Some More Thoughts on AI Sovereignty.

AI sovereignty is having a moment. Honestly, it’s been having a moment for a while now, and for good reason! Across Southeast Asia (and, frankly, most places that have discovered the joy of strategic anxiety), ‘AI sovereignty’ is increasingly treated like one of those policy phrases that come preloaded with heaps of moral charge—and specifically, virtue. Like ‘human-centric AI’, or ‘responsible innovation’, or even something as straightforward sounding like ‘trust’ (seriously: emphasis on ‘sounding’ straightforward). These are words you’re not supposed to disagree with, unless you spend your spare time twiddling a mustache while tying young women to railroad tracks.

And, let me be clear out of the gate: I am not against AI sovereignty.

I am challenging the assumption that AI sovereignty is unambiguously positive, politically neutral, culturally harmless, or automatically anticipatory just because it is framed as defensive in a world where tech-sharks circle the weak. The risk here is not merely that sovereignty projects fail or otherwise fall short. The risk is that they succeed, and succeed at the wrong things. See, there’s a potentially thin line between being defensive and being protectionist. The latter may very much serve to quietly harden a particular version of ‘the nation’ or ‘the people’ into a narrowly-defined construct that feels and appears natural, inevitable, and is, suddenly, technically validated.

Recently, I had the honour and the pleasure of speaking at a workshop in Kuala Lumpur, on Sovereignty in AI(-generated) times: Exploring artificial intelligence, government, and power in Asia at Monash University, KL, and this post incorporates the material I presented at the workshop, alongside some post-workshop reflections, as well as conversations with friends and colleagues. In many ways, this represents a work-and-thinking-in-progress, which is why this is on Substack and not being submitted to a journal. In other words, every word written here is incomplete, and these are all conversations I am looking forward to continuing.

At the workshop, I spoke of the slide from AI sovereignty towards what I’ve called AI nationalism. This is a concept I have been playing around with for a few months now, and it’s not one I see being floated much elsewhere. This slide—characterised by inclusive, otherwise reasonable, and even overdue sovereignty talks begins drifting into something more inward-facing, exclusionary, and culturally reductive—is not unique to AI, but, as with so much else, AI systems present yet another avenue for these kinds of reductive exclusionary projects to play out, whilst also potentially supercharging them.

I am also not necessarily only talking about ethnic nationalism here; these developments may very well be far more subtle than jackboots and flags (though I am sure these will be available in a premium tier somewhere). I think there is a risk for something far more subtle to emerge: the kind of nationalistic project that is emergent through design defaults, governance shortcuts, and branding decisions. The kind where ‘our values’ becomes a single set of values, and ‘our language’ becomes the language, and cooperation becomes something we only do when it flatters our self-image.

Here’s the guiding question I want to table, in other words:

Through what mechanisms does inclusive AI sovereignty slide into exclusive AI nationalism?

And I want to qualify this further by saying that I am not looking at making any moral accusations here. Rather, I am approaching this as a governance problem, and governance problems rarely politely confine themselves to conference panels, or clearly identify themselves when requested.

And this is exactly what I think this (often implicit) slide towards exclusionary protectionism is: a governance challenge that is inherent to an otherwise important tool many countries ought to otherwise engage with.

So, why am I even talking about this? Almost everybody working in this space here in Southeast Asia (and, I am sure, elsewhere) agrees that AI sovereignty is important, but there is not even a shadow of agreement on what it actually means.

Over the last eighteen months or so, I have spent substantial time working on cultural alignment, mostly with engineers and developers. This represents the micro-level version: how is data collected, how do you identify harmful model behaviour after specific use-cases and contexts, and how do you avoid building systems that treat whole populations as an edge-case, a stereotype, or both? It’s messy, with clear doctrinal fault lines being exposed between us bleeding-heart humanities and social sciences folks who just want to make the world a better place, and the stone-cold engineers busy playing god. This work is also deeply necessary. It has also shown me more than ever why governance is required: not to ruin anyone’s fun, but to answer questions that go beyond optimising the loss function, and thus engineers are unprepared to effectively approach (though it doesn’t necessarily stop them from trying).

Then, alignment started showing up elsewhere: in policy roundtable discussions, in national strategies, in the glossy language of ‘values’ and ‘culture’ and—you guessed it—’sovereign AI’. When alignment becomes a state-level aspiration; when ministries want to operationalise at scale, something changes. Suddenly, you’re not asking how to align a model. You’re asking:

• Who gets to define what we’re optimising for?

• Who gets to be included?

• Who doesn’t?

• And why are we even bothering?

This is the nexus I am speaking of specifically here: AI sovereignty and how it intersects with capital-C Culture, and the mechanisms that can lead it down less than savoury pathways.

A very valuable insight from a recent email exchange with a colleague working in the same space as I, but in a different country, is, simply put, that Sovereign AI isn’t one thing—and pretending that it is, creates a whole slew of problems. Sometimes, sovereignty is framed as full-stack autonomy: you own the model, the data, the compute, the weights, the whole stack. We can call this sovereignty-as-production.

Other times, sovereignty is framed as institutional capacity: the ability to evaluate, audit, govern, and ultimately take responsibility for AI systems and enforce regulatory boundaries—even when those systems are built elsewhere. This matters because it changes the narrative. You don’t have to insist on ‘home-grown’ models to claim sovereignty if your institutions can constrain externally built models. This is sovereignty-as-a-regulatory-and-evaluative-capability as opposed to a from-farm-to-table supply chain.

Here is where things get spicy, though: in many contexts, a change in political leadership or narrative prioritises shifts in the rhetoric from ‘world’s best’ to ‘independence’—and with independence, a whole new logic emerges, not particularly different from military self-reliance or critical infrastructure security. Crucially, independence implies boundaries: it necessitates an inside and an outside. It also potentially implies suspicion, or pride; a national narrative, credibility, and, a performance of, seriousness.

Then you get the debates that come with all this baggage:

• Is a model sovereign because it’s ‘home-grown’?

• What even counts as a home-grown?

• Does it have to be trained from scratch with randomly initialised weights?

• Is heavy use of open-source components compatible with sovereignty, or is this ideological impurity?

• Are we measuring sovereignty by technical genealogy or by governance capacity?

Two things are noteworthy here. Firstly, these are not abstract questions. Their answers shape funding decisions, eligibility rules, and who gets to be crowned as The National Model. And because we can’t have nice things, there is also a more cynical read floating around policy circles: that sovereign AI narratives drive a massive compute race, accelerating GPU demand, and—surprise—benefiting global infrastructure suppliers.

Secondly, taken as a whole, these questions do not speak to sovereignty as some binary characteristic that elusively exists out there, in the natural world. Instead, what we get here is sovereignty as an ideological construct, contingent on answers to questions that, on their own, do not exist in a vacuum, either. As we say in STS-circles, there is no view from nowhere, and the deeper we seem to dig here, the more subjective it all gets.

Whether or not you fully buy this interpretation, it’s at least worth noting that sovereignty discourse can align very neatly with global AI capitalism. Equally, things like national pride is a strong motivator. It gets budgets approved. It gets data centres built. It makes it easier to justify billion-dollar infrastructure plans as existential necessity rather than industrial strategy. So yes: sovereign AI can be about self-determination. It can also be about a politically useful story that happens to require a lot of NVIDIA hardware. And it can be about protectionist narratives.

All three can be true—though, perhaps not all at once.

I realise I have been spending a lot of time qualifying the background to this discussion—this is because everything surrounding AI sovereignty is deeply contested, so I think this background matters to see how I approach the discussion. With this in mind, let me put my cards on the table. Firstly, let’s look at the typical framing of why sovereignty is necessary in places like Southeast Asia.

• Silicon Valley Imperialism is nothing new.

  Simultaneously, Silicon Valley’s dominance over digital infrastructure is not new (shout-out to Erin McElroy’s great book). What’s potentially new is how mask-off some of it has become. The result is that sovereignty looks not just appealing, but as a strategic and national necessity.

• AI Sovereignty is framed as defensive, emancipatory, and overdue.

  Data is increasingly framed as a strategic resource; capacity to use it becomes equally strategic. You can see variations of this everywhere: local compute, local data, local models, ‘our own stack’.

  It’s not hard to sympathise, either. The belief that American tech corporations will spare a single piece of a single shit for Laos or Timor Leste is not a belief held by anyone who has ever watched American tech corporations behave like American tech corporations.

• Sovereignty discourse is typically framed as a reasonable reaction to the above.

  The world is being reorganised whether we like it or not. Strategic autonomy is not guaranteed. The US is (phrasing this diplomatically) less committed to predictable partnership than it once was (pretended to be??). As a result, sovereignty projects are presented as long overdue.

  Again: fair, but it equally doesn’t mean that this is the default position of all actors in the space, nor does it necessarily mean that the results of all such policies will play out as intended.

All of the above tracks. I cannot emphasise enough how my position here is not anti-AI sovereignty-as-such.

Yet, there are friction points.

See, sovereignty is not a neutral container you pour policies into. It’s a boundary-making object. Even when framed as ‘merely’ defensive and with the explicit aim of being inclusive, it necessarily always involves decisions about:

• inclusion and exclusion,

• authority

• legitimacy

And here is the bit that is often conveniently forgotten: sovereignty isn’t only about drawing lines. It is equally about enforcing those lines. Thus, the moment we start talking about ‘sovereign AI’ or ‘AI sovereignty’, we’re also talking about:

• Who gets to be represented

• Who gets to decide who gets to be represented

• and, by extension, whose existence becomes administratively awkward

This is where, conceptually, AI sovereignty becomes to circle the drain towards AI nationalism—not necessarily as intent, but as effect.

Everyone agrees that the concept is contested—and anyone who says otherwise is trying to bamboozle you. So, when I use it here, I effectively mean the following, cobbled together from the above discussion. I also want to make clear that this is but one approach, and I by no means intend this to be a universal definition:

Classical sovereignty is about territory and monopoly on violence: Digital sovereignty shifts this towards data, compute, infrastructure, and models. AI sovereignty becomes even more specific: control over an ‘AI stack’ [which is a whole other can of worms I will not open here], deployment, oversight, capacity, and so forth.

Yet, what gets lost in policy translation is the social reality: sovereignty becomes a matter of technical control, rather than social negotiation. As STS-folks have been saying for decades, once opacity, expert control and institutional mediation gets embedded into systems, the scope for ordinary agency shrinks (see: Langdon Winner’s classic essay).

In other words, sovereignty is always relational. It’s defined against something. It’s also always selective: it empowers some and marginalises others. Opacity and reduced agency on the one hand, and necessary exclusion on the other. So where do these two developments intersect?

I am reminded by something a colleague here in Singapore said during a roundtable discussion a while back: that we won’t really know what AI sovereignty is until it is enacted in practice by enough entities. In effect, sovereignty only really becomes legible once it’s enforced.

I find this framing compelling—and it intersects with some previous anthropological work on similarly amorphous terms, like universal rights. These are not singular things, and perhaps a singular definition is not always the most important thing. I want to approach sovereignty in this context more like a bundle of practices—technical, legal, bureaucratic, institutional, symbolic—through which authority over AI is asserted.

In other words, AI sovereignty is less about what it is, and more about what it does.

I realise all of this might seem a bit diffused, and you wouldn’t be wrong. So let’s ground this for a second with some examples.

Let’s talk about Japan.

Japanese national AI strategy has focused on AI alignment, specifically emphasising creating AI systems that are aligned with Japanese values, frequently underpinned by well-established discourse tied to human-centric AI. The value language often circles around ideas of harmony, social order, trust, politeness, social cohesion—which, speaking like an ethnographer here—comes across a bit more as a civilisational vibe than technical specifications, or carefully drawn out cultural definitions.

The reason for choosing Japan as an example here is because its AI alignment discourse is explicit, state-linked, and non-ethnic. It’s useful because it’s a comparably coherent example, and subtle enough not to be reliant on crude ethnic signalling—a point I think is key to highlight.

At first glance, too, this framing looks defensible. Japan’s strategy asserts the right to self-representation. It positions itself against external domination (often Western; often Silicon Valley). It anchors the discussion to established governance discourse around human-centricity. It also avoids explicit ethnic nationalism. There is no open hostility. No ranting about bloodlines.

However, once you decide your models should be aligned with ‘Japanese values’, you’re already doing some serious work. Firstly, you’re beginning to define who is in and who is out—this requires that you define the object. Second, you’re also turning this idea of ‘Japanese values’ into a stable object: something that can (and should) be encoded.

Unsurprisingly, certain dominant norms become the default:

• Language expectations

• Etiquette norms

• Conformity patterns

• ‘Appropriate’ social performance

• and so on…

Again, it depends on how you define all the above, but crucially to my point here, minority groups do not have to be explicitly targeted. They can simply… not appear at all. The few Auni communities that still exist on Hokkaido, the Okinawans, migrants—these complexities are ironed out, and disappear without outright hostility.

This is where Jennifer Robertson’s very good and very relevant work on robotics in Japan becomes worthwhile. There is a tendency to assume that new technologies are inherently progressive (even if they are imperfect and mistakes can happen). Yet, as she notes: Japan often does renovation rather than innovation. Futuristic tech becomes a vehicle for renewing older social norms. The future is imagined through familiar moral orders. AI isn’t inventing new politics; it’s reactivating existing ones.

So, let’s dwell on that point for a moment—and while we do, let’s return to Southeast Asia for another example.

Malaysia announced their ILMU model at the ASEAN Summit in Kuala Lumpur last August. I was there, in the audience experiencing all the pomp and circumstance. The Prime Minister was there! ILMU is marketed as being “for Malaysians and by Malaysians”, and though it has been announced that it will support multiple languages—English, Chinese, Rojak language (a mix of Malay, Chinese, English, and other languages), alongside Malay. This does sound great: a multilingual model for a multilingual country.

Yet, branding matters. In the public material I’ve seen, it has leaned heavily into Malay-ness: framed primarily as a model that speaks Malay; its image generation is predominantly generating images of Malays, and Malay food, and so forth. This is all done in a way that implicitly equates Malaysian-ness with being Malay. From an external view, this looks suspicious. Malays only make up around 60% of the population, with the other major groups being Malaysian Chinese, Malaysian Indians, and Orang Asli (especially in Sabah and Sarawak).

The point is that Malaysia is structurally multiethnic and multilingual. And the way ILMU is presented narrows this down to predominantly a majoritarian (and not by much…) identity, with the rest being presented as an afterthought.

Yet, I’ve had to frame this around how ILMU is being presented. The challenge is: ILMU is not publicly available, so there’s not yet been any opportunity to test it in practice. It is very possible that once released, it will be much more balanced than presented here. Yet, the risk of nationalism hiding as sovereignty very much remains—which is what this piece fundamentally is about—especially in light of Malaysia’s historical backdrop, and some contemporary political decisions being made around language.

There is also an irony worth noting: ILMU’s development is reportedly outsourced to a major Malaysian company—YTL Labs. As such, there is some lack of clarity around ownership, YTL being a private company and all, and how this might undercut the purported sovereignty of the project. However, this signals to me that the matter is less about economics and really more about politically controlling the narrative around what constitutes ‘real’ Malaysian culture.

This is not to mention the otherwise complex relationship between languages in Malaysia. It is not as straightforward as: the Malays speak Malay; Chinese speak Chinese; Indians speak Indian; and everybody speaks English. Overlooking, for the moment, that Malaysian Chinese speak more than one language (e.g. Mandarin and Hokkien), and that most Indians speak Tamil (among other languages), there’s also a matter of the performance of languages in different settings. Take an example—again from a colleague of mine—of a Malaysian Chinese civil servant: he drafts prompts in English because English is where current LLMs perform best and where he personally feels more productive. Yet, when he needs official text—for bureaucratic purposes where Malay is normatively required—he asks the model to translate or re-render it into formal Malay.

Think about what this implies: LLMs begin to function as mediators of public linguistic performance. English is the language of drafting; thinking. Malay becomes the language of official legitimacy—ritualised, symbolically indispensable, but increasingly relegated to something more performative (a version of which already exists in so-called Bahasa Istana; or ‘Palace Language’).

So sovereignty as discourse (Malay-ness, national identity, cultural alignment) and sovereignty as practice (English prompts, Malay outputs, performance gaps across languages) diverge, and in that gap, older hierarchies can be reproduced—now refracted through digital systems. Multilingual citizenship doesn’t have to be denied. It can simply be smoothly translated away.

So, let’s look beyond some examples here, and try to formalise this (if informally) into whatever set of mechanisms come into play for this sovereignty-to-nationalism drift. I have a working set—incomplete, certainly, but a way to begin the conversation:

1. Narratives

   Sovereignty framed as protection invariably comes with some story. It’s a narrative about who is being protected, from whom are they being protected, and why this is worthwhile. Whether presented in heroic terms, or as a matter of course, narratives are unavoidable players in this. The key is to make ‘the nation’ a stable category—or, indeed, a category that ought to be stable—and thus protection becomes a boundary-making exercise.

   Of course, this is not to say that narratives are inherently bad; rather, this is one mechanism through which this kind of sociopolitical slide can be engaged.

2. Design

   The design of systems—implicit or explicit—presents another mechanism to look out for in these contexts. This includes:

   • Training data choices: which communities are represented? Which dialects count?

   • Language prioritisation: whose language is the default interface?

   • Normative defaults: What social assumptions are baked into the system?

   Crucially (and a way in which this intersects with narrative mechanisms) is how many of these choices are justified. Yet, let’s not forget that many will be justified as purely technical decisions; this is ‘the best way’ to do it, according to engineers and scientists. Yet, technical decisions are governance decisions when they begin to structure social or cultural legibility.

3. Governance

   This forms the classic challenge of ‘quis custodiet ipsos custodes?’; which is pretentious for ‘who watches the watchers?’. This includes questions around:

   • Who defines ‘local values’?

   • Who gets consulted in the design process/deployment?

   • Who has veto power?

   • How are disagreements handled?

   A lot of sovereign projects skip participatory processes because they’re slow and often perceived as politically messy. So ‘values’ are defined by whoever is already authorised to define ‘stuff’—which, in some cases, might be the result of laziness, and in other cases, it is absolutely the whole point.

4. Temporal

   In these kinds of models, ‘culture’ and ‘society’ is treated as temporally static. However, as we all know, they are now: they’re always contested, and they’re constantly evolving. In these cases, alignment becomes an exercise in encoding a snapshot (and often a very carefully curated snapshot). This snapshot then becomes infrastructure, which—at the best of times—simply becomes more and more outdated, and in worst cases, a means of enforcing a very specific, often conservative vision of society and how it ought to be.

Though a lot of these can be harnessed by those with less than noble intent, I think an important takeaway for me here is the opposite; all of these mechanisms can come into play without such explicitly exclusionary intent, be it out of laziness, lack of resources, skills mismatch, something else, or a combination of all and more.

You don’t need mustache-twirling nationalism for this slide to occur. You need a handful of defaults, some shortcuts, and a good story—and maybe an arsehole or two mixed in for good measure.

This, of course, begs the question of what can be done about all of this? As I’ve mentioned before, this is not a post trying to ‘take down’ AI sovereignty or otherwise undermine the idea. Rather, I want to make the point that AI sovereignty as a solution to a very real problem also comes with its own governance challenges that we ought to try to get ahead of.

I think the best way of side-stepping this is not to treat sovereignty as an object(ive), but as a process. It ought to be understood as an ongoing negotiation rather than fixed control; procedural rather than categorical. This is not about ‘this model for the nation/people/etc.’, but it ought to be ‘this governance arrangement maintains agency, accountability, and contestability’. The problem, of course, is that from a narrative perspective, it sounds far less sexy. Which is a recurring problem with governance: it sounds boring, because reasonable solutions often are, at least on paper.

Furthermore, if we’re serious about cultural alignment (believe me, I am), pluralism has to be treated as a design challenge, not some vague virtue. By extension, thus means:

• Multiple cultural registers: remember! Culture is fractal.

• Contestability built into systems: ways to challenge outputs, update norms, surface disagreements.

• Resisting universalisation at all costs: no ‘one model for all’ fantasies.

• Focusing on specific contexts and deployment: think through use-cases rather than sweeping national abstraction.

Sovereignty need not mean sameness and autonomy does not require homogeneity. However, it does require resisting the temptation to treat the nation as a single cultural object that can be encoded and deployed.

Furthermore—returning to the two examples above—essentialism, stabilisation and erasure through (il)legibility are not country or region-specific. They can be observed across liberal democracies, in postcolonial contexts, and other contexts alike. The takeaway here, really, is that AI systems do not introduce these dynamics, but potentially amplify and scale existing sociopolitical structures by embedding them within technical and governance decisions.

Once ‘Japanese values’ or ‘Malaysian-ness’ or any other such construct becomes the alignment target, the system stabilises that target; it makes it reproducible and scalable. It also often makes it feel, a bit perversely, like common sense—because the machine and benchmarks say so!

AI doesn’t merely reflect culture. It crystallises a particular version of culture into infrastructure.

So here we are.

AI sovereignty is contested, yes. It’s also increasingly treated as a default good, and in many contexts, it is a reasonable response to very real asymmetries of power, representation, and infrastructure. However, it is also not a neutral value, and once sovereignty gets optimised through AI systems—such as through alignment, through value chains, through cultural branding—the risk is that you don’t just build autonomy. You can also build cultural fixation by stabilising a singular national imaginary. You then scale it, naturalise it; make it appear inevitable.

If you’re a policymaker, the uncomfortable part is that this can happen without any particular intention. If you’re a developer, the uncomfortable part is that the slide can happen through the very decisions often called ‘technical constraints’. If you’re anyone who lives in a multilingual, multiethnic, and internally contested society (like… you know, basically all of us), the uncomfortable part is that AI can make those contestations less visible—not by repressing them, but by rendering them administratively irrelevant.

AI should be treated as a site of negotiations, not resolution. Sovereignty is relational, unfinished, and politically and socially alive.

As it should be.