First of all, you definitely don't want any chance of incorrect recognition of handwriting coming from a nuclear weapons plant.

Loving this, this sounds like the plot of a bad sci-fi movie.

It reminds me of the start of "Brazil", bugs and all.

Sounds like a plot for a really decent sci-fi book with an AI as the antagonist.

He can use it such that the model's predictions are still manually verified by him before actually submitting them or whatever. Seems pretty harmless to me.

In his current situation, if he zones out, he produces nothing. In the situation where he's aided by the potentially faulty described system, if he zones out, he produces erroneously transcribed data.

What? In his current situation, if he zones out, he can produce erroneously transcribed data too. What's your point?

What? In his current situation, even if he doesn't zone out, he can still produce erroneously transcribed data too. Why even go to work?

Manual verification is likely just as tedious as doing it by hand in the first place.

Probably. But maybe he'll get a kick out of modeling and automating it. Challenging his intellect seems to be the key point here.

I'd first go for a pareto approach. There's probably some easy automation that can aid the manual work, which is easy to implement and provides visible performance gains. Stuff like pre selecting interesting image parts (aim for high false positive rates and zero false negatives) or even just an interface for speeding up dull manual work like pulling documents from the repository, associating relevant documents and presenting them in a fast acting UI.

I've done something similar, automating a similar process. Essentially the screen was split on half. The left was the actual image, the right was the data of what the program thinks it is after OCR and some magic interpretation. The user can scroll around and both panes would move together. User clicks accepts on each value. Otherwise user can edit each value, or add a new value (if it was completely missed) and then accept. Once everything is accepted they can save the data.

On paper it saved a lot of man hours but it was never fully rolled out. The project didn't have a long lifespan and we had enough people sitting around doing nothing to just manually do all the work. But I got a nice award for it LOL.

one could make a similar argument that no one should have tried making a car that could drive better than a human

This is more like some guy in the early 1900s having the job of hand-delivering important information on a horse.

Then deciding on his own on the advice of HN that he should try a car for that task instead, the car breaking down, and the guy getting fired because his employer never approved of this whole "car" thing and he shouldn't have introduced that into his job workflow without having talked to his employer about it.