I regularly use programs that generate truly new designs through the use of Genetic Algorithms. These programs can start with several sheets of window glass and turn them into a Canon 800mm f5.6 L IS USM lens, and they need exactly zero experience to do this. The Wynne corrector for parabolic telescope primary mirrors was "found"/"invented" by one of these programs. The man who ran the program said that the Wynne solution that the program came up with would never have been discovered by humans, because it was so far from the productive areas of search.
Technically, these programs aren't AI, but rather use the tricks of genetic evolution to generate their results. But the output of these programs is definitely new and original. Humans can understand genetic algorithms; we just can't do them as well as machines can.
So machines can invent new things without knowing anything about existing experience. I see no reason why AI programs couldn't do this, too. The fact that the solution sets they come up with are being truncated by human experience (they are presently being "trained" to act like humans and non-human behavior is being rejected) does not mean that they are inherently limited to human-understandable solutions.
But, because of the inherently complex manner in which AI machines process information, it seems unlikely that we humans will be able to "understand" how they reached a decision. These things are not James Watt's steam engine governors.
*EDIT*
And now, for a little entertainment, this:
https://www.wired.com/story/future-o...-martha-wells/