Jake Sullivan and Tal Feldman have just published an article in Foreign Affairs that proposes a method for making decisions about artificial intelligence when the future is radically uncertain. Their starting premise is straightforward: nearly all debates about AI rest on hidden assumptions that no one makes explicit.
Multiple paths
Some assume superintelligence is near and that whoever arrives first wins permanently; others believe AI will be powerful but uneven, with clear limits. Some think copying breakthroughs will be trivial—through espionage, leaked weights, model distillation; others argue that leadership depends on a complete technological stack (hardware, talent, infrastructure, know-how) that cannot be replicated quickly. And then there is China: does it genuinely compete for technological frontier, or does it prefer to let others innovate before copying and scaling globally?
The matrix of eight worlds
Sullivan and Feldman construct a 2×2×2 matrix with those three binary variables. The result: eight possible scenarios, each with radically distinct strategic implications. In a world where superintelligence is achievable, hard to copy, and China is competing fiercely, it makes sense to invest in frontier capability and shield exports. But if China plays a waiting game—letting others innovate before copying and scaling through techniques like model distillation—and AI proves easy to replicate, then advantage goes to whoever diffuses faster. The emphasis shifts to governance, alliances, and global adoption.
What does this mean for Europe?
Though the article proposes a framework for the United States, it serves equally as a reference for Europe. And it signals something important: copying American strategy without understanding which world we inhabit can prove costly. If Washington bets on dominating the technological frontier whilst Europe bets on robust regulation and democratic adoption, that simply means we are betting on different worlds.
Powers that matter
The text also challenges the myth that governments control little. They have far more to do with shaping AI's future than commonly assumed. The United States does not formally own OpenAI or Anthropic, yet its export controls, political signals, public procurement, and decisions on energy, permitting, and immigration shape the entire ecosystem. Europe has equivalent levers (financing, regulation, public procurement, standards), but deploys them in fragmented fashion and without a clear theory of change.
The operational recommendation
The article proposes that these powers choose a "base case" (the most likely scenario), test each policy against it, and build hedges for higher-risk worlds. This demands institutions that think probabilistically and adjust course as signals shift. It is not about predicting the future correctly, but rather about avoiding being trapped in a single forecast whilst reality takes another path.
Choose wisely
Europe needs to build its own AI matrix. And do it now, thinking across multiple possible futures. Because the final scenario will almost certainly be far more complex than any single prediction, and whoever prepares for only one path will be following almost certainly the wrong one.
This article is republished from Futuribles. Here's the original article in Spanish and English.
