The Off-Switch Problem: AI Inclusion Through Interoperability

Digital and AI policy experts increasingly recognize that access to artificial intelligence is unevenly distributed, and that closing the gap between those who can use these tools and those who cannot is one of the defining development challenges of the decade. There is less recognition of a quieter vulnerability beneath that gap: the terms of access can change overnight, by decision of a government in which users have no voice.

That vulnerability became concrete this month. A United States export-control directive, citing national security, suspended access to Anthropic’s most capable models, Fable and Mythos, for all foreign nationals. The technical details of the ban are important, but for users, the lived reality is more significant. A fraud-detection team at a fintech, an agricultural advisory startup, a public-service pilot building a citizen helpline: all of them were building on these tools on a Friday and locked out by Saturday. Not because they could not afford the service. Not because the infrastructure failed. Because of a policy choice made by someone else.

This directive highlights something the inclusion agenda has tended to treat as settled. The movement to insert AI into everyday workflows across the Global South has framed the divide largely in terms of affordability and infrastructure: the cost of access to AI tools, the compute to run them, the skills to use them. The Fable episode adds a third axis, arguably the most destabilizing: revocability. You can pay for access, embed it in a live service, train staff around it, and still lose it to a directive you never saw coming. Put differently: these users had access, but they never had sovereignty. And access, it turns out, can be switched off arbitrarily.

The instinctive response is to reach for ownership. If dependence on a foreign model is the problem, the reasoning goes, then build your own national champion, sovereign stack, or public money for domestic capacity. This instinct is understandable, but it is largely a trap.

As David Eaves argues, owning every layer of the technology stack is not possible, even for the United States, and building a  “sovereign stack” carries real and lasting costs. A domestic champion may protect you from a kill switch, but it will not help businesses scale, and it quietly taxes your own economy for decades. The point generalizes beyond the cloud to the model. A sovereign AI model that locks in its users just as tightly as the foreign one it replaced has not solved the dependence problem; it has merely changed the flag on it. This is sovereignty theatre: the appearance of control without the substance of agency.

The substance lies in a less headline-friendly word: interoperability. The real goal for those pursuing sovereignty, as Eaves and Mike Bracken have put it, is agency, to reshape the market so its layers behave like a utility: interoperable, substitutable, something every country, company, and individual can draw on and, crucially, switch away from. Eaves makes this case for cloud, the substrate that compute and AI run on. The same logic applies one layer up. If models, the tools that use them, and the data they hold can be moved between providers, if an application built on one system can run on another without being rebuilt, then no single directive in any one capital can strand the users downstream. Substitutability is what turns a kill switch from an arbitrary death sentence into an inconvenience.

This is why interoperability is not, at the core, a technical matter. It is a governance question wearing technical clothes. Markets do not commoditize themselves; incumbents have every incentive to make switching costly, and left alone they will. Interoperability must be required: written into procurement rules, competition policy, data-portability standards, and the conditions attached to public money and partnership agreements. The decision to mandate open standards, to refuse contracts without exit rights, to treat lock-in as a cost rather than a convenience are all governance choices, made by states and standards institutions. Interoperability is the technical form that governance takes, and governance is what makes interoperability more than an aspiration.

Seen this way, the three regions that are now hedging against dependence are really making bets about governance, not hardware.

Europe’s Tech Sovereignty Package is the most muscular effort, but it risks confusing the goal. Much of the political energy goes into ownership: national clouds, domestic champions, the quiet pursuit of “buying European.” The more important choice  is whether Europe builds a utility or merely a stack. A European alternative that is as hard to leave as the American incumbent protects against the remote kill switch, but reproduces the tax.

Canada provides a strong illustration of the trap of dependence. As Eaves and Curtis McCord document, three US firms hold around 85 percent of the Canadian cloud market, and standing up domestic providers risks trapping users in what they call  “maplewashed dependencies” unless those providers are required, by rule, to be interoperable. As a close US ally, Canada also assumed proximity to Washington was a form of cover. The Fable directive, which caught both enemies and allies in the same net, is a reminder that substitutability is a more durable protection than friendship.

Africa shows the more interesting answer, one with interoperability at its core. Across the continent the response to external control has been not retreat but construction, AI that is built to be shared.

Governance connects these layers: the African Union’s Continental AI Strategy, the digital provisions of the continental free-trade area, and the practice of South-South exchange are the mechanisms through which shared standards and collective bargaining power are built, and what makes interoperability a regional asset rather than a national hope.

These mechanisms are the durable answer to the off-switch problem. They show how  interoperability can be governed into existence. Through diversified supply across proprietary, open-weight, and regional options; through shared and substitutable compute; and through continuity, exit rights, and open standards written into the contracts and policies from the outset. This is what intelligent dependence looks like. The goal of sovereignty efforts should be agency rather than an illusory independence: the institutional capacity to choose what to depend on, under what conditions, and for how long, and to change your mind without having to rebuild everything. The path to real sovereignty, if the word is to mean anything, runs through the commoditization of stack elements, not only through building the stack. The aim is to be sovereign, but not alone.