Under what conditions should copyright-protected content be used for AI training?
I'm so tempted to say “never”, given my disdain for the current state of affairs. But on a less emotional note:
how about under the terms specified in the licence the content is available under?
And yes, by that I mean mandatory attributions for stuff like #CreativeCommons and mandatory open sourcing for GPL'd stuff. If the training cannot do that, it is breaking the author's rights, by definition. And if the company is willing to pay for a different licensing terms and all the authors agree, so be it (at least that way the profits are returned back to the authors).
(it's funny that Anthropic can comply with licences when it comes to [destroying] printed books, hmm?)