Your Microsoft example is only equivalent if Microsoft was also scanning hotmail and MSN messenger for threats against the company.

I think I see where you're coming from, but to me, the analogy falls short.

I think it's because email has the postal system metaphor to work with: I put my letter in an envelope, the post office just forwards it. Similarly, Hotmail does not _need_ knowledge of the contents of my email to forward it to its destination. If you asked a user who had access to their email, they might reasonably expect only the recipient would.

With LLMs, it is quite clear to the user that you are interacting with _something_ that is inspecting your messages and reacting to them. If you asked the user who had access to their inputs, they'd probably say "Claude" or "Anthropic". Thus, when Claude rats them out for threatening violence, they probably wouldn't be that surprised.

To me, this feels much closer to someone phoning up Microsoft's customer support phone number and mentioning that they have a gun and know where Bill Gates lives, and that then resulting in a visit from the police. I do idly wonder if it was intentional that this user had this conversation with Claude, and not, say, with ChatGPT or Gemini.