Can the Liar Paradox Prove AI Can't Be Controlled?

Can the Liar Paradox show that AI cannot be controlled?

In this video, I examine an argument from AI safety researcher Dr. Roman Yampolskiy, who appeals to the classic Liar Paradox as part of an argument for the uncontrollability of artificial intelligence.

The basic idea is intriguing: if we explicitly tell an AI to "disobey," what happens? Yampolskiy suggests that this creates a kind of paradox that reveals a fundamental problem with explicit AI control.

But does it?

To find out, we'll take a detour into the philosophy of the Liar Paradox. What exactly makes the Liar paradoxical? What role do principles like the Law of Excluded Middle, the T-schema, and explosion play? And do any of these features carry over when we're dealing with an imperative like "Disobey!" rather than a declarative sentence?

I'll argue that they don't—and that the Liar Paradox doesn't provide the argument Yampolskiy needs. That doesn't settle the larger question of whether AI can ultimately be controlled, but it does give us reason to reject this particular argument.

We'll also look at the logic of imperatives, the distinction between satisfiable and unsatisfiable commands, and the familiar puzzle of whether an omnipotent being can create a rock so heavy that it cannot lift it.

And if you stick around after the conclusion, there's some bonus material on the Genie Paradox, the Aligned Genie Paradox, and the Master-Servant Paradox.