> There's no way to know if/what an AI experiences
Getting the state of neural net, at a given point in time, is easy. There are many ways to see exactly which neurons activate, why they activate, how much they activate, etc. For smaller neural nets, this is actually easy to do - here's a blog post about it:
https://distill.pub/2020/circuits/visualizing-weights/
As neural nets get larger and larger, interpretability gets harder and harder. However, I wouldn't say it's impossible.