Calling It 'Super Intelligence' Doesn't Make It So

There is a quiet sleight of hand in the way the US government now talks about artificial intelligence. An executive order has moved federal usage toward the term super intelligence, a new Super Intelligence Force has been stood up to coordinate policy, and the White House Accord and the Kyoto science declaration both carry the phrase. Repeat a word often enough in official documents and it starts to feel like a description of reality. It is worth pausing on whether the reality has arrived.

The claim worth examining is not a government one, exactly. It is the impression the vocabulary creates: that the systems we use today are already superintelligent. That impression is running ahead of the evidence.

What the word is supposed to mean

In the field that coined it, superintelligence has a demanding definition. The standard formulation, associated with the philosopher Nick Bostrom, describes an intellect that dramatically outperforms the best human minds across virtually every domain, including scientific creativity, general reasoning and social skill. The key word is every. A system that is superhuman at one task and brittle at the next does not qualify.

By that measure, today's models are something stranger and narrower. They are genuinely superhuman at recall, at speed, at drafting and summarizing, at pattern matching across more text than any person could read in a lifetime. They are also, still, prone to inventing citations, fumbling multi-step reasoning they have not seen before, and failing at tasks a competent professional handles without thinking. The capability profile is jagged, not uniformly above us.

The evidence points sideways, not up

Watch where the systems break and the picture gets clearer. Enterprises have spent the year discovering that agents which sail through internal tests still fail in production, which is why many are pulling back from letting them act unsupervised. Robots that can reason about the physical world remain clumsy and years from the home. Benchmarks climb, and then someone builds a slightly novel version of the same test and scores collapse. None of that describes an intelligence that has comprehensively surpassed us. It describes a powerful, uneven tool.

The experts are not settled, and honesty requires saying so. Some serious researchers believe transformative systems are close, and at least one walked away from DeepMind over the pace of the race. Others regard the super intelligence framing as premature by years or decades. That genuine disagreement is precisely the point: when the people who build these systems cannot agree on whether superintelligence is near, a government declaring it present by renaming a category is making a political choice, not reporting a technical fact.

Why the label matters

Words shape budgets and laws. If the public believes superintelligence is already here, the policy conversation drifts toward science-fiction fears and away from the mundane harms that are real today: unreliable agents, opaque decisions, concentrated control of frontier compute. A grand name can make a technology sound both more miraculous and more inevitable than it is.

The systems are remarkable. They are reshaping how work gets done. But remarkable and uneven is not the same as superintelligent, and no amount of official rebranding closes that gap. The honest description is the less flattering one, and for now it is also the accurate one.