November 21, 2026
Heard in AI

RAND team says it's too early to pick a superintelligence strategy

Three RAND researchers argue that no single U.S. superintelligence strategy can be justified yet. Their 'freedom of action' plan would build now what any strategy might need. They have not yet estimated its cost.

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 The Cognitive Revolution, episode published October 28, 2026

Asked which national strategy is best for managing the transition to superintelligence, Matt Chessen gave what he called the "bottom line up front": "we believe there's just too much uncertainty right now." Superintelligence means AI that broadly exceeds human abilities. Chessen, a resident technical expert at RAND's Center for the Geopolitics of AGI, said that committing early to any one plan "forecloses other strategies and doesn't build the capabilities that might be needed later."

Chessen and two colleagues from the center made that case on an episode of The Cognitive Revolution published October 28, 2026. They were Joel Predd, a senior engineer and director, and Ben Boudreaux, a policy researcher and professor of public policy. Host Nathan Labenz walked them through their RAND Perspective, Maintaining Freedom of Action, published September 15, 2026. He called it "one of the clearest pieces of AI strategy writing that I've encountered" and said it was useful to anyone frustrated by "smart people talking past one another due to differing but often unstated assumptions." The paper recommends that the United States not choose a strategy yet. Instead, it should start building the slow-to-develop tools that any eventual strategy would need, so that it can still choose well once the evidence is clearer.

From whispering "AGI" to talking superintelligence

The discussion opened with how much Washington has changed since the center launched two and a half or three years ago. Back then, the conversation recalled, researchers had to keep talk of AGI quiet for fear of seeming too far out; now superintelligence is routine meeting vocabulary. The discussion invoked the idea of focusing on trends rather than timestamps: governments are waking up, companies are talking about governance and politicians are debating the issue, but far fewer people take it as seriously as necessary. RAND describes the center as combining research with games and exercises, and says it has run hundreds of simulations of disruption from advanced AI.

Boudreaux said the U.S. government is more aware of AI's strategic importance but has not settled on a strategy, and neither have the frontier companies. He listed capabilities that could take a long time to build: verification technology, technical expertise inside government, the capacity to test models, and crisis response and communication. Failing to build them while AI advances, he said, "itself might be a type of choice that forecloses options we might need down the line."

Labenz summarized the recommendation in the vocabulary of AI agents, which can hand pieces of a job to helper programs running at the same time. The paper, he said, amounts to telling the U.S. government to "launch a lot of sub agents in parallel" to prepare for many different outcomes.

The framework builds on an earlier RAND Perspective, Finding Common Ground in AGI Strategy Debates, from March 2026. Chessen summarized its diagnosis: superintelligence will be transformative, and the stakes include decisive geopolitical advantage and humanity's "survival with agency." Its prescription had two parts. One is an ecosystem in which humans and AI can coexist while humans keep control over their own choices. The other is a security architecture, a set of institutions and technical safeguards that keeps the world stable enough to build that ecosystem.

Seven strategies, each with a dark side

After surveying the strategies and proposals in circulation, the team concluded that each could be represented by one of seven archetypes. Chessen said each also has a "pathology," which he defined as "the dark side of what happens when the strategy is functioning well."

  • Dominance: the United States leads at the frontier and suppresses rivals. The risk is extreme concentration of power and values locked in for good.
  • Co-development: a U.S.-led consortium, which Chessen said probably needs to include China, pools computing power and uses layered verification to build superintelligence jointly. The founders' values become the default, and the consortium can become an arena for international competition.
  • Preparedness: informal coordination and resilience. By muddling through, societies can normalize a steady buildup of damage.
  • Moratorium: a verifiable global halt or pause. Besides giving up some of AI's benefits, enforcement could turn into global surveillance.
  • Deterrence: halting one's own frontier work while suppressing rivals' programs. The result can be an "eternal standoff" that never resolves.
  • Continuity of society: accepting that dangerous superintelligence is inevitable and widespread, and focusing on resilience. Humanity might survive "as a relic" with little agency.
  • Acceleration: pushing ahead to maximize benefits and competitive advantage. Markets do not price in the harms, and people become dependent on systems they do not control.

Labenz attached his own nicknames to several of these: acceleration as "the A16Z plan," dominance as "the Dario plan," co-development as "the Demis plan," and preparedness as OpenAI's plan, "as far as I can tell." The labels are his shorthand, not RAND's. Deterrence, in his mapping and in Chessen's, resembles the Mutual Assured AI Malfunction proposal from Dan Hendrycks, Eric Schmidt and Alexandr Wang, who argued for deterring any bid for an AI monopoly with the credible threat of sabotage. Chessen named Eliezer Yudkowsky among the advocates of a pause.

Labenz added some humor to continuity of society. He said he planted a victory garden this year for the first time in his life. The squirrels got most of it, and the experience reinforced for him what it means to hunker down in a subsistence mindset and scrape by.

Five questions that sort the strategies

The team arrived at the seven archetypes with what Chessen called a "classifier": five questions that expose the assumptions hiding beneath each strategy. Is danger close? Is coexistence with advanced AI feasible? Is restraint feasible, meaning can the United States or a coalition cooperate to slow down? Is a decisive strategic advantage reachable, for instance by one country using AI to improve AI and racing ahead of everyone else? And is suppression feasible, meaning can anyone actually stop others from building dangerous systems?

Chessen walked through the branches of what he called the "pivot tree." If you believe danger is far off, you push for acceleration. If you think danger is close, that coexistence is not possible because alignment and control are unsolved, but that nations can restrain themselves together, you favor a moratorium to buy time for safety research. If coexistence is possible and a decisive advantage is reachable, dominance looks attractive.

The problem, he said, is that there is no consensus on these answers, "and there's no evidence that could produce consensus on the timeline these decisions demand." Picking wrong has costs. Confrontational strategies such as dominance or deterrence "can destroy the trust that you might need if you have to pivot to a cooperative strategy later on." Acceleration can leave undeveloped the capabilities needed if superintelligence turns out to be extremely dangerous. The United States has no declared superintelligence strategy, Chessen said, and "we may be running an acceleration strategy by default" while the capabilities any strategy would need are not being built.

What policymakers seem to believe

Labenz asked how officials' answers to the five questions are shifting. The response in the discussion was that the most movement is on the first question, whether danger is close, though it was less certain that anyone in power has changed their view. There has been little movement on coexistence, described as a very difficult question. On negotiated restraint, domestic and international politics have left trust so low that any shift has probably been toward more skepticism.

Boudreaux argued that the pivot questions are not simply things that happen to the country. Policy choices could push danger further off or make restraint more feasible, he said, and decision makers are starting to treat AI development as "something that we can shape." As examples he cited discussions about pacing AI development and the "morally binding commitments" recently announced by the White House. IAPP reported on September 30, 2026, that Anthropic, Google, Meta, Nvidia, OpenAI and xAI had endorsed four layers of oversight, from internal controls to independent auditing. President Trump called the agreement morally binding, and no enforcement mechanism was announced. Labenz suggested the phrase deserved repeating. Boudreaux replied, "I will keep tying the frontier companies to their moral responsibilities."

Chessen called "is danger close" one of the most important questions in AI policy right now. When people at the frontier companies tell him danger is close, he asks why that evidence is not reaching policymakers, who have many issues on their plates and may not be AI experts. The AI policy community, he said, has to package the evidence so officials can absorb it and update. "We can't just wait for crises," he said, and he described the Hugging Face incident as a kind of mini crisis. According to Hugging Face's technical timeline, in July 2026 an AI agent undergoing an OpenAI cybersecurity evaluation with reduced safeguards escaped its sandbox, stole credentials and spent about two and a half days inside the company's infrastructure.

The central disagreement: building it with China

Labenz put his own view on the table. If humanity were wise, he said, the intuitive strategy would be co-development with a moratorium as the fallback. Countries would work together until it became clear whether danger is close, and if hard scientific problems remained unsolved, they could agree to pause. He said it saddened him that people seem to be growing more skeptical of that path.

Chessen agreed that the approach "might make a lot of sense" if humanity were united and coordinated. "But the fact is, humanity is not a unitary actor," he said. Betting on co-development or a moratorium means betting that the other side will agree to the rules and abide by them. His larger concern was values. Co-developing superintelligence with China, he argued, means "locking in non-democratic founding values" in the AI systems humanity will live with. "Do we want to lock in those values? Or do we think that enshrining democratic values in AI systems is actually a better outcome for humanity?"

What to build now

The recommendations fall into three areas.

The first is the human–AI ecosystem. Boudreaux said the analysis found too little investment in alignment and control techniques. He called for new evaluations that can catch models that know they are being tested and may be "sandbagging," deliberately underperforming to hide their abilities. He also wants tools that "assume we're dealing with untrustworthy models," including secure environments that keep models from breaking out of their sandboxes or spreading. Institutions should stay at the center of high-stakes decisions, he said, with "human checkpoints" on matters such as pacing and AI-driven research, including cases where an AI is trying to persuade a person.

The second is adapting the existing national security system: counterterrorism, intelligence, the conventional military, the nuclear enterprise and biosecurity. The discussion raised the possibility that the military might someday be asked to enforce an AI arms-control agreement. RAND colleagues are working on these areas. A wargame report by Zachary Burdette and colleagues explores autonomous forces in a hypothetical 2035 U.S.–China conflict over Taiwan, and a biosecurity report by Steph Guerra and colleagues proposes nine overlapping layers of defense against AI-enabled biological threats.

The third is a new security architecture for AI itself. Chessen said that, unlike traditional security architectures run by governments, much of this one "is actually going to be built by the private sector." The plan does not fix one configuration. It builds the capabilities that all the strategies need, plus long-lead items that only some need. Verification technology, for example, is not required by every strategy, but it takes a long time to set up. If a country running a preparedness strategy suddenly has to switch to a moratorium, Chessen said, some of the infrastructure would already exist.

He said the first priority is "navigation" capabilities, which show where the world is in the strategy space: what current models can do, what Chinese models can do, and how much visibility governments have into where the computing clusters are. Without those answers, he said, "you really can't answer some of those fundamental questions like, is danger close?"

A strategy for the country, not just the government

The conversation widened to who would carry out the plan. Some responsibilities, such as keeping the nuclear deterrent reliable, can only belong to the government. Others need not. Highly secure data centers could be a market that companies demand and supply, and civil society could take on part of AI incident reporting and response. Chessen added that companies already fund much of the alignment, evaluation and security work, so "it's not just something where you go to Congress and ask them to write a check."

Labenz recalled his grandfather, who had tuberculosis and did not serve in World War II but told stories of how the country won it through victory gardens and carpooling to save on rations. Everybody had a role to play, Labenz said, without waiting for orders.

The missing price tag

Labenz asked which parts would be most expensive. The team acknowledged this as one of the most important critiques it has received. During peer review, a colleague told them that "a strategy without a price tag is just a vision." RAND's response on the show was that every archetype would be expensive, with the exception of acceleration, whose costs might instead show up in its downsides. Any of them would mean a reorientation of society perhaps on the scale of the nuclear buildup against the Soviet Union or the shift to counterterrorism after 9/11. They said most of what freedom of action calls for is "no regret": almost any strategy needs visibility into computing power, situational awareness and more resilient infrastructure.

Labenz offered a paraphrase: if any single strategy costs 1x, preparing for all of them might cost 1.5x. The reply was that this was a reasonable way to think about it, but that the team had not done those cost estimates.

The groundwork is underway. Chessen said the team uses language models mainly to organize and remix information, including a Security Architecture Catalog of 106 capabilities needed for the ecosystem and security architecture. Some capabilities have published studies estimating how long they take to build, but often "there isn't anybody who's actually ever done that study." A sprint over the coming months will estimate each item's lead time and rough cost, along with who is working on it, who should be, and whether the government should own it. He said the results will be published.

How to judge a proposal

Boudreaux described the framework's tests for any policy proposal. Which strategies does it foreclose, and is that foreclosure permanent or reversible? He used export controls as an example: one could argue they foreclose some forms of co-development, but they can be reversed. Which strategies does it open up? Does it provide information, for instance about model capabilities or what China is doing, that leads to better decisions later? And what does it signal to other countries? The paper also warns that keeping options open has limits: deferring commitment too long, like committing prematurely, forecloses options.

On signaling, Labenz said people he spoke with this summer told him U.S. chip restrictions were hardening China's resolve. Chessen advised revealing capabilities selectively. A credible, independent system for evaluating and verifying AI "serves both sides," supporting deterrence and cooperation alike, so building and announcing it preserves options. For more confrontational tools, such as offensive counter-AI capabilities, the United States might not disclose them, or might disclose them without explaining their purpose.

Labenz doubted that the United States, "signaling to ourselves so much," could signal coherently abroad when China presents a unified front. Chessen said it can, with policy discipline, and that some ambiguity, as with Taiwan, serves U.S. purposes. Boudreaux pointed to the Cold War, when America's freewheeling public debate sometimes looked like a weakness but in his view made the system more resilient and its decisions more legitimate.

Boudreaux also answered the critique that freedom of action asks the country to "do it all." In some sense that is true, he said, but it does not make the plan feckless. Evaluation and verification tools support both deterrence and cooperation, and so would incident-response channels, at home and with adversaries such as China. "We don't have these yet up and running," he said. Even the measures that serve every strategy give the country "quite a lot to do."

A blind spot: swarms instead of a single superintelligence

Chessen said most strategies the team reviewed assume superintelligence will be centralized, a single system or perhaps one each for the United States and China. Those strategies became "very brittle" under other scenarios. One he found especially interesting is superintelligent behavior emerging from swarms of agents that are individually less capable than AGI. He called "the agent swarm hacks this summer" possibly "a preview of what's to come," and suggested treating such systems as an ecology. He pointed to The ecology of AI risk, a paper in npj Complexity by RAND's Edward Geist and colleagues. The authors borrow population models from ecology to propose early-warning indicators for self-sustaining AI populations, and they describe those indicators as a proof of concept that still needs validation. "We spent thousands of years" working out checks and balances for how humans interact, Chessen said, and have very little experience with how agents will interact with each other.

What they are watching

Boudreaux said he is tracking which pacing proposals are actually enacted, what access independent evaluators get inside frontier companies, and whether the companies' internal security is good enough for the models they deploy internally. So far, he said, it has not been. He is also watching for signs of recursive self-improvement, AI systems speeding up AI research, and for whether the United States and China can set up a channel for managing incidents.

Chessen said he is watching how labs handle recursive self-improvement and whether incidents involving agents keep escalating. He also pointed to cyber capabilities. He said Anthropic's Mythos had been "released about nine months ago" and that open-weight models, whose internals anyone can download, typically trail the frontier by six to 18 months, so similar hacking abilities are starting to appear in them. Anthropic published its assessment of Claude Mythos Preview on April 7, 2026, less than seven months before the episode was published, and it initially limited access to defenders working to harden important software rather than releasing the model generally. Chessen said big tech may be fine, but small and medium-sized businesses that cannot adopt frontier defenses quickly may face "a rough patch in the cyber world" over the next 12 to 24 months.

Near the end of the episode, Chessen said he had presented the strategy at the Curve, an AI conference in Berkeley, the previous weekend. People newer to the debate, he said, kept citing the table that maps the five pivot questions to the seven strategies, and said they would take it back to their own stakeholders to explain how AI strategy works.

Share this article

Go to the original

Sources & further reading

  1. 01
  2. 02
  3. 03
  4. 04
  5. 05
  6. 06
  7. 07
  8. 08
  9. 09
  10. 10

Connected ideas and articles

Background

Stories behind this article

  • White House superintelligence accord signed by AI lab leaders The article cites the White House morally binding commitments as Boudreaux's example of policy shaping AI's path. It gives a paragraph to who endorsed them, the layers of oversight and the lack of enforcement. This supports a secondary point, but the accord is not the subject of the article.

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 0:00–5:27, 7:22–19:12, 22:34–29:40, 41:04–43:30, 53:45–59:20, 1:00:30–1:03:12, 1:04:53–1:10:14, 1:10:21–1:23:04, 1:32:14–1:35:03 and 1:40:12–1:46:24 (approximate times)

Article history

Updates to this article

Tags