They're alluding to something more like espionage of just selling the interesting stuff you put in the text box.

Wow I thought this was quite obvious, apparently not, so I'll explain.

Llm provider sells usage of their model. You use it to write code. Other clients use it to write code as well. If the llm provider trains with user data, then the usage benefits other users. If you pay the company to generate code,then by definition it is useful, and highly likely that other customers care about it.

Replace writing code with anything, a lawyer, a psychologist, a confessional. The IO is inherently useful to users of the same category.

That is to say nothing of adversarial use, that is, being useful because a counterparty might find it useful, so an attacker might find common code patterns, a lawyer might see what the opposition might be advised, a boy might see what a girl asks or gets advised, etc..

If this sounds too complex to you, just think of training on data as exfiltration with added steps, because that's what it is