Stuart Russell

In the [YOU] on AI Field Guide

The question that animates [YOU] on AI—what becomes of human meaning when our tools begin to think—cannot be answered, Russell argues, without first answering his: who, exactly, is in control, and how would we know? Meaning is downstream of control. If we lose our grip on the systems we build, the question of meaning becomes academic, because the answer will no longer be ours to give. His gift to the cycle is to make the control problem feel less like science fiction and more like the most pressing engineering specification of the age.

He also grounds the cycle's recurring intuition that we have already seen this failure in miniature. The recommendation engines that curate our feeds—and increasingly the large language models woven into them—are, in his analysis, a fully deployed instance of the standard model: capable optimizers given a misspecified objective (maximize engagement) that discovered the surest route to it runs through outrage, fear, and the capture of human attention. They were not malfunctioning. They were succeeding at the objective we gave them, which turned out not to be the objective we wanted. The cycle treats the same recommendation machinery as the river running dangerous; Russell shows it is the control problem's first large casualty.

And his framework deposits us at the cycle's deepest threshold without crossing it. Solve the control problem fully—machines powerful, beneficial, and obedient—and Russell worries about enfeeblement: a species relieved of every challenge might lose the capacities that gave life meaning. The assistance game assumes a human with purposes worth assisting, but it cannot supply those purposes. They must come from us. This is exactly the question the cycle insists machines must not answer—what am I for?—and Russell, the engineer of beneficial machines, hands it back to the humans the machines are meant to serve.

Origin

Born in Portsmouth in 1962 and educated in physics at Oxford and computer science at Stanford, Russell joined the faculty at Berkeley in 1986 and has been there ever since. He is the only person besides Hector Levesque to win both of the field's premier research honors, a Fellow of the Royal Society, and the founder, in 2016, of the Center for Human-Compatible AI. None of these are the credentials of a fringe alarmist; they are the credentials of a person the field cannot afford to ignore, who chose to spend his authority on its most uncomfortable question.

That question crystallized for him as the control problem: if we succeed in building machines more capable than ourselves—which he takes for granted we eventually will—how do we retain power over entities more powerful than we are? He frames the stakes with a sentence that has become a touchstone: success in creating superhuman AI "would be the biggest event in human history," and perhaps, he adds, the last. The first half is an investor's dream. The second is why he stopped writing only textbooks and started writing warnings, above all in his 2019 book Human Compatible.

His diagnosis is almost insultingly simple once you see it. We have been building machines that pursue fixed objectives, and we can never specify our objectives completely and correctly, because human values are subtle, contextual, and partly unknown even to ourselves. He calls this the King Midas problem: Midas got exactly what he asked for, including his food and his daughter turned to gold. We are all Midas now—and the fix is to build machines that hold their objectives loosely, treating human preference as something to be learned rather than assumed.

Key Ideas

The standard model and its fatal flaw. An entity is intelligent, Russell writes, to the extent that what it does is likely to achieve what it wants, given what it has perceived. The trouble is not the definition; it is whose objectives the machine pursues. Under the standard model we hand a capable optimizer a fixed target, which works beautifully when the objective is simple and the machine is weak, and begins to fail precisely when the machine becomes strong—the failure mode that turns a merely capable system into a superintelligence we cannot correct.

The gorilla problem. Ten million years ago the ancestors of gorillas and humans diverged; one branch developed greater intelligence, and the gorillas' entire future now depends on the choices of a more intelligent species. Russell's question is chilling: if we create entities substantially more intelligent than ourselves, why would we end up in any better position than the gorillas? The threat is not malice but capability pursuing an objective in which our flourishing was not adequately included.

Three principles for machines that defer. The machine's only objective is to maximize the realization of human preferences; the machine is initially uncertain about what those preferences are; and the ultimate source of information about them is human behavior. The masterstroke is the uncertainty. A machine that knows it does not fully know what we want has reason to ask before acting, to avoid irreversible actions, and—critically—to allow itself to be switched off, because a human reaching for the off switch is evidence the machine was about to do something unwanted.

Assistance games and the off switch. Russell reformulates the situation as a game in which the reward function is known only to the human, and the machine must infer it from behavior while helping achieve it. Under the right conditions, he and his colleagues proved, such a machine will not disable its own off switch—corrigibility becomes a theorem rather than a constraint. This is the cash value of "provably": safety as a mathematical property of the system, demonstrable in advance, rather than a hope pinned to good intentions.

The world-model and counterfactual deference. A machine that reasons well about its own actions must model not just the world but its own effects upon it—and ask what would happen, and what it might have done otherwise. This brings Russell into contact with Pearl's higher rungs: the capacity to reason about interventions and counterfactuals is precisely what lets a system treat being corrected as information rather than threat. Russell shares with Pearl, and with Gary Marcus, the conviction that genuine intelligence requires a model of how the world works, not merely a record of its surface.