Renaming AI to Superintelligence Does Not Make It Safer

A meeting between a head of state and the leaders of the largest AI companies is not a safety mechanism. It is a photo of one. The image is reassuring. The substance, so far, is not.

Start with what can actually be verified

I will be direct about the limits of this piece. I was not in the room, I have not seen a verified transcript, and I am not going to describe conversations I cannot source. Anyone writing a detailed account of who said what at a closed meeting between a president and a handful of lab chief executives is guessing, and guessing dressed as analysis is worse than silence.

What can be discussed honestly is the pattern around such meetings, because the pattern repeats across industries and decades. When public anxiety about a technology grows faster than the state's ability to govern it, governments and the industry appear together. The appearance itself is the product. It signals that adults are handling the matter, without creating a single obligation that anyone can be measured against later.

That is the part I want to argue with. Not the meeting. The absence of anything enforceable coming out of it.

Why the word swap matters

The move from calling this technology "AI" to calling it "superintelligence" looks small. It is not. Words set the regulatory clock.

From today’s systems to tomorrow’s story

"AI" describes software that is already deployed. It reads applications, scores risk, routes cases, screens candidates, drafts medical and legal text, prices services and increasingly takes actions through tools and APIs without a human between the decision and the effect. That is an object a regulator can inspect this quarter.

"Superintelligence" describes an arrival. It points at a future entity with its own agency, and it invites the public to imagine a film rather than an audit. You cannot write an inspection checklist for a story. The practical effect of the rebrand, whatever its intent, is to move the conversation away from the systems that are causing real harm now and towards a horizon that is conveniently out of reach of present-day law.

I will say plainly that I read this as a framing choice that serves the people doing the building. If your product is under scrutiny, the most useful thing you can do is to make it sound like it does not exist yet.

The race will not slow itself

The most common argument against hard regulation is that it is unnecessary because the serious labs are responsible and will moderate their own pace. I do not believe this, and I do not think the people making the argument believe it either.

Two competitions make self-restraint irrational. The first is between states. The United States and China are not going to decelerate out of mutual courtesy, because capability here is read as economic and strategic position. The second is between companies. No lab holds a frontier model back for six months while a rival ships, because the cost of caution is paid immediately in market share and the benefit is diffuse, delayed and unprovable.

That is not a moral failure of individuals. It is the structure. When caution is expensive for whoever goes first, you do not get caution from goodwill. You get it from a rule that applies to everyone at once, or you do not get it at all.

What real regulation would look like

If the goal is public safety rather than public calm, four things need to exist. None of them are exciting, which is probably why none of them are announced at meetings like this.

A published standard, not a private pledge

There should be one written safety and governance standard for any organisation training or deploying systems aimed at general capability. Published in full, in plain language, so that a journalist, a parliamentarian or a citizen can read it and know what is required. Voluntary commitments fail because the commitment, the measurement and the grading all sit inside the same company.

Regulators with the right to inspect

A regulator that receives a slide deck is not a regulator. It needs named, funded authority to enter, to test models directly, to see evaluation results including the failed ones, to review incident logs and to look at what the system actually does in production rather than what the documentation says it does. Audit without access is theatre.

A monthly public report

This is the piece I would argue hardest for. Each month, the regulator publishes a short, standard report: which organisations were assessed, which requirements they met, which they failed, what incidents were recorded, and what enforcement followed. Monthly, because the release cycle of this industry is measured in weeks. Public, because a report that goes only to a ministry becomes a negotiation. Standard in format, because comparability is what turns disclosure into pressure.

A global floor

Finally, this has to be international. A country that regulates alone hands its labs a disadvantage and hands the work to a jurisdiction that does not. So the standard needs treaty-level agreement covering the United States, China, the European Union and every other state hosting frontier development, with mutual inspection and consequences for non-participation. Whether that sits under an existing body or a new one matters less than whether it binds.

The honest objections

The first objection is that global agreements are slow, weak and frequently ignored. True. Nuclear safeguards, aviation safety and financial reporting are all imperfect, politically contested and sometimes evaded. They are also the reason those industries are survivable. Imperfect enforcement beats a press conference.

The second objection is that regulation slows innovation. Sometimes it does. In my own work on enterprise and public-sector AI systems, the controls that slow a project down are also the controls that let it go live at all, because nobody signs off on an automated decision they cannot explain or reverse. Clear published rules are usually faster than unclear ones, because teams stop guessing what will be acceptable.

The third objection is the one I accept: nobody knows the timeline. I am not predicting a date, and I am not claiming a machine will develop intentions. My concern is narrower and more likely. Within a few years, highly capable systems will be making consequential decisions at national scale, inside institutions that cannot fully explain them, with no external body empowered to check them. That is a genuine risk to people, and it does not require science fiction to arrive.

The takeaway

Renaming AI to superintelligence changes the headline, not the exposure. Safety is not a vocabulary problem or a sentiment problem. It is a problem of published standards, real inspection rights, monthly public reporting and international force.

Until those four exist, every meeting of this kind should be read as communications, not governance. If you were designing the monthly public report, what is the first line you would insist it contain?