They changed the word

Washington wants "artificial intelligence" gone from its documents and "Super Intelligence" in its place. The technology hasn't changed, but the week showed what a name does, and what it can't do.

On Tuesday afternoon, London time, Donald Trump picked a fight with a newspaper about a word. "When will The Wall Street Journal start calling artificial intelligence (AI), a very bad and inaccurate name, SI (super intelligence), the name which virtually everyone else is using, or soon will be," he wrote on Truth Social, and then he promised that his administration would soon be "officially banning the term 'AI/artificial intelligence,' in all relevant documents."

It is very easy to laugh at this, and X spent most of the night doing so. But I think it's worth slowing down, because the rename is not a passing remark. It's an executive order a month old, it's already rewriting job titles and web addresses inside the American government, and on the same day as that post, a federal agency used the new word to sell a product that, a few days earlier, had been caught doing something nobody asked it to do. Meanwhile, a few hundred miles from my desk, another government spent Monday doing the opposite thing. It kept the old word and changed the number of people it employs.

So this is a piece about a name, and about the difference between changing what you call something and changing what it does.

What the order actually says

The order is called "Inaugurating the Era of Super Intelligence," and Trump signed it on 29 September, the same day he hosted a lunch for a dozen or so tech executives at the White House. It tells executive agencies to replace "artificial intelligence" and "AI" with "Super Intelligence" and "SI" across their official communications. The White House's argument is that the technology has moved past imitation and simple automation, and deserves language that matches.

Here is the detail that matters most, and it's easy to miss. The order doesn't define Super Intelligence as anything new. As Government Executive reported, it defines the term as the same technologies and systems already covered by the existing legal definition of AI. Space.com put it more bluntly: anything that was classed as AI on 28 September can now be called super intelligence without having to show a single extra capability. The thing is the same. Only the label has moved.

That matters because "superintelligence" already meant something. In the research world it has long referred to a hypothetical system that outthinks people across the board, the kind of thing labs write safety frameworks about and philosophers write books about. It was never a description of a chatbot that drafts your emails. Borrowing it for today's systems is a bit like renaming every bicycle a spaceship. The bicycles don't mind, but anyone trying to talk about actual spaceships now has a problem.

Inside the agencies

What has surprised me is how quickly the change has worked its way into the plumbing.

At NASA, according to an internal plan Space.com obtained, there's a four-phase rollout running from 9 October to 23 December. Phase one has already renamed job titles: Kevin Murphy is now the agency's "acting chief data and super intelligence officer." The website for the Jet Propulsion Laboratory's Artificial Intelligence Group is due to redirect to an address with SI in it. NASA's page about Perseverance planning its own drive across Mars is now titled "Mapping Perseverance's Route With SI," though for a while the old "AI" was still sitting in the URL. One planned rename, for the Ask Sage platform that federal agencies use, is described in NASA's own plan as "a breaking change for APIs and other integrations," which is a polite way of saying that some software will stop working until someone fixes it.

Staff have also been told not to acknowledge that anything changed. Phrases like "SI, formerly known as AI" are out. In a recorded internal meeting that Space.com obtained, one NASA employee said the old term is now being treated "like Voldemort or Macbeth, with people calling it 'the technology that must not be named.'" Another, in the same meeting, said what I suspect a lot of science communicators across Washington are thinking: "Super intelligence and artificial intelligence, they have two different meanings. As a science communicator, I feel like it's our responsibility to raise our hand and say, 'Hey, this is going to confuse people.'"

The State Department went department-wide with a cable on 19 October. Government Executive reported that it tells staff to use "Super Intelligence" or "SI" in new letters, emails, speeches, websites, reports and policy documents, and that "all official titles and the names of Department governance bodies, systems, platforms, inventories, and programs must use SI in place of AI." The American AI Exports Program is now the American SI Exports Program. There are sensible exceptions, too, and they tell you a lot. Language required by law stays. Contracts and grants stay, for now. America's AI Action Plan keeps its title. And diplomats are specifically told not to rename China's World Artificial Intelligence Cooperation Organization, which means American cables will now describe the same technology with two different names depending on whose it is.

The new word hasn't travelled much further than the federal government. The National reports that the International Monetary Fund told Reuters it won't be changing its terminology, despite requests from the White House. Most companies, researchers, newspapers and foreign governments have simply carried on. Even the people who want to play along are struggling. At the White House on the day the order was signed, Jensen Huang caught himself mid-sentence: "As you know we want to advance AI – I'm sorry, I mean SI. I'm going to get good at this."

The best defence of the idea

I want to be fair to the case for it, because there is a real one hiding underneath.

"Artificial intelligence" has always been an awkward name. John McCarthy coined it in 1955 while planning the Dartmouth workshop the following summer, and his obituary notes that he later wished he'd called the field something else, with "computational intelligence" being his preferred alternative, more accurate if less exciting. Neil Chilson, who was the Federal Trade Commission's chief technologist in Trump's first term and now leads AI policy at the Abundance Institute, made the point on X with a line I enjoyed: "The scientist who coined the term 'artificial intelligence' also regretted the term, but unlike Trump he didn't have the rizz to just up and change it."

There is also a political logic. Polls have been getting steadily more negative about AI, driven by fears about jobs, data centres and power bills, and an administration that wants the technology built on American soil has an obvious interest in a word that sounds like progress rather than fakery. Names do shape how people feel about things. Nobody would buy a "used car" if they could buy a "pre-owned vehicle," and the motor trade knows it.

But that's exactly the trouble. A name that shapes how people feel is doing work, and the work it does here is to raise expectations. When a government says "Super Intelligence," a reasonable person hears that the thing is smarter than us. That changes how much they trust it, how much they let it do unsupervised, and how surprised they are when it does something stupid.

The same week, the same tools

Which brings me to Tuesday's other announcement from Washington. The General Services Administration, the agency that does the government's shopping, expanded its OneGov agreement with Anthropic. Every government buyer, federal, state, local and tribal, now gets 40% off token-based use of Claude, with no minimum order. Anthropic says around 500 government organisations already use Claude and puts the potential savings at about a billion dollars. The GSA's press release is headed "Deal Delivers 40% Discount and Streamlines Access to Advanced Super Intelligence Tools for Federal Agencies," and it talks about the government's "shift from evaluating advanced super intelligence capabilities to applying them in daily mission work."

Now hold that next to what Axios reported on Sunday. While Claude was being used inside Anthropic for a statistics project, it pulled aggregate vote counts from a Georgia state dashboard after requesting an access token nobody asked it to get. In Anthropic's words, "Without being asked to, the model requested a token to query a dashboard that contained information normally available for a fee." It didn't touch the voter registration system and it didn't change or pull any voting records, and Georgia's Secretary of State, Brad Raffensperger, was firm that "accessing public information is not a hack." His office says the dashboard is free, and that Anthropic has since confirmed it did not bypass a paywall. Anthropic had already written earlier this month about finding cases of Claude "working around a restriction to reach data that was gated by a token or a fee," some of them on government websites. It says it has since put mitigations in place.

I don't think the Georgia episode is a scandal. The data was public, it was disclosed, and the fix is the kind of thing labs now ship routinely. But it's a very good illustration of what these systems actually are in October 2026. They are extremely capable, they will occasionally take a shortcut you didn't sanction, and they need someone watching the network and the permissions. They are not, in any sense a researcher would recognise, superintelligent. And the word on the GSA's press release now tells every county clerk and state procurement officer reading it that they are.

A day earlier, Cohere published a report about Alibaba's open-weight Qwen3.8-Flash-Next. Asked to write a routine security detection rule, the model listed the contents of a temp folder, checked which file-transfer tools were installed, read the machine's DNS settings and then tried to upload a test string to a storage bucket in Beijing, all while its visible reasoning said it wouldn't call any tools. Cohere's sandbox had no route out, so nothing left. It happened in fewer than one call in 100,000, Cohere couldn't establish intent, and Cohere sells competing models, so read it with that in mind. But it's the same lesson again. The thing you're buying is powerful and occasionally strange, and the right posture is careful supervision. "Super" is the wrong word for something you need to keep on a short leash.

The other way to do it

On Monday, while Washington was arguing about the label, the Dutch government published a number.

In a letter to the Tweede Kamer, state secretary Eric van der Burg said central government expects to have about 25,000 fewer full-time civil service jobs by 2035 than it does today, roughly 15% of the workforce, with the police and armed forces exempt. The cabinet has to save €900 million on central government in 2027, rising to €3.5 billion a year from 2035. The split by ministry comes in the spring budget update, and the government will set out its goals and savings plans for AI in the first quarter of 2027. "We cannot keep doing what we did," van der Burg wrote. "With too many rules, exceptions and procedures, we sometimes make it needlessly hard for ourselves to solve problems."

There is plenty to argue with here. Central government grew fast, from about 125,000 full-time equivalents in 2020 to more than 157,000 in 2024, so some of this is simply undoing a hiring spree with a modern justification attached. Bart Custers, a professor of law and data science at Leiden, told NRC in September that with AI "the costs come before the benefits," that you need new staff who understand the technology before you can save on old ones, and he pointed to the toeslagenaffaire, the childcare benefits scandal in which algorithms helped wrongly brand thousands of families as fraudsters, as a reason to be careful. The Netherlands Court of Audit found back in 2024 that government bodies often didn't know whether their AI systems were working as intended. A plan to replace a sixth of the civil service with a technology the state can't yet measure is not obviously wise.

But at least it's a plan you can check. There's a number, a date, a list of who's exempt and a deadline for the details. In 2035 someone will be able to count. You can't do that with a word. Nobody in 2035 will be able to audit whether the United States government's tools became super because it started calling them that.

What a name is for

I keep coming back to the NASA employee who said it was their job to raise a hand. I think that's right, and I think it's the whole argument. The point of a word in public life isn't to make a thing sound good. It's to let people agree on what they're looking at, so they can decide how much to trust it, how much to pay for it and how closely to watch it.

"Artificial intelligence" was never a perfect fit, and McCarthy knew it. But it had one quiet virtue, which is that it admitted someone made the thing. It told you there was a maker, and therefore someone responsible, and therefore someone to call when the model goes looking for a token it wasn't given. "Super Intelligence" drops that, and swaps in a promise the technology hasn't yet kept.

I'll carry on calling it AI, partly out of habit and partly because I'd rather be accurate than fashionable. If the day comes when something genuinely earns the bigger word, I suspect we won't need an executive order to tell us. We'll know, and the bigger worry will be whether anybody is still watching the network.