RAND urges the US to keep its options open on superintelligence, as scientists grow human brain tissue in mice
RAND has published a paper arguing that, because so much about the next phase of AI development remains unknown, the US's best strategy is to preserve options rather than commit early to one path. It calls this a "Freedom of Action" strategy, aimed, in RAND's words, at building and preserving the ability to secure US geopolitical advantage and ensure humanity's survival with agency through the transition to superintelligence. The strategy rests on four ingredients: building a human-AI ecosystem that invests in safety and shares the technology's benefits; building an AI-security architecture with visibility into compute and tools to verify agreements; adapting national-security institutions for the AI era; and investing in the capacity of citizens, firms and governments to respond. RAND lays out seven archetypal strategies across three families: coexistence (dominance, co-development with other powers including China, or a looser preparedness effort built on voluntary commitments and incident reporting), denial (a verifiable global moratorium above certain thresholds, deterrence that would use national power to disable foreign programs, or, as a last resort, building distributed, self-sustaining settlements), and acceleration, which treats rapid, market-driven progress itself as the source of safety. Which of these fits best, RAND says, hinges on five uncertainties: how close real danger is, whether humans and AI can reach a stable coexistence, whether restraint is achievable through cooperation, whether any single actor can secure a decisive advantage, and whether rival AI programs can be technically suppressed at all. The Import AI author reads current US behavior as closest to the acceleration strategy already, and compares it to driving a car while spending everything on a faster engine and nothing on seatbelts or brakes.
Separately, a group of researchers reports growing human-like brain tissue inside mice bred with deliberately depleted brains, to build an experimental platform for studying the human brain. The team first depleted glutamatergic neurons from the mouse neocortex and hippocampus, creating what they call "apallial" mice, then grafted human stem-cell-derived cortical organoids into newborn animals' cortical cavities to produce "xenocortical" mice (XCX). The human neurons integrated with the mouse nervous system: calcium imaging and electrophysiology recordings showed organized activity resembling developing neural circuits, and gene analysis found the graft's neurons carried markers seen in the human motor cortex and in a class of cells linked to social cognition. Motion-tracking software applied to the animals' spontaneous behavior found the XCX mice's behavior patterns sat between the control mice and the depleted-brain-only mice, distinct from both. In a maze test designed to require memory of previously explored environments, control mice and XCX mice performed above chance while the depleted-brain-only mice did not, indicating the grafted human tissue could support memory function. Import AI's author frames the main use as testing medical therapies for human brains inside a mouse model, while flagging the broader and more unsettled question of how far such chimera experiments, mixing human and other tissue, might eventually be taken.
A third paper, from nine institutions including ACS Research, the University of Toronto, Arb Research, the Paradigm 3 Institute, the Wharton School at the University of Pennsylvania, Goodheart Labs, the Trajectory Institute, Harvard University and the University of Cambridge, argues that pacing, deliberately managing the speed of AI progress rather than trying to halt it outright, deserves its own research agenda rather than being an afterthought to the stop-or-don't-stop debate. The authors write that the world will pace AI progress one way or another, and that without better tools it will likely do so haphazardly, through improvised reactions that fail to address the real risks. They lay out arguments against pacing (it delays benefits like healthcare breakthroughs, can concentrate power, and can create a false sense of security that lets capability overhangs build up as compute gets cheaper) alongside arguments for it (more time to prepare for AI-driven cyber and biological threats, and avoiding situations where governments hand consequential decisions, including military ones, over to AI systems). The paper stresses that any pacing intervention's effect depends on how the parties it targets respond, and singles out semiconductor and cloud infrastructure providers as having the strongest incentive to keep pushing ahead and the most to lose from any slowdown, which the authors say pushes many contexts toward harder, coordinated pacing rather than voluntary steps by individual actors. The excerpt Import AI provides breaks off mid-sentence as the authors turn to how such coordination could be improved.
Key facts
- RAND's "Freedom of Action" strategy for US superintelligence policy rests on four ingredients: a human-AI ecosystem, an AI-security architecture, adapted national-security institutions, and stronger citizen and government capacity to respond.
- RAND maps seven archetypal AI strategies across three families (coexistence, denial and acceleration), and says the right choice depends on five uncertainties, including how close real danger is and whether rival AI programs can be technically suppressed.
- Researchers depleted glutamatergic neurons from mouse neocortex and hippocampus, then grafted in human stem-cell-derived cortical organoids to create "xenocortical" mice whose neurons integrated with the host nervous system and showed organized electrical activity.
- In maze tests, control mice and xenocortical mice performed above chance while mice with depleted-only brains did not, suggesting the grafted human tissue supported memory function.
- A nine-institution paper, including Harvard, Cambridge and the Wharton School, argues pacing AI progress deserves its own research agenda, warning that infrastructure providers like semiconductor and cloud companies have the strongest incentive to resist any slowdown.
Why it matters
RAND's paper reframes the AI-safety debate as a portfolio problem rather than a single choice: instead of picking dominance, a moratorium or all-out acceleration, it argues the US should spend now to keep every option open, because it is still unclear which of the underlying uncertainties, on danger, coexistence, restraint, advantage and suppression, will resolve which way. Import AI's author reads the country's actual behavior today as already closest to the acceleration strategy the paper describes, with little invested in the safety side of that choice. The mice work matters on a longer horizon: a mouse whose brain can carry human tissue that integrates electrically and supports memory is a new kind of research platform, and a step toward a much larger space of possible human-AI-biological experiments the author explicitly flags as unresolved. The pacing paper matters because it tries to turn a slogan, slow down AI, into a research field with real tools, arguing that without them the world paces AI progress badly rather than not at all.
Who it affects
US policymakers and national-security institutions, the direct audience for RAND's strategy paper. Medical researchers who could use xenocortical mice to test therapies aimed at the human brain, and the bioethics community that will have to weigh in as such chimera models advance. Governments, AI developers, and especially semiconductor and cloud infrastructure providers, who the pacing paper says have the greatest incentive to keep development moving fast and the most to lose from any coordinated slowdown.
How to use it
RAND's paper is a framework, not a policy: it gives the four ingredients, seven strategy archetypes and five deciding uncertainties as a structure for evaluating which superintelligence strategy fits a given set of facts, rather than a single recommendation to implement. The xenocortical mouse platform is offered as a tool for testing medical therapies aimed at the human brain in a living model. The pacing paper functions as a starting research agenda: it names the incentive structures researchers and policymakers would need to account for, including the outsized incentive infrastructure providers have to resist pacing, and argues that coordinated rather than voluntary mechanisms will be needed in many contexts.
How solid is it
The RAND paper and the Nature study are described from Import AI's own summary, which quotes their stated goals and methods directly. No publication dates are given for either the RAND paper or the Nature study, and neither piece names its individual authors: the RAND paper is attributed only to RAND as an organization, and the mouse study only to "the authors." The pacing paper does name its nine contributing institutions. The excerpt Import AI provides is cut off mid-sentence partway through the pacing section, before the newsletter's account of a third research topic ("machine hermeneutics") even begins, so this retelling covers only what the available text reaches.
Risks and caveats
RAND's paper is the organization's own analysis and recommendation; nothing in the text indicates the US government has adopted any of the seven strategies it describes, including the "Freedom of Action" approach RAND favors. The xenocortical mouse work raises the ethical questions the author raises directly: how far chimera experiments mixing human and animal tissue should be allowed to go, including whether comparable experiments could ever be proposed in other species. No count of mice used in the study is given. The pacing paper's own list of arguments against pacing, including that it can concentrate power and create a false sense of security while capability overhangs build up, applies as much to future pacing efforts as any argument in favor of them.
“to build and preserve the ability to secure U.S. geopolitical advantage and ensure humanity’s survival with agency through the transition to Superintelligence”
— RAND, on the purpose of its "Freedom of Action" strategy