US President Donald Trump says warnings about catastrophic AI risk are a ‘hoax’. Now he wants the US government to rename artificial intelligence ‘Super intelligence’ and has declared that ‘whoever wins it, wins’, collapsing established distinctions between today’s AI capabilities and hypothetical systems that would exceed human intelligence, while intensifying the rhetoric of a global technological race.
We have to take a minute to look at what happened during a speech given by the US president to the United Nations General Assembly on the 22 September. Donald Trump used his address to announce that from that moment on, the United States would be renaming artificial intelligence “Super intelligence”.
Artificial intelligence, he said, would henceforth be known in US government documents as “Super Intelligence – SI”. The White House subsequently reproduced the wording in its official account of his speech. Notably, the White House renders Trump’s new term as two words – “Super Intelligence” – while the established technical concept is conventionally written as one: superintelligence.
His reasoning appeared to be straightforward. The word ‘artificial’, he argued, makes intelligence sound fake, and ‘super’, apparently, is more accurate. He welcomed his audience, with characteristic bombast, to the new world of “SI”, though a round of applause and subsequent cheering was nowhere to be heard. And quite rightly.
I wish there were only one significant problem with this! The implications run much further, especially given the debates that have continued across the world over the past two weeks about the impending AI apocalypse. But one issue absolutely has to be pointed out.
Superintelligence already means something
Let’s just get into the academics of it first of all.
Artificial intelligence is a broad category encompassing many different types of machine-based systems and capabilities. ‘Superintelligence’ occupies a very different place in the vocabulary of AI. And given Trump’s declaration yesterday, it is really, really important to understand why.
The terminology is not perfectly standardised, and arguments about the definitions and boundaries of AI and artificial general intelligence have been debated for years. But it’s important to understand how different forms of artificial intelligence are defined and understood.
Artificial intelligence is the broadest term here, and even that has never had a single universally accepted definition. One influential tradition, associated with Stuart Russell and Peter Norvig, understands AI in terms of intelligent or rational agents: systems that perceive their environment and act in pursuit of goals. More recent scholarship continues to emphasise that definitions vary according to whether intelligence is understood through reasoning, behaviour, human-like performance or rational action. I have to point out too that even human intelligence is not fully understood or defined, hence the difficulty in comparing and defining artificial intelligence. Nevertheless, the important point here is that AI names the wider field and class of systems. It does not define a particular level of intelligence.
Within AI sit very different technical approaches. Machine learning is concerned with building computational systems that improve their performance through experience or data. Michael Jordan and Tom Mitchell describe it as a field concerned with how computers can “improve automatically through experience”. Deep learning is a subset of machine learning, using neural networks with multiple processing layers to learn increasingly abstract representations of data, which is the formulation set out by the AI Godfathers Yann LeCun, Yoshua Bengio and Geoffrey Hinton in their landmark 2015 Nature review. Much of contemporary generative AI, including large language models (LLMs), is built using deep learning techniques.
So, in very simple terms: deep learning sits within machine learning, which sits within a much broader field of artificial intelligence.
A completely different vocabulary is used to describe the scope and capability of AI systems. This is where we need to understand artificial narrow intelligence (ANI), artificial general intelligence (AGI) and artificial superintelligence (ASI).
Ben Goertzel and Cassio Pennachin distinguish the ‘narrow AI’ that dominates practical AI research – systems exhibiting intelligence in specialised domains – from artificial general intelligence, intended to operate successfully across a wide variety of problems and domains.
That boundary is contested, especially now that general-purpose foundation models, such as ChatGPT, Claude and Gemini, can perform many different kinds of task, and there is no universally agreed point at which a system suddenly becomes AGI. Recent work by Gilles Gignac and Eva Szodorai argues that present evidence is better understood as demonstrating artificial achievement or expertise than established general intelligence.
Beyond AGI lies artificial superintelligence (ASI). Here the claim becomes much stronger. Nick Bostrom’s influential definition describes superintelligence as an intellect “much smarter than the best human brains in practically every field”, including scientific creativity, general wisdom and social skills. More recent academic work similarly uses the term for hypothetical systems that surpass human intelligence across domains including creativity, strategic thinking and problem solving.
A common capability taxonomy places ANI, AGI and ASI on a spectrum of increasing breadth and capability. Machine learning and deep learning are technical approaches that cut across those categories; they are not levels of intelligence in themselves.
The boundaries between ANI, AGI and ASI may still be contested, but they describe very different levels of computational intelligence.
What would that look like in practice?
ANI describes AI whose competence is limited to particular tasks or domains. Think recommendation algorithms, facial-recognition systems, spam filters, systems built to play games like chess and Go. Some narrow systems can outperform every human alive at the task for which they were designed without possessing anything close to general human intelligence.
AGI would involve broadly human-level capability across a wide range of cognitive tasks, including unfamiliar ones, with the ability to adapt, transfer knowledge and perform reliably across domains. Current frontier models can already write code, analyse legal problems and explain complex scientific concepts to some degree. Assessing general intelligence therefore requires looking at how broad, robust and consistent that competence is across different domains and genuinely novel situations. Researchers increasingly distinguish between different degrees of generality and performance rather than treating AGI as a single threshold.
ASI goes considerably further. ASI would substantially outperform the best human experts across virtually all relevant cognitive domains: scientific research, engineering, strategy, creativity and much, much more.
So, when we talk about superintelligence, we are talking about systems capable of developing scientific theories beyond the reach of the world’s leading researchers, solving engineering problems humans could not solve and consistently outperforming expert human judgement across disciplines. Essentially, the term refers to intelligence that would substantially surpass the best human cognitive performance across virtually every field.
We are not there
Admittedly, today’s frontier systems, including general-purpose models such as ChatGPT, are much more difficult to fit into the old ‘narrow AI’ box than earlier generations of AI. Large foundation models can perform an extraordinary range of tasks, and their capabilities are advancing extremely quickly.
But let’s put today’s AI capabilities into context. The 2026 Stanford AI Index records frontier models reaching or exceeding human baselines on some PhD-level science, multimodal reasoning and competition mathematics benchmarks. Yet the same report illustrates what researchers call “jagged intelligence”: a model can achieve gold medal performance in International Mathematical Olympiad problems while still performing markedly worse than humans on something as simple as reading an analogue clock. AI agents have also become substantially more capable while continuing to fail around one third of attempts on structured computer-use benchmarks, and robots remain terrible at many ordinary household tasks.
Yes, perhaps we’re currently in a conceptually untidy position with categorising AI. Current systems have moved well beyond the traditional picture of narrow, single-purpose AI. But have we even reached AGI yet? Some researchers describe those systems as precursors on a path towards AGI, but there is no established scientific consensus that AGI has been achieved, let alone ASI. In fact, researchers have explicitly cautioned against equating impressive performance with established general intelligence.
So, when Trump takes the term at the far end of that spectrum – superintelligence – and applies it to artificial intelligence as a whole, he is literally collapsing the distinctions between the AI systems we have, the general intelligence researchers are still trying to define and measure, and a hypothetical form of intelligence that would surpass us and may never be achieved. And why? Because he thinks artificial does the technology no favours; it sounds fake.
To be clear: ‘artificial’ does not mean ‘fake’. ‘Artificial’, in this context, describes intelligence instantiated in an artificial, machine-based system. Artificial light is still light. An artificial limb is still real. Nobody needs protecting from the terrible misunderstanding that AI might secretly be imaginary.
Using the term ‘super’, however, makes a significant claim about capability. That is misleading and particularly poorly judged given the extraordinary amount of public discussion over the past couple of weeks about AI posing an existential threat to humanity. Adoption by one of the world’s leading AI powers gives that terminology far more weight than a casual misuse of the word.
A generative AI system capable of producing an image from a text prompt is AI. A recommendation system is AI. A medical diagnostic model may be AI. A large language model is AI. None acquires superhuman intelligence because somebody changes the heading on a government document.
Yet the language Trump is using now alters the story being told about those systems. Calling all AI ‘superintelligence’ rhetorically upgrades existing technology. Systems with well-documented limitations begin to sound extraordinarily capable, broadly autonomous and inherently beyond ordinary human cognition. That feeds the capability inflation and anthropomorphism already surrounding generative AI. And he couldn’t have timed it worse. The terminology risks creating the impression that current AI has actually already reached superintelligence, at a moment when public discussion is saturated with warnings about what superintelligent systems might do.
“Whoever wins superintelligence wins”
“Whoever wins superintelligence wins” turns AI development into a binary contest for technological supremacy. There is a race, there is a winner, and possession of the most powerful technology is framed as synonymous with national advantage. Trump immediately placed that contest in relation to China, saying that the United States was leading and intended to keep it that way. At that point, the rebranding starts doing serious geopolitical work.
The Cold War analogy has obvious limits. AI involves very different technologies, actors and risks. But the political logic is extremely familiar, especially after the past fortnight of arguments about existential risk and human control. Geoffrey Hinton has backed calls to slow AI development, while representatives of 20 countries, together with the European Commission, have called for frontier AI to remain under human direction, oversight and control.
The contrast at the UN was difficult to ignore. Trump chose that moment to present technological development as a race for decisive strategic superiority, while later that day Finland’s President, Alexander Stubb, used the very same UN podium to argue that states faced a choice between competition, co-operation and slowing development, concluding that “co-operation is the best way forward”. Trump had already rejected international efforts to control AI and presented superintelligence as something to be won. In that framing, restraint becomes harder to contemplate.
Trump has dismissed warnings about catastrophic AI risks as a ‘hoax’. Yet he has simultaneously adopted terminology that makes the AI systems we interact with today sound more like the hypothetical systems those warnings concern. We have not achieved superintelligence.
I have written recently about the way discussions of AI can become dominated by extraordinary futures. One story promises systems capable of transforming science, medicine and economic productivity – a utopia – while another anticipates uncontrollable superintelligence, catastrophic risk or even human extinction – the apocalypse.
Both keep pulling our attention towards what AI might eventually become.
Meanwhile, governments have decisions to make about technologies operating today: the use of copyrighted material in training; transparency around datasets and models; automation and its effects on work; the replication of voices and likenesses; accountability for AI-generated outputs; competition and market power; and how people whose labour, data and creative work contribute to AI systems participate in the economic value being created. Those questions really require a reasonably accurate account of the technology we actually have.
There may come a time when genuinely superintelligent AI exists. Researchers profoundly disagree about whether that will happen, what such a system would look like and how far away it might be.
If that moment arrives, we are going to need a word for it.
Unfortunately, the President of the United States has just decided to use that word for everything we currently call AI.