Um. I've still failed to be clear.

The nature of AI is that it is inherently sufficiently complex, that, although we may well be able to get better at predicting the kinds of characteristics that might result from implementation, the actuality of implementation will likely not just surprise us, but confound us.

I'm saying that any attempt to develop approaches that lead to Friendly AI , while they are surely interesting and as worthwhile as any other attempts to push understanding, cannot be relied on by implementers of AI as more than hopeful pointers.

It's the relationship between the inevitable surprise and the attitude of researchers that is at the core of what I was trying to say, but having started out attempting to be contrarian, I've ended up risking appearing mean. I'm going to stop here.