Malta - OpenAI: When a State Becomes a Customer

The agreement between the Maltese government and OpenAI must be read on two levels: the official one, made of digital literacy and innovation, and the deeper one, which raises legitimate questions about technological lock-in, cultural influence, and infrastructural dependence. Because the real question isn't whether a free year of ChatGPT Plus is good for Maltese citizens. The question is what happens after that year, and who is truly in control.

On May 17, 2026, Malta became the first country in the world to sign a national agreement with OpenAI. The content, in its official formulation, is simple: every Maltese citizen or resident aged at least 14 can obtain a free one-year subscription to ChatGPT Plus. Not unconditionally, however. To access the benefit, one must first complete a course of about two hours called "AI for Everyone," developed by the Malta Digital Innovation Authority in collaboration with the University of Malta, available on the ai4all.gov.mt platform in Maltese and English, without requiring prior technical skills.

Those who do not want ChatGPT Plus can alternatively choose Microsoft 365 Personal Copilot, which makes the initiative formally pluralistic, even if Microsoft's presence is not entirely neutral in the reasoning that follows. The offer also extends to Maltese citizens residing abroad, a detail that expands the scope of the initiative well beyond the physical borders of the island.

Behind this lies a specific economic and political context. The agreement is part of the 100-million-euro digitalization plan announced in the 2026 Maltese Budget, which covers artificial intelligence, cybersecurity, robotics, and emerging technologies. Deputy Prime Minister Ian Borg defined the program as a concrete commitment to ensure that no one is left behind in the AI era. Minister for the Economy Silvio Schembri stated that the initiative will transform artificial intelligence from an "unfamiliar concept" to a "practical tool for families, students, and workers." George Osborne, head of the OpenAI for Countries division, praised Malta as a country that "leads Europe and the world in bringing AI to all its citizens."

Some things, however, have not been made public. The financial terms of the agreement remain confidential: it is not known how much the Maltese government is paying, nor according to what formula. The nominal cost of ChatGPT Plus is 20 dollars a month, and if all eligible Maltese residents signed up, the theoretical value would exceed 130 million dollars annually. The real cost for OpenAI is presumably much lower; digital distribution infrastructure does not scale linearly like physical infrastructure, but without public data, any calculation remains speculative. What is certain is that the lack of financial transparency is the first element that invites interpretative caution.

Who Gains What: Stated Benefits and Real Interests

The advantage for Malta, in the official narrative, is clear: to accelerate the digital transition of the population, reduce the gap between those who know how to use AI and those who do not, and position itself as an innovator country within the European Union. It is a legitimate goal, and the chosen mechanism—first training, then the tool—is methodologically more sensible than blind distribution. Teaching people not only how to use a tool but also what it cannot do is a commendable policy choice.

On the political front, Malta is also playing an image game. The island has a long history of strategic positioning as a hub: financial in the nineties, crypto-friendly in 2018, and now potentially a European laboratory for the public adoption of AI. Being "the first country in the world" to do something has a signaling value that goes beyond the merits of the initiative itself. It is both domestic and international soft power.

The advantage for OpenAI is equally concrete but operates on different levels. The most obvious is the growth of the user base: bringing tens of thousands of new users to the platform, even temporarily, means usage data, behavioral feedback, and retention metrics. The "OpenAI for Countries" plan, which includes agreements already signed with Greece and Estonia, reveals a systematic strategy of institutional penetration into European markets. This is not technological philanthropy: it is commercial development wrapped in public interest, a model the tech industry has known well at least since the days when Google distributed Chromebooks in American schools.

George Osborne is a non-negligible detail. Former UK Chancellor of the Exchequer, now head of government relations for OpenAI: his presence signals that the company is building lobbying capacity and institutional relations at the highest level, not just commercial agreements. It is the type of move that recalls, on a smaller scale, Microsoft's strategy in the years when Bill Gates became a privileged interlocutor for world governments on education and health. The question is not whether this is legal—it is—but what the nature of the power being built is.

The Unspoken: Three Levels of Dependence

Here the reasoning becomes more complex, and it is worth distinguishing precisely between what is established, what is plausible, and what is speculative.

The first level is economic. Technological lock-in, a term borrowed from business literature, describes the condition in which the cost of abandoning a tool exceeds the cost of continuing to use it, even when better or cheaper alternatives exist. In professional software, this is a well-documented mechanism: SMEs that built their processes around certain proprietary management software in the nineties discovered a decade later that migrating was more expensive than staying, even when paying increasing license fees. The analogy with AI is not identical; language models do not have the same type of data dependency as structured data, but the underlying logic is similar. If Maltese students, teachers, and public employees spend a year building habits, skills, and workflows around a specific tool, the real cost of changing it afterward is not zero. It is composed of retraining, procedure adaptation, prompt and automation rewriting, and loss of familiarity. None of these costs are catastrophic individually, but summed up, they create inertia.

The second level is organizational. Institutions—schools, public offices, universities—do not use tools individually. They integrate them into procedures, workflows, and evaluations. If a Maltese high school adopts ChatGPT Plus as a teaching support tool for a year, at the end of that year it has teachers who have structured their lessons around certain outputs, students who have developed certain expectations, perhaps even evaluation rubrics designed for certain types of responses. Changing models is not like changing browsers: it is like changing the working language in the middle of a project. Not impossible, but costly in an organizational sense. And the longer the adoption lasts, the more it settles.

The third level is the one that deserves the most attention because it is the least visible: it is cultural. A language model used massively in an institutional context is not a neutral tool. It has response styles, preferred registers, and implicit hierarchies of what counts as a "good response." OpenAI's Collective Alignment project, a survey conducted on over a thousand people in 19 countries to understand how an ideal model should behave, revealed an 80% alignment with the company's internal guidelines. This result can be read in two opposite ways: either OpenAI has done an excellent job of aligning with global values, or the detected global values were measured with tools designed by those who already had a preferred answer. Both are likely partially true, and this is exactly the point.

Ivan Illich, in his 1973 work Tools for Conviviality, a text that has not lost an ounce of relevance, distinguished between tools that amplify human capabilities and tools that replace them, creating dependence. He did not argue that the latter were inherently evil, but that their massive and uncritical adoption tended to change not only behaviors but the categories with which people think about problems. A language model that becomes the default filter through which students and public employees process information, draft documents, and make decisions is not just a productivity tool. It is a shared cognitive model. And shared cognitive models, as anyone who has studied the sociology of knowledge knows, are not neutral.

The Data Issue and the GDPR Knot

There is a technical element of the agreement that deserves separate attention. Access to ChatGPT Plus for Maltese citizens is not anonymous: it is verified via the national digital identity system. This means that users' interactions on the platform are linked to their civil identity, not to an alias or a generic account.

Researcher Miranda Bogen raised this concern directly, observing that collaborating with nation-states poses serious questions about how to protect human rights against potential government requests for user data. The concern is not hypothetical: the intersection between AI usage data, verified identity, and national government is a configuration that lacks sufficient documented precedents to know how it behaves under stress—in the event of a change of government, a judicial request, or a unilateral change in terms of service.

Malta is a member of the European Union and therefore subject to the GDPR, which offers significant protections. However, the specific terms of how data generated by Maltese users is treated, where it is stored, for how long, and for what secondary purposes, have not been made public. In the absence of this information, it is not possible to assess the real risk, and this is in itself a transparency problem that should be addressed before large-scale implementation, not after.

The European Precedent and the Geopolitical Reading

The Maltese model is not a local story. It is a proof of concept for something bigger. If the experiment works, if course completion rates are high, if user satisfaction is measurable, if the political narrative holds, other European governments will have a case study to imitate or challenge. OpenAI knows this, and "OpenAI for Countries" is exactly the program designed to replicate this type of agreement on a continental scale.

This raises a governance issue that goes beyond Malta. The European Union has been building a regulatory framework for AI for years—the AI Act, discussions on algorithmic transparency, the Commission's public consultation on rules for interacting with AI systems—with the explicit goal of maintaining democratic control over high-impact technologies. An agreement like the Maltese one does not necessarily violate this framework, but it occupies a space that the framework itself had not foreseen: that of the State that does not regulate a technology from the outside but adopts it as infrastructure from the inside.

The difference is not subtle. A regulator imposing transparency requirements on a private company maintains a third-party position. A government that distributes subscriptions to that same company to its citizens becomes a customer, with all the relational constraints that entails. This is not necessarily wrong—governments have been purchasing private software for decades—but it is a structurally different relationship and deserves a proportionate level of public scrutiny.

There is also a dimension of soft power worth mentioning. OpenAI is an American company, headquartered in San Francisco, subject to US laws and the political pressures of the context in which it operates. When a European state adopts its infrastructure as a public policy tool, it is also making an implicit geopolitical choice. Not necessarily a wrong one, but it is a choice that should be conscious, not incidental.

It Is Not Inevitable

It would be wrong, however, to conclude that the Malta-OpenAI agreement is inevitably destined to produce monopoly, dependence, or cultural colonization. The risk exists, but its materialization is not written in the DNA of the initiative: it depends on how it is managed over time.

Lock-in can be reduced with specific policy choices: imposing open interoperability standards, ensuring data portability, requiring by law that public institutions use multiple models in parallel, and investing in critical training for AI use that explicitly includes understanding the limits and biases of any single tool. These are not utopian measures: they are digital procurement techniques that some European administrations already adopt, at least in part.

Competition between models still exists and is genuinely intense. Anthropic, Google, Mistral, open-source models: the ecosystem is plural, and nothing prevents Malta from adopting a multi-vendor strategy as early as next year. The risk is not "the end of choice," but something more subtle: the progressive erosion of real choice when habits, training, and procedures settle around a single provider and the cost of changing becomes too high to be practiced, even when theoretically possible.

The television series Halt and Catch Fire, four extraordinary seasons on the hidden history of the computing revolution, told better than many essays how the technological choices of an era are never purely technical. They are choices about who controls infrastructure, who writes the grammars of work, who decides which questions are normal to ask and which are not. Malta is not deciding the fate of European AI. But it is writing a page of a manual that others will read carefully. It is worth ensuring that page contains not only the benefits of the agreement but also the questions to which the agreement has not yet responded.