The Eight Worlds Framework is a 2x2x2 strategic planning matrix proposed by Jake Sullivan and Tal Feldman for making US AI policy decisions under uncertainty. Rather than betting on a single prediction about AI's future, it varies three binary axes — the nature of AI progress, the ease of catching up to a frontier lead, and China's competitive posture — to make the assumptions behind a given policy explicit. The authors introduced it in Geopolitics in the Age of Artificial Intelligence (Foreign Affairs, 2026-01-27) and developed it further in The Tech High Ground (Sullivan, Foreign Affairs, April 2026).
The three axes
The framework's first axis is the nature of AI progress. One pole, superintelligence, describes recursive self-improvement far surpassing human capability, where even a narrow lead could prove decisive. The other pole, bounded and jagged intelligence, describes impressive applications that do not constitute a singular break with history — uneven capability that is strong in math and coding but weaker in judgment and creativity.
The second axis is the ease of catching up, the question central to the fast-follow problem. Where catch-up is easy, breakthroughs are copied quickly through espionage, leaked weights, distillation, or innovative training on older hardware. Where catch-up is hard, frontier capability depends on the full stack — proprietary hardware, institutional expertise, unique datasets, and a talent ecosystem — so that the model may be copyable but the infrastructure behind it is not.
The third axis is China's strategy. Under a racing posture, China aggressively funds large training runs and competing labs. Under a not-racing posture, it instead prioritizes deployment, adoption, and the commoditization of US breakthroughs.
The eight worlds
Combining the three binary axes produces eight scenarios, each with a different strategic implication identified by the authors:
| World | AI Trajectory | Catch-up | China | Strategic Implication |
|---|---|---|---|---|
| 1 | Superintelligence | Hard | Racing | Arms race — possible Manhattan Project 2.0 |
| 2 | Superintelligence | Hard | Not racing | Unipolar AI moment — lead responsibly |
| 3 | Superintelligence | Easy | Racing | All-out proliferation — resilience focus |
| 4 | Superintelligence | Easy | Not racing | Fleeting window — act decisively |
| 5 | Bounded | Hard | Racing | Grinding innovation + diffusion race |
| 6 | Bounded | Hard | Not racing | Comfortable lead — prosperity + safety |
| 7 | Bounded | Easy | Racing | Diffusion race — deployment is king |
| 8 | Bounded | Easy | Not racing | Resembles 5G contest — scale and adoption |
Using the framework
The authors describe applying the matrix through a sequence of steps. A policymaker first selects a base case — the world believed most likely — and tests each policy proposal against it. The next step is to hedge by identifying policies that work across multiple worlds, and to determine which policies are reversible if the base case proves wrong. Decision-makers then track signals of movement along the three axes, such as the pace of progress, replication speed, and shifts in Chinese investment, updating their odds and adjusting strategy as evidence accumulates.
Sullivan and Feldman argue that disagreements about AI policy often stem not from different values but from different assumed futures, and that mapping arguments onto this matrix reveals whether a real disagreement is about policy or about predictions.
Policy implications by world
The authors present the framework as a tool for stress-testing policy rather than for prediction, drawing out distinct implications for each pair of worlds. In Worlds 1 and 2 (superintelligence with hard catch-up), they argue that existential stakes justify drastic measures such as compute controls, talent restrictions, and possibly international agreements to slow racing dynamics, on a logic resembling nuclear non-proliferation. In Worlds 3 and 4 (superintelligence with easy catch-up), they hold that frontier restrictions become futile and the priority shifts to resilience — detection, incident response, and international norms.
In Worlds 5 and 6 (bounded intelligence with hard catch-up), standard competition policy applies, with investment in workforce, deployment infrastructure, and standards bodies; the techno-federalism analysis fits this region most closely. In Worlds 7 and 8 (bounded intelligence with easy catch-up), adoption speed matters more than frontier capability, which the authors liken to a "5G contest" in which whoever deploys first shapes standards, markets, and norms.
Relation to other forecasting frameworks
Sullivan and Feldman position the matrix against other AI forecasting approaches, each of which they read as implicitly assuming a particular world. AI 2027 (Kokotajlo et al.) implicitly assumes World 1 or 3 (superintelligence with contested catch-up), tracking, in the framework's Worlds 1 and 2, the scenario where "even a narrow lead could prove decisive." Situational Awareness (Aschenbrenner) makes a similar assumption of a US frontier lead with China racing, placing it in Worlds 1 or 5. Why AI Take-Off Is Relatively Slow (Cowen) implicitly assumes World 7 or 8 (bounded intelligence with easy catch-up), prioritizing adoption pace over frontier racing, and AI as Normal Technology (Narayanan & Kapoor) assumes World 8 (bounded intelligence, easy catch-up, China not racing competitively).
The authors observe that most US national security discourse sits in Worlds 1 and 2, while most tech-economic discourse sits in Worlds 5 through 8, and present the framework's contribution as making that implicit choice explicit.
Limitations
The authors acknowledge several limitations. Reality is more complex than binary axes, since each axis is really a spectrum, and China may pursue a middle path rather than either pole of the third axis. They write that the framework should evolve, with resolved questions replaced by new axes, and note that policymakers' own decisions shape which future emerges. They also note that the framework does not capture third-country dynamics, including AI investment by the EU, India, and the Middle East.
Relationships
- depends-on: Fast-Follow Problem, Export Controls (AI), AI Race Dynamics
- related: AI 2027, Situational Awareness: The Decade Ahead, Why I Think AI Take-Off Is Relatively Slow (Cowen), AI as Normal Technology
- related: Compute Governance, Semiconductor Supply Chain (axis 2 determinants)
- related: US-China AI Competition: Different Races, Different Metrics (analysis page)
- supports: The Tech High Ground (Jake Sullivan, Foreign Affairs) (same authors, same framework)
Sources
- Geopolitics in the Age of Artificial Intelligence (Sullivan & Feldman, Foreign Affairs, 2026-01-27)
- The Tech High Ground (Sullivan, Foreign Affairs, April 2026)