The most frightening thing is that the more unpredictable things we do, the faster the AI will accumulate statistics on them and predict them by simply lowering their priority.

It doesn't usually work quite like that. Even the slightest modification in inputs can alter outputs significantly. So the threshold between randomness and predictability is still quite fragile, hence, we still have a chance to stay on top of the situation, as long as we manage our data footprint correctly. Including "planting" data outrageously different from what we usually do.