Remember, estimates of revenue growth needed to sustain the model are 80%, Which will be impossible without pricing this far higher than people might think on the current cost per token.
Quite a few "dis-intermediation" changes show this behaviour. We do indeed now have a paperless office, as presaged by the dawn of the computer. We didn't reduce the cost of "paperwork" significantly as a result. We just shifted it into various externalities which turn out to be costly.
The paper measures productivity. There is a tendency to try to make this equate to direct labour savings, even if it's not explicit.
The paper is maths heavy. Economists don't like making qualitative statements when quant is there. Arguing for a 3.6% rise in GDP is bold. I think they haven't measured counterfactual outcomes like the inflationary effects of the rise in costs of computing inputs across the economy, or the cost of electricity, or the cost effects on other industries like construction sector as hyperscalers hoover up the inputs.
I think it's possible they're right. I'm not numerically minded, can't really fathom the inputs but they do make big claims of being responsive to live market data and trends which I guess is interesting. If the approach is right for tendency if not scale, then thats something useful.
A legitimate critique of my comment includes "you're not qualified to comment"