Op-Ed: Is the whole idea of AI sovereignty incredibly naïve or absolutely essential?

“AI sovereignty” is a hot issue. It boils down to the idea of not being dependent on external AI platforms and technologies. It’s seen as a survival option as much as a more practical way of managing AI in general. Geopolitics, market forces, and technological realities make this a very complex issue.

At the geopolitical level, it’s become another “US or China” scenario. The recent blocking of external access to Anthropic’s Claude flagship AIs is the truly brutal sticking point for the pro-sovereignty arguments. Users outside the US were effectively locked out of frontier-level AI overnight.

At this level, sensitivities are very high. It doesn’t take much imagination to see what the lockout could have done in any real-world operational environment. Nor was anyone even slightly impressed with the “national security” rationale for the shutdown. Claude is supposed to be the finder and fixer for system vulnerabilities, so they shut it down?

National responses have varied. The response to AI sovereignty has been a bit confused but with common objectives:

Japan is working on a sovereign AI model.

Australia is in a state of understandable but annoying total indecision about AI sovereignty at its most basic level.

Canada is taking an appropriate if demanding stance on managing AI at the sovereign level, in the face of an even more demanding situation with the US relationship. The Anthropic shutdown emphasized the risks of dependency on external sources.

The EU has a broad theory of AI sovereignty, but implementation is likely to be costly and difficult in practice.

Data sovereignty is a common and critical sticking point in the broader arguments for sovereignty, underpinning the case for national jurisdiction.

Market forces are generating as much turbulence and irritation as the AI. This market is extremely neurotic and very changeable. The commercial AI services market is in total skank mode, selling whatever is available to anyone and anything worldwide. Sovereignty clearly isn’t an issue to the companies selling the services. This tidal surge of external AI capacity crosses borders with ease. How much sovereignty can you actually have?

Technologically, the “breakthrough every second” scenario is charging ahead, hype or not. Just keeping up with it is hard work. In practical terms, AI has to deliver value, not press releases. That’s a big obstacle course for those deploying and adopting AI in a blurry mess of expectations and actual capabilities.

A tough and utterly thankless geopolitical environment

Lumbering along like a slow-moving shopping trolley in the center aisle is the US and China dichotomy. You’re being asked to pick a side between two antagonistic superpowers, and you have few options. Sovereignty looks like a much safer approach, for good reason.

There’s no way of knowing or even guessing which way this dichotomy will roll. Add to this the fact that nations have to govern AI in their own jurisdictions, and “administrative straitjacket” is a euphemism. In this scenario, sovereignty is seen as essential.

A less obvious but inevitable problem is that nearly all AI evolution is effectively coming from the US or China. Nobody else is getting a word in edgewise. Nobody wants to get involved in the various conflicts.

This also locks out external AI development, and forces nations to follow whatever whimsical trail of breadcrumbs comes from the two primary sources. It’s a much bigger problem than it looks and can be seen as a limiting factor on healthy AI development.

The emerging reality of amorphous AI networks

AI creates pan-national environments and ad hoc networks simply by plugging into the Cloud or the internet. A single transaction can be global. Daily operations generate a lot of transactions. Demand for services alone creates instant diversity and demands. This shapeless, intense state of billions of transactions undercuts theoretical sovereignty in real time.

There’s something incredibly unrealistic about assuming that this definitive state of AI operations can be conducted in a pristine sovereign environment. The legalities create issues for administration through multiple jurisdictions.

This is also the key argument for sovereignty, trying to clarify the legalities. The argument for sovereignty also equates to rules governing AI behaviour and illegalities, the vast scope of which could be a PhD subject for the future. This is the truly non-negotiable part of the sovereignty argument, and it’s anything but naïve. Having two diametrically opposed primary sources of AI makes it very difficult to get consistent rules in place.

Can AI be governed without sovereignty?

The short answer to that question is an unequivocal “No”. Any sort of legal governance must be based on jurisdictional authority, and the default authority is national sovereignty. An international version would have to be something like trade treaties, based on agreements between individual nations. That would take a long time to implement between the parties. An international charter would be nice, but probably very difficult to enforce in a crowded environment of disputes. Sovereignty is also the only real way of enforcing rules at the national level.

The theory of sovereignty seems naïve in such a complex mix of issues, but in practice, it’s the only visible option for any real governance of AI. You can’t “assume” that everyone will be a good actor. You can’t assume that global organized crime won’t exploit AI because it’s already doing that.  Again, sovereignty is the default for law enforcement.

The bottom line is this:

You can’t have a whole class of technology that is beyond the law.