Adding to this, the concern is that the human might have to already know how to program an aligned recursive meta-programmer.

I think a variation on ALBA could actually address this fairly minimally, by ensuring that A_N is aligned even if the human has no programming skill. The idea would be to train A_i to judge itself well compared to noise, so that it is at least a moderately good B_I. This involves an expensive training procedure where it is trained to do certain things in bootstrapped mode, though. Also, this potentially harms the ability for A_i+1 to be better than A_i, which would be a serious problem.