Training is transformative use.

It’s not like this ends with authors being paid every time a neural network gets used - we’d see a properly open model emerge. Smaller networks, trained on hand-pruned data, already look like the way forward. Generalization from sparse examples is how any of this works.

If people can’t simply let the robot read every book in the library, they’re not gonna pay for every book in the library. They’re gonna switch to some above-board torrent of public domain / Creative Commons / permissively-licensed content. That didn’t happen first because scaling up worked better, to a point. Aaaand scale prevents competition. Also it was a great excuse to spy on as much private data as possible.

When the bubble bursts, the tech will remain and flourish, and there’s not gonna be much money involved.