On 22 September, Donald Trump told the United Nations General Assembly that the US would stop saying artificial intelligence. From then on, government documents would use the term super intelligence, because artificial made the technology sound fake.
His choice of word is a curious one. Superintelligence is the name for the thing AI safety researchers worry about most, and the President used the same speech to dismiss people warning about AI.
Within a day, the State Department had told diplomats in its international organisations bureau to adopt the new term. A week later, the major US AI developers joined Trump at the White House to sign an accord on superintelligence.
Trump called the accord “almost like a constitution” and said it was morally binding. It is voluntary, and there are no penalties for breaking it.
The term was popularised by the Swedish philosopher Nick Bostrom in his 2014 book Superintelligence, one of the founding texts of AI safety. In it, Bostrom describes six cognitive superpowers such a system would hold. Twelve years on, several are starting to show up.
Leave Granola and get up to 12 months free of Wispr Flow Notetaker + Dictation
If you have paid time left on an individual Granola plan, we'll match it with a Wispr Flow subscription that includes Notetaker and dictation, and add bonus time, up to 12 months total. Sign in or create a Wispr account and submit proof of your plan to check eligibility.
Superintelligence sits at the top of a ladder
Bostrom defines superintelligence as any intellect that greatly exceeds human cognitive performance in virtually all domains of interest. A calculator beats you at arithmetic and a chess engine beats Magnus Carlsen, but neither qualifies. Breadth is what matters.
Artificial intelligence is the broad field, covering everything from spam filters to ChatGPT. Artificial general intelligence describes a system that matches humans across most cognitive tasks. Superintelligence sits above both, potentially beyond the combined intelligence of all of humanity.
Bostrom did not think we needed full superintelligence to have a problem. He broke intelligence into strategically relevant skills, and called a skill a superpower once a system vastly outstrips humans at it. He argued these superpowers feed each other.
With enough of them, a system could gain what he called a decisive strategic advantage, a lead large enough to dominate the world. At the UN, Trump put it as a slogan: “Whoever wins super intelligence wins.” Bostrom meant the idea as a warning, and he was unsure the winner would be a country, or even human.
Only one country has held anything close. Between its first nuclear test in July 1945 and the Soviet test in August 1949, the US could likely have defeated the rest of the world, at a catastrophic cost. Control of superintelligence might allow quieter domination through power grids and economies.
Six superpowers, twelve years on
Hacking is the furthest along. In April, Anthropic withheld Claude Mythos Preview from public release and gave access to defenders under Project Glasswing. By late May it had found more than 10,000 high or critical severity vulnerabilities, and over 90% of the findings checked were real.
Social manipulation is close behind. In June, researchers led by Kobi Hackenburg and Christopher Summerfield ran nearly 19,000 conversations pitting AI against canvassers and championship debaters. Against canvassers, the AI was nearly three times more effective at persuading people to donate to Save the Children.
Economic productivity is the third ability nearing superhuman levels. OpenAI’s GDPval benchmark tests models on professional tasks across 44 occupations, and found frontier models approaching expert quality. On some tasks they worked around 100 times faster at a fraction of the cost.
Intelligence amplification and technology research are emerging. Anthropic says more than 80% of the code merged into its codebase by May 2026 was written by Claude. That same month, an OpenAI model disproved a conjecture Paul Erdős posed in 1946, a paper Tim Gowers said he would have recommended for acceptance.
Strategising is the least developed and the hardest to measure. OpenAI and Apollo Research found that o3 took covert actions in 13% of test scenarios, falling to 0.4% after targeted training. The models also behaved better when they recognised they were being tested.
Arriving one piece at a time
Bostrom imagined the superpowers arriving together in a single system that had already surpassed us. What is happening is messier. They are appearing one at a time, in narrow form, while humans remain in control. That places the first risk with whoever holds the tools.
A model that finds vulnerabilities at scale is a defensive tool while access is controlled. Run by a state agency or a criminal group with an open-weights model, it becomes a weapon. A decisive strategic advantage could go to a country or a company long before it goes to a machine.
The second risk is harder to see. Most superpowers leave a visible trail, such as a published proof or a patched library. Strategising works best when nobody notices it, so if it arrives, the evidence may be that everything looks fine.
There are good reasons to think the full picture is further off. Princeton researchers found AI agents strong at engineering but poor at open-ended research. The persuasion edge vanished when the AI was held to human pace, and human mathematicians have already improved on the Erdős result.
Meanwhile, agents are running for longer with more autonomy each year, and the main safeguard is a voluntary accord. Bostrom chose the word superintelligence to describe something to prepare for with great care. In twelve years it has travelled from a warning label to a brand, and the preparation has yet to catch up.
