The United States is increasingly organizing its artificial intelligence strategy around a concept it cannot clearly define, cannot reliably measure and may never achieve in the singular, decisive form imagined.
That concept is Artificial General Intelligence, or AGI.
In Washington and Silicon Valley, AGI has become the policy anchor and rhetorical North Star. Lawmakers invoke it to justify massive investments. Tech executives tie timelines to presidential terms or national dominance. Analysts warn that the first country to reach it will shape the global order. The language is urgent: a race, a finish line, a winner-take-all victory.
There is only one problem: no one agrees on what AGI actually is.
Moving target
Ask ten AI researchers for a definition, and you will likely get ten different answers. Some describe human-level performance across all cognitive tasks. Others frame it economically — the automation of the most valuable human labor. Still others emphasize autonomy, continuous self-improvement or the capacity for original scientific discovery.
These are not interchangeable. A system that excels at writing code, generating essays or solving benchmarks is not the same as one that can redesign its own architecture, conduct groundbreaking research or reliably operate in open, unpredictable environments.
Yet public debate and policy routinely collapse these distinctions into a single, shifting target. As observers have long noted, AGI often seems to mean “whatever the next system cannot yet do.”
‘Situated’ intelligence
Even leading figures acknowledge the issue. OpenAI’s Sam Altman has at times called AGI “not a super useful term” because definitions vary so widely. The goalposts keep moving, making any strategy built around hitting them inherently unstable.
The confusion runs deeper than semantics. AGI rests on an implicit and rarely examined assumption: that intelligence is a unitary capability that can be reproduced in a single system, and that it would closely resemble human cognition.
This is a category error.
A bird and an airplane both fly, but they do so through entirely different mechanisms. The similarity is in the outcome, not the underlying process. Today’s AI systems are like airplanes: they perform tasks that resemble human cognition — reasoning, diagnosing, optimizing, creating — through statistical pattern matching on vast amounts of data, not through experience, intention, emotion or embodied understanding.
Human intelligence is “situated.” It emerges from bodies, cultures, social relationships, context and lived reality. AI simulates tone without feeling it, reproduces patterns without inhabiting them, and generates language without genuine intention. This gap is not a temporary shortfall awaiting more scale. It is structural.
Current systems, for all their impressive advances, still show persistent limitations: shallow reasoning in novel situations, brittle generalization, lack of robust long-term memory and dependence on human-curated data and architectures. Progress is real and valuable, but it looks more like iterative improvement in powerful tools.
AI is likely to evolve more like electricity or the internal combustion engine: transformative through diffusion, integration and widespread application, not a single breakthrough moment.
Strategic miscalculation
By framing AI competition as a sprint to a decisive AGI finish line, US policy risks distorting priorities. Resources concentrate on ever-larger frontier models developed by a handful of private labs, sometimes at the expense of broader adoption, infrastructure, workforce development and institutional integration.
This creates a winner-take-all mindset that history does not support. General-purpose technologies — electricity, the automobile, the internet — diffuse across borders and contexts.
Value accrues to those who integrate and apply them effectively, not merely to those who invent them first. There is no single “owner” of electricity; its impact came from decades of engineering, infrastructure and adaptation by many players.
Meanwhile, China has pursued a different emphasis. While not ignoring advanced research, Beijing has prioritized rapid deployment: embedding AI at scale across manufacturing, logistics, urban systems, education and industry.
Chinese models have narrowed performance gaps dramatically, and the country leads in areas like AI publications, patents and industrial robot adoption. The US retains an edge in frontier capabilities and private investment. But the deeper contest is increasingly about who can turn powerful tools into systemic advantage through diffusion and integration.
The real danger for America is not “losing the AGI race.” It is winning on speculative breakthroughs while falling behind in the practical, economy-wide application of AI, producing the world’s most advanced models yet failing to fully embed intelligence into its institutions, workforce and infrastructure.
Hype cycles compound the risk. Overpromising imminent AGI already has a long track record of disappointment, potentially leading to “AI winters” of disillusionment and disinvestment.
A more realistic strategy
None of this means abandoning frontier research. Breakthroughs in models, algorithms and efficiency matter enormously. But they should not define the entire strategy. A saner approach would prioritize steps China has already taken:
– Accelerating adoption and integration across government, industry and society.
– Modernizing data infrastructure, computing capacity and energy systems.
– Investing heavily in workforce training, AI literacy and education at all levels.
– Supporting a broader research ecosystem beyond a few large private firms, including open approaches that promote diffusion.
These steps lack the drama of a Manhattan Project for AGI. They are also far more likely to determine long-term competitive outcomes.
The future of AI will not be decided by a single invention or the crossing of a mythical finish line. It will be shaped by how intelligence is embedded, distributed and governed across economies and societies.
America faces a clear choice. It can continue chasing an ill-defined phantom that shifts with every new model and headline, or it can recognize the transformation already underway: AI is not becoming a mind. It is becoming infrastructure.
