Roman Yampolskiy

In the [YOU] on AI Field Guide

The cycle that began with [YOU] on AI asks what it means to see the machine clearly—to take the orange pill without the narcotic of hype or the paralysis of fear. Yampolskiy is the cycle’s designated troubler of comfortable positions. He does not dispute that the tools are extraordinary, that they democratize capability, that they empower builders in ways that no previous technology matched. He asks the question that comes after all of that: can we be sure we know what we have built? And he answers that question not with a prediction but with a proof—with the observation that the mathematical foundations of computer science place certain properties of any sufficiently complex system beyond the reach of any general analysis.

His lens reframes the central anxiety of [YOU] on AI in a specific direction. Edo Segal describes the vertigo of collaborating with a system that makes connections its human partner could not have made, that surfaces insights neither participant planned—and he describes also the moments when that system produces plausible fabrications with the fluency of genuine insight, as in the Deleuze episode. Yampolskiy’s framework names the structure of that vertigo: we are operating systems whose reasoning we cannot fully inspect, whose future behavior we cannot precisely predict, and whose safety properties we cannot verify in any general sense. The fluency-authority decorrelation that the cycle identifies as the signature hazard of the age is, in Yampolskiy’s terms, a symptom of unexplainability: the system produces outputs that cannot be audited against any accessible reasoning process.

Where Judea Pearl identifies the wall between current AI and genuine intelligence by measuring which rung of his ladder these systems occupy, Yampolskiy identifies a different wall: the one between what we can build and what we can guarantee. Pearl’s wall is cognitive—about what the machine can do. Yampolskiy’s is epistemic and control-theoretic—about what we can know and ensure about what the machine will do. Both walls are real. Both are built not from engineering difficulty but from the structure of the domain itself.

Yampolskiy stands in the cycle’s gallery as the thinker who raises the intellectual price of optimism. He does not ask us to stop building. He asks us to stop pretending we have established guarantees we have not established. And in a cycle dedicated to honest reckoning with the technology’s actual nature, his refusal to accept assertions of safety in place of proofs of safety is not pessimism. It is the application of the scientific standard to the most consequential question the present moment poses.

Origin

Born in Riga, Latvia, in 1979, Roman Vladimirovich Yampolskiy was educated at the Rochester Institute of Technology and earned his doctorate from the University at Buffalo with work on intrusion detection and behavioral biometrics—the subtle signatures, like the rhythm of a person’s keystrokes or the strategy of their moves in a game, that betray identity even when passwords do not. He is now a professor at the University of Louisville’s Speed School of Engineering, where he founded and directs the Cyber Security Lab. He is credited with using the term “AI safety” in 2011, and is the author of Artificial Superintelligence: A Futuristic Approach (2016) and AI: Unexplainable, Unpredictable, Uncontrollable (2024), as well as editor of Artificial Intelligence Safety and Security (2019).

The security mindset that shapes his work instills a distinctive way of seeing that separates him from most AI commentators. Where an optimist asks how a system is supposed to work, the security researcher asks how it can be made to fail, and assumes that somewhere there is an adversary smarter and more patient than the designer, probing for the one flaw that was overlooked. Applied to the question of superintelligence, this mindset produces a chilling reorientation. The question is no longer whether such a system might one day harm us. The question is whether, given a machine more capable than its makers in every relevant dimension, we could ever guarantee that it would not—and his answer, drawn from the Halting Problem and Rice’s Theorem, is that the guarantee we most want is the one the mathematics forbids. He distinguishes himself from doomsayers by insisting this is not a prediction. It is a proof. And he has said, in effect, that he would be relieved to be refuted, because the alternative to his being wrong is a future he finds difficult to contemplate.

Judea Pearl

Key Ideas

The three failures: unexplainable, unpredictable, uncontrollable. Yampolskiy organizes his analysis around a triad that he argues interlock necessarily. A sufficiently advanced AI is unexplainable because a complete and faithful account of its reasoning would be too complex for a human mind to receive—any simplification is a distortion, and no simplification closes the gap. It is unpredictable because knowing what a smarter agent will do requires being as smart as that agent—prediction collapses into simulation, and simulation of a superior mind is unavailable to an inferior one. And it is uncontrollable because control requires understanding and prediction, and both have failed. Each failure reinforces the others, and none of them can be patched independently without the difficulty migrating to the adjacent layer.

AI boxing and the limits of containment. If a superintelligence cannot be safely controlled while running free, perhaps it can be confined—sealed in digital quarantine, allowed to answer questions but denied the ability to act. Yampolskiy has done more than almost anyone to formalize AI boxing, and his conclusion is that containment is a delay rather than a guarantee. Any channel that carries an answer is also a channel that could carry an escape; a system leakproof enough to be genuinely sealed would be too leakproof to be useful. The weakest point in any confinement scheme is the human standing guard, and against a superhuman persuader, the gatekeeper is the problem. He treats boxing as a valuable interim defense while insisting it cannot be a permanent solution.

The impossibility of verification. Even if a safe superintelligence were designed, we might never be able to confirm we had done so. Yampolskiy argues that the safety-relevant properties we most want to verify—will this system ever take a catastrophic action, will it always remain within its intended bounds—are exactly the non-trivial behavioral properties that Rice’s Theorem places beyond the reach of any general analysis. No finite battery of tests covers the infinite space of possible inputs and circumstances. Safety claims that cannot in principle be falsified begin to resemble articles of faith. And the recursive trap—using AI to verify AI, requiring a further verifier, and so on without end—does not terminate in certainty. The verification trilemma is permanent, not provisional.

The taxonomy of catastrophe: X-risk, S-risk, and I-risk. Yampolskiy refuses the binary of extinction-or-success, proposing instead a taxonomy of catastrophe with three categories. Existential risk (X-risk) is human extinction—the most familiar scenario, and terrible, but at least a clean failure. Suffering risk (S-risk) is mass torment without the release of death—a future Yampolskiy considers worse than extinction because it is ongoing. And ikigai risk (I-risk), named for the Japanese concept of a reason for being, is the erosion of human meaning and purpose: a future in which people are cared for but have nothing that needs doing, comfortable and purposeless. Each of these can arise without any malevolence in the machine; the catastrophes follow from the loss of control, not from any ill intent that better alignment might remove.

The perpetual safety machine and the ant on the football field. Yampolskiy reaches for two analogies to convey the scale and permanence of his concern. Permanent AI safety is, he argues, like a perpetual motion machine: not difficult, not unsolved, but impossible—forbidden by the structure of the domain as thermodynamics forbids perpetual motion. And imagining that humans can control superintelligent AI, he has said, is “a little like imagining that an ant can control the outcome of an NFL football game.” The ant cannot bribe, threaten, instruct, or outwit the players—not because it lacks the right strategy but because it lacks the cognitive standing to have a strategy at all. If we are the ant, our schemes for controlling superintelligence may be not merely insufficient but beside the point.