I propose a litmus test: "Would it be okay if this LLM was actually a client-side tool, trained and hosted on the user's machine?"
This framing helps highlight that (A) the data used to train and configure it is not reliably secret, and (B) a determined user can manipulate its output to fit their own goals.
1 comment
[ 5.3 ms ] story [ 16.0 ms ] threadThis framing helps highlight that (A) the data used to train and configure it is not reliably secret, and (B) a determined user can manipulate its output to fit their own goals.