I don’t understand how you can have preferences that you use to decide what ought to count as a “moral justification” without already having a moral reference frame.
Well, consider an analogy from mathematical logic: when you write out a formal proof that 2+2 = 4, at some point in the process, you’ll end up concatenating two symbols here and two symbols there to produce four symbols; but this doesn’t mean you’re appealing to the conclusion you’re trying to prove in your proof; it just so happens that your ability to produce the proof depends on the truth of the proposition.
Similarly, when an AI with Morality programmed into it computes the correct action, it just follows the Morality algorithm directly, which doesn’t necessarily refer explicitly to “humans” as such. But human programmers had to program the Morality algorithm into the AI in the first place; and the reason they did so is because they themselves were running something related to the Morality algorithm in their own brains. That, as you know, doesn’t imply that the AI itself is appealing to “human values” in its actual computation (the Morality program need not make such a reference); but it does imply that the meta-ethical theory used by the programmers compelled them to (in an appropriate sense) look at their own brains to decide what to program into the AI.
Well, consider an analogy from mathematical logic: when you write out a formal proof that 2+2 = 4, at some point in the process, you’ll end up concatenating two symbols here and two symbols there to produce four symbols; but this doesn’t mean you’re appealing to the conclusion you’re trying to prove in your proof; it just so happens that your ability to produce the proof depends on the truth of the proposition.
Similarly, when an AI with Morality programmed into it computes the correct action, it just follows the Morality algorithm directly, which doesn’t necessarily refer explicitly to “humans” as such. But human programmers had to program the Morality algorithm into the AI in the first place; and the reason they did so is because they themselves were running something related to the Morality algorithm in their own brains. That, as you know, doesn’t imply that the AI itself is appealing to “human values” in its actual computation (the Morality program need not make such a reference); but it does imply that the meta-ethical theory used by the programmers compelled them to (in an appropriate sense) look at their own brains to decide what to program into the AI.