"AI Safety" (as defined by clients' corporate leadership being able to say "Yes, this AI is safe for us to use and won't cause repetitional harm to our company") is both worthwhile to approximate, but impossible to achieve.<p>There exists content which will offend some important constituencies if the content is generated, but it will offend other important constituencies if the content cannot be generated.<p>I think a huge part of "AI Safety" (as I defined it above) has to involve driving society towards a consensus on what content they should expect AI's will generate vs. will not generate. It's more okay if people are not happy about it, if they are also not <i>surprised</i> by it and it doesn't generate newsworthy headlines because the behavior is expected and mundane.
Pretty nice and straight-forward. It doesn't say much but it does indicate that they understand that people found the deliberate lack of accuracy a real problem.
Sorry but I expect more from Google than a 10 person startup. You don’t get to talk about how much time and resources you spend on making safe, trustable AI while also clearly not testing and noticing that extremely basic prompts are broken. Their incompetence here has fueled the fire for the political right who were already saying that all tech companies are controlled by hardcore democrats that are trying to brainwash you and your children with liberal propaganda.