November 21, 2026
Heard in AI

RAND's Boudreaux: a credible AI off switch needs life without AI

RAND's Ben Boudreaux argues that a threat to shut down advanced AI works only if society can still function without it. He wants human agency measured and offline fallbacks kept and practiced, a habit he admits he often fails at.

A briefing reports one development at a point in time. We may correct or clarify it later; a new development gets a new briefing. How our formats work

Based on 2 episodes of The Cognitive Revolution, published October 17–28, 2026

Nathan Labenz, host of The Cognitive Revolution, described a small scare. He was preparing public versions of his personal AI setup when Claude, Anthropic's AI model, rotated the passwords on the agent console he uses to talk to his agents. Labenz said the console is "increasingly how I'm doing all the work." For a moment he was locked out of his own system. He got back in by opening a terminal and asking Claude for the new password. He called it "a little wake up call." He had already put AI between himself and "the actual levers" of action in the world, he said, and "if Claude turned bad on me, it could go really bad." Labenz first told the lockout story on the show's October 17, 2026 AI:AM episode.

He then asked Ben Boudreaux, a researcher at RAND's Center for the Geopolitics of AGI, which habits of mind people should cultivate. Boudreaux tied the everyday problem to national security. A threat to shut down an AI system is believable, he argued, only if the people making the threat can live without the system. That makes the ability to work without AI something to keep, measure and practice.

Boudreaux appeared with RAND colleagues Joel Predd and Matt Chessen in the episode, published October 28, 2026. They discussed RAND's framework for US strategy on the path to superintelligence, meaning AI that would surpass human abilities. The framework compares seven broad strategies, ranging from dominance and deterrence to co-development with rivals and a moratorium on development. It argues against committing early to any one of them.

Why RAND talks about coexistence

Boudreaux said that managing our dependence on AI is part of carrying out any of the seven strategies. Even a moratorium, he said, "will really only be feasible if we're able to live without advanced forms of AI." That means preserving human capacities at two levels. Individuals need to keep thinking things through for themselves. Societies need to keep running critical infrastructure and developing and implementing strategy.

He said this is why the RAND team uses the word "coexistence" and not just alignment. Alignment is the effort to make AI systems pursue the goals people intend. Boudreaux said that word alone can make it sound as if the only job is training models to conform to a set of established values. He described a two-way process instead: "a process of adaptation and perhaps even co-evolution." As AI becomes more capable, people become more capable of living with it. They also keep the ability to live without it, which helps manage the vulnerabilities that dependence creates.

He wants institutions to stay at the center of high-stakes decisions so that "humans are still calling the shots." People will have to keep deciding which decisions to keep in human hands, he said, and how much autonomy and authority to delegate to models. He named two candidates for human checkpoints: pacing, meaning how fast AI development proceeds, and AI-enabled research and development, the use of AI to help build more advanced AI. Those checkpoints should also catch cases where an AI is trying to persuade a human, he said.

Measuring how much has been handed over

The conversation turned to an idea that Boudreaux and his RAND colleague Alvin Moon have written about: the agency evaluation. Most AI evaluations test whether a model can do a discrete task, including dangerous ones such as helping to engineer a bioweapon. An agency evaluation looks at a model inside its real deployment. It asks whether people have crossed a threshold where they have given up some agency, perhaps without realizing it.

Boudreaux described the method as "mapping the human decision space." Evaluators then look at the roles AIs play when they enter those decisions. An AI might push some humans out of a decision, persuade them, or control the agenda they deliberate over. He said such evaluations are not typically done now and will be hard to do, because each individual decision to delegate something to an AI can be rational and easy to overlook. His answer is to track delegation over time and look for trends and trajectories instead of single decisions. "You really have to, I think, think longitudinally," he said, about the roles humans and AIs play in high-stakes choices such as whether to deploy a new model. He said this work is under way and that he hopes to publish a paper on it soon.

An earlier RAND report by Moon and Boudreaux, published April 20, 2026, lays the formal groundwork. It draws on social-choice theory, the mathematics of how individual preferences combine into group decisions. The report describes a collective choice through its decisive coalitions, the groups whose agreement is enough to settle it. It proposes three measures: how those coalitions are distributed, how small the smallest one is, and how much of each is human or AI. It identifies three pathways by which AI can erode human agency: fewer humans in decision-making roles, AI systems gaining decision-making power, and AI systems shaping which options reach human decision-makers. Within the model, power can concentrate in a single minimal coalition that decides every choice. The authors propose agency evaluations, minimum human-participation thresholds for high-stakes domains such as military applications and critical infrastructure, and tracking the human makeup of decisive coalitions over time.

Holding AIs at risk

When Labenz asked for neglected policy ideas, Boudreaux named holding AIs at risk. Humans are vulnerable to AI, he said, and AIs should likewise "remain vulnerable to us." That requires the capacity to shut models down and to threaten credibly that humans will do it. Part of that credibility, he said, comes from resilience and the ability to live without the systems. The rest comes from having the right technical approaches, legal authorities and crisis processes to act quickly. If the concern is the most extreme risks from advanced AI, he said, the ability to hold AIs at risk and shut them off is "pretty crucial." He added that the point "maybe doesn't come out as explicitly in this particular piece," meaning RAND's framework.

Habits for individuals

Boudreaux said he thinks less about which passwords Claude manages than about which kinds of analytical work, writing and critical thinking he hands to AI. He is a teacher and a professor and has small kids, he said, so he thinks about which abilities they need to keep. His practice starts with his own judgment. Before asking a model anything, he wants to be clear about his question and about what good and bad answers would look like. Then he lets "the model cook," expanding his thinking or building tools he couldn't build alone. At the end, his critical judgment comes back in, so he isn't just putting his name on the result and passing it on. The human comes in at the beginning and again at the end, he said, and he tries to instill that cycle in himself, his students and his kids.

Beyond that, he said, it is hard to recommend a single best practice for everyone. He tries to make sure he can still operate without these tools. Some of this is what he called "cutesy," such as finding his way to a friend's house without GPS.

Backstops for institutions

For institutions, Boudreaux pointed to human backstops: fail-safes that allow systems to be run manually, such as paper copies of health records. They help a society withstand an incident, he said. They also make threats to shut AI systems down "much more credible," because a shutdown would cost society much less. His approach is to identify the dependencies, reduce them or at least keep fallback options, and then practice using those fallbacks so they are actually available when needed. He doesn't claim to have mastered this. "It's very much a practice," he said, something he strives to do "and often fail at."

An AI Sabbath

Labenz proposed "an AI Sabbath day": a day each week "where we just like, don't do AI stuff." Boudreaux agreed and took the idea further. He said the Los Angeles Unified School District and New York schools are now AI-free by the districts' explicit choice. He asked whether communities need more AI-free spaces, such as community centers where people exercise or meet without AI in between. Such spaces would let people practice human abilities and prepare for a time when AI is unavailable or has to be stopped. He also suggested a human reserve labor force for critical infrastructure and hospitals. He tries to take an AI Sabbath himself, he said, but often fails.

"Well, that might be something I need to start," Labenz said.

Share this article

Go to the original

Sources & further reading

  1. 01

Connected ideas and articles

From the conversation

Podcast episodes

The Cognitive Revolution

Freedom of Action: America's ASI Strategy, w/ RAND's Joel Predd, Ben Boudreaux & Matt Chessen

Episode published This article draws on 59:21–1:00:30, 1:02:06–1:04:53 and 1:24:24–1:32:13 (approximate times)

Article history

Updates to this article

Tags