" is much more straightforward than most of what GPTs have been doing.

How so? I don't understand at all.

I'm assuming that: OCR is something that's in a GPT's wheelhouse, and that OCR is something that can get much better with training. What is fact and fiction is hard on the internet for humans, look at how we have to teach youngsters how to not cite Wikipedia. But the difference between a T and an A seems much more cut and dry? I'm thinking OCR errors can be fixed / trained much easier than most domains.