Ben Werdmuller writes about the fallout from an attempt to train AI on Bluesky posts:

So the problem Bluesky is dealing with is not so much a problem with Bluesky itself or its architecture, but one that’s inherent to the web itself and the nature of building these training datasets based on publicly-available data.

I also like Tantek Çelik’s proposal to add a “no-training” flavor of Creative Commons. I blogged about that a couple months ago.