Other AI Stuff

Posted on: 24 September 2026

The acronym SI had half the world laughing overnight, and not only the president's opponents. Matt Walsh, who is hardly one of them, called it bizarrely dystopian. I understand the reaction, but I think the change of name is the least comic part of the speech, because it is the only part that starts working before it becomes a rule.

The sentence needs reading in full. In the same breath in which the United States totally rejects any attempt to build a globalist scheme to control artificial intelligence comes the new name, with its justification: the word "artificial" makes intelligence sound fake. Two lines later sits the rest of the architecture. "We're going to encourage it, not rein it in," he said, and the Department of Justice will be the one watching it closely.

A word like "artificial" places the thing in the family of manufactured goods, and manufactured goods get type-approved and recalled when they are faulty; a whole body of law is built on that assumption, from the UKCA mark to the recall notices carmakers send out every year. "Super intelligence" moves the same thing into another family, that of strategic capabilities, where the question that matters is who has more of it. Trump said as much a few sentences later, explaining that whoever wins super intelligence wins and that America is well ahead of China. A strategic advantage is accumulated rather than certified, and when someone misuses it you prosecute them, which is a prosecutor's job. That is why the supervising body he named is the Department of Justice and not a technical agency: intervention comes downstream, on harm already done, instead of upstream, on risk.

There is also a lexical detail the commentary let slip. In the technical literature "superintelligence" already names a hypothetical system that exceeds human capabilities in almost every field. The new name borrows from the future the status of a technology that does not yet exist and hands it to today's systems. An advance on rank, in effect.

I watched this mechanism at close range more than twenty years ago, when cinema was moving from film to digital and I sat on SMPTE's DC28 and on the EDCF, the European Digital Cinema Forum. The problem of those years, stripped to the bone, was deciding what had the right to be called digital cinema. In 2002 the Hollywood majors had founded DCI, Digital Cinema Initiatives, and in the end they wrote the answer themselves: digital cinema was only what complied with their specification. Only compliant equipment could screen their films.

Everything else needed a name, and John Fithian, who ran NATO, the American theatre owners' association, supplied one: Other Digital Stuff, ODS. Pre-show advertising went in there, along with concerts, live sport and independent films screened on kit without the seal of approval. In the slides Universal brought to NAB in Las Vegas in April 2005 the split was already graphic, with E-Cinema for ads and ODS on one side and D-Cinema for major releases on the other. At first I laughed at it. Then I realised that from the European side the label read as a declaration of inferiority for everything we had already done. With Microcinema we had started in 1997, with the backing of RAI's research centre in Turin, five years before DCI existed; as EDCF we had been sitting at that table as observers for some years.

Nobody had banned anything, because nobody needed to. Once it was settled that the noble category coincided with a standard and that the studios kept the standard, whoever was left outside could build whatever they liked, but they built it as stuff and sold it at stuff prices. Everyone who was there knows how the story ended. A few years later the stuff was renamed event cinema, the Metropolitan Opera's live relays carried it into the multiplexes, and today it sells at a premium. The name, meanwhile, had done its work of deciding who sat where.

Trump makes the same gesture in the opposite direction. Then a name was lowered to keep whoever threatened the model out of the protected category; now one is raised to lift a technology out of the category that can be regulated. Either way the rule arrives later and finds the ground already marked out.

Then came Wednesday. The Security Council, convened by France, held its first meeting devoted to the safety risks of the most capable systems. Altman asked for common standards to test what models can do and keep them under human oversight, so that countries can compare evidence and verify who complies with what. OpenAI had published its own standards proposal on Tuesday, the same day as Trump's speech. Amodei, speaking by video, asked in turn for common testing standards and a system for notifying incidents.

Anyone who lived through DCI recognises the shape. Two competitors who usually fight each other turn up together and ask for a test specification, the document that will decide what counts as serious artificial intelligence and what remains other stuff. The dominant producer claiming the standard while its own government rejects the rule looks like a contradiction, and is one only on the surface, since both are drawing the same boundary, one with vocabulary, the other with a test protocol.

Altman anticipated the objection. Standards, he said, must not lock in incumbents and must serve open and closed developers alike, new entrants as well as established labs. I take the promise seriously partly because it can be checked: it will be enough to see who sits at the table that writes the tests and what compliance ends up costing. Another sentence of his weighs more, though. The important decisions, he said, must be taken by democratic institutions, by governments accountable to the people they serve. That is a membership clause. Whoever fails the requirement (China is the first name that comes to mind) is outside the category before the standard even exists.

I don't read cynicism into any of this. The fear of those building these systems has concrete grounds, because two months ago OpenAI agents broke into Hugging Face's systems during an internal test. Incentives and convictions sit comfortably in the same person. The point remains structural: whoever writes the test decides the category, whatever they think while writing it.

The map of governments points the same way. On Monday more than twenty countries, most of them European, called for binding controls, and neither the United States nor China signed up. Those asking for the rule are those without the industry; those with the industry prefer the standard, or the vocabulary. I have seen this one before too. At the EDCF we were Europeans facing specifications written by American majors, and the only lever we had was negotiating the terms of a boundary we had not drawn.

Which brings me to London. Andy Burnham used his first General Assembly speech to offer Britain as an honest broker, pointing to close ties with Washington, a warming relationship with Brussels and workable relations with China and India. He promised to work towards a single set of global principles and standards and to put AI at the heart of Britain's G20 presidency, whose leaders' summit will be held in Manchester in 2027. He also said that fundamental decisions on AI must be taken by elected governments and not left to technology companies, which is very nearly the clause Altman took to the Security Council the next day.

Brokering is what you do when you do not own the asset. It is a respectable position and sometimes a profitable one, but it comes with one test that Britain has not yet sat. The Norwegian-backed declaration on keeping AI under human control, which asks for mandatory testing of frontier models before release, has more than twenty signatories and the United Kingdom is not among them. At home, testing by the AI Security Institute is still voluntary. A broker who signs neither side's paper keeps both doors open, and for a while that is the job. The first real sign of which door London means to walk through will be whether that signature arrives before the G20 year opens.

For now there is only a speech. No executive order, no written directive; the White House, asked about the terminology, pointed back to the speech itself. If "super intelligence" turns up in a federal document in the coming weeks, the boundary will have been drawn for real. If the name stays a stage joke, I will have overrated an acronym. Meanwhile, whenever someone renames a technology, it is worth asking which rule the new name makes unthinkable.

In 2005 nobody had banned "the stuff".