Tomorrow on my usual Monday evening national network radio tech segment, I'll be briefly discussing how a new study shows that AIs can be tricked into telling lies over the long term, despite efforts at "safety training" that try to prevent this from happening.
@lauren yeah, the solution isn't to try to make AI safe. That's an exercise in futility.
The solution is to let people know AI isn't safe, and to respond accordingly.
It's the exact same as accepting that Wikipedia will not be authoritative and therefore to treat it with skepticism. Exactly the same with AI.
Heck, it's like talking about making alcohol safe. Nope: it's toxic and altering. It is not safe and cannot be made safe. Therefore, treat it as an unsafe element.
@lauren ha, equally unpractical as making Wikipedia safe then :)
It strikes me as the exact same problem.
Neither Wikipedia nor AI output is reliable.
Neither can be made reliable.
So the way we address one is likely very related to the way we address the other.
@volkris To some extent.
@volkris Not practical. Even now, most people assume Wikipedia info is accurate. Millions of people burned by this all the time. AI is even more convincing when it's wrong.