Seems like it's up to Anthropic to teach it's AI model not to pimp itself out.

If every AI company steals the public data separately, it means massively increased costs for everyone who is getting their data stolen. If the AI companies "steal" from each other it's much better for everyone else.

No difference. Distillation is a valid and useful way of generating data to improve or make new models. It's still just example data to be trained on. Anthropic is doing the same with their own models, and inadvertently every other model through web scraping.

The legal difference is that this data is uncopyrightable. At most it's a TOS breach, nothing major.

Claude is trained on stolen data (the whole Internet), so I can't have any sympathy for Anthropic when someone steals from them.

Why the fuck would they do that if Anthropic is being kind enough to just give them that data (regardless of how it makes them be butthurt)?