Safety is also safe for people trying to make use of the technology at scale for most benign usecases.
Want to install a plugin into Wordpress to autogenerate fun illustrations to go at the top of the help articles in your intranet? You probably don’t want the model to have a 1 in 100 chance of outputting porn or extreme violence.
I wrote a random password generator once. I was a naive young developer, and I thought it was helpful to generate memorable passwords, so I threw a dictionary of words into it without really checking the content, beyond the obvious swearwords. First day in production, it generated an inappropriate password and suggested it to a user.
When I replaced it with a different non-word based alphanumeric algorithm that couldn't issue someone a password of 'fat cow 392' ever again, I considered that a 'safe' implementation.
Want to install a plugin into Wordpress to autogenerate fun illustrations to go at the top of the help articles in your intranet? You probably don’t want the model to have a 1 in 100 chance of outputting porn or extreme violence.