In case you haven't heard of them before, here's an explanation of Counterfactual Oracles. If you erase the answer, you're not supposed to ask the same question to another oracle.

My first assumption was that this could lead to self-fulfilling prophecies such as where you fail because the oracle tells you fail so you lose motivation. However, I don't see how this would happen if we asked the question twice.

Let's assume you'll succeed unless you're told that you will fail. The second counterfactual oracle is evaluated based on the case where both counterfactual oracles fail, so it'll tell you that you'll succeed and then you'll actually succeed because you dodged the self-fulfilling prophecy. The first counterfactual oracle will be evaluated by how well it matches the weighted average of the case where neither oracle answers and where the first oracle answers. In either case, you succeed, so it'll tell you that you succeed and again you'll dodge the self-fulfilling prophecy.

So given this, can anyone explain how this could go wrong?

I'll note that it's not clear why you would use this technique. If you were to use it infinitely, then there would arguably be no probability where none of the oracles gives you an answer which could ground the process (I'll concede that some people might argue this technique would work for infinitesimal probabilities). If you use it a finite number of times, then you could have just adjusted the probability of the original oracle engaging in erasure instead.

In case you haven't heard of them before, here's an explanation of Counterfactual Oracles. If you erase the answer, you're not supposed to ask the same question to another oracle.

My first assumption was that this could lead to self-fulfilling prophecies such as where you fail because the oracle tells you fail so you lose motivation. However, I don't see how this would happen if we asked the question twice.

Let's assume you'll succeed unless you're told that you will fail. The second counterfactual oracle is evaluated based on the case where both counterfactual oracles fail, so it'll tell you that you'll succeed and then you'll actually succeed because you dodged the self-fulfilling prophecy. The first counterfactual oracle will be evaluated by how well it matches the weighted average of the case where neither oracle answers and where the first oracle answers. In either case, you succeed, so it'll tell you that you succeed and again you'll dodge the self-fulfilling prophecy.

So given this, can anyone explain how this could go wrong?

I'll note that it's not clear why you would use this technique. If you were to use it infinitely, then there would arguably be no probability where none of the oracles gives you an answer which could ground the process (I'll concede that some people might argue this technique would work for infinitesimal probabilities). If you use it a finite number of times, then you could have just adjusted the probability of the original oracle engaging in erasure instead.