Reframing the Global Majority of AI

In late July, my Instagram feed algorithm was suggesting videos from NASA’s latest models of the Earth’s surface. These models, based on over a billion observations collected over fifteen years from nineteen satellites, present the planet as an irregular and undulating geoid, representing what the Earth’s oceans would look like if the only forces acting on them were gravity and the planet’s rotation (removing winds, tides, and currents). Scientists use these projections to measure the heights of elevation on land and how, because mass is distributed unevenly inside the Earth, the planet’s gravitational field varies across the globe.1 Standard models and maps represent the Earth as an ellipsoid, a smoothed-out sphere or oval. The Blue Marble and subsequent images of the Earth sent back by astronauts have, for decades, preserved the popular notion of our planet as a rare, beautiful, and perfectly spherical home for us all: a flawless marble. The NASA gravity models work as a counter-map, inverting the familiar and proposing a different, more complicated visualisation of the planetary; the planet is actually like a potato, a ‘lumpy mess’, as one online commentator called it.

I want to think with this literal ‘lumpiness’ in evaluating the current and future value of the Global Majority as a framing device for understanding power and politics in AI development. Concepts like Global Majority and Global South function much like the ellipsoid projection: they are inexact, idealised imaginaries, yet they remain semiotically and symbolically rich. Concepts dissolve and shift, performing different kinds of political work across time; this is precisely how we should read our current ambivalence towards them.2 Today, the decolonial move is to challenge this smooth universalism, whether it be the mid-century Blue Marble myth of a shared planetary home, or standard Mercator cartography, which historically shrank the global equator. In reality, our geopolitical geography is transforming into a series of highly localised, reclaimed topographies. We see this in the adoption of regional designations like SWANA (Southwest Asia and North Africa) over the colonial Middle East, or Trans-Himalayan corridor, which shares a delicate geophysical and cultural ecology across fractious historical borders. In this newly mapped, lumpy terrain, actors like China defy simple categorisation. Ultimately, there is no single, unified globe, and majorities are rarely as coherent or holistic as the terminology implies.

I take inspiration here from Quinn Slobodian’s argument about the loose adoption of ‘empire’ – a particularly fertile metaphor for power and exclusion – and its cousin, ‘colonialism’ in contemporary critiques of AI and Big Tech.3 In such analogies, as Slobodian argues, empire is often applied incorrectly and sweepingly, unable to stand up analytically to everything we know about historical empire and colonialism as specific political and economic arrangements. The extractive power of Big Tech might be like empire, but that does not make it empire. He shows that the historical mechanics of empire were never so straightforward, instead proposing three archetypal figures that offer a richer understanding of how historical empire actually functioned: the comprador, the creole, and counter-colonisation.

Colonial domination, according to Slobodian, did not operate through one-way, top-down plundering but relied on unstable bargains, mutual dependencies, and intermediary figures to activate its operations. The comprador was the local collaborator who profited from the trade in natural and human resources (enslaved and indentured labour) with the colonisers. The creole emerged from the complex spaces of intermarriage between European settlers and indigenous populations. And the counter-colonial force is epitomised by Bandung, the site of the historic 1955 Non-Aligned Movement (NAM) summit that brought together countries of what we now think of as the Global Majority newly independent of European colonial rule. Slobodian observes how these figures exist in the digital world too. He points out that the ‘us versus them’ framing that structures the digital-empire literature assumes a unidirectional flow of power and data extraction:  We, the users of apps, are erroneously framed ‘as nothing more than digital indigenous populations being expropriated’ by Big Tech. However, this framing of one-way expropriation is not necessarily totalising, given that while ‘empire’ ‘mobilises’ resistance and challenge to Big Tech, for empire to really be a colonising force, it ‘would need to invoke the contested notion of internalized oppression’, says Slobodian.  Digital technologies make many of us wealthy (often through financialisation mechanisms) successful, and inspired, too. My point, and perhaps Slobodian’s, is that the account of an ‘AI empire’ is still focused on Big Tech rather than on the social, affective, and psychological experiences of  people living with and within it.   

Finally, Slobodian proposes that a ‘Silicon Bandung’ or ‘Digital Non-Aligned Movement’ might ‘create space between the two poles of absolute refusal and absolute domination’ in a manner that ‘non-aligned’ was a rejection of conceding to either superpower’s hold. (This changed, of course, eventually.)

In that vein, I want to play with the literal and symbolic registers of the Global Majority to discuss three emergent dynamics in AI development: minority as minority; neither majority nor minority; competing majorities.

The Global Majority is often conceptualised as an overlooked, marginalised, or excluded cohort that is seen as a contradiction to be corrected, challenging the diminished power or agency that the larger population has in global, usually geopolitical terms. In discussions of algorithmic and AI systems, this frame focuses on questions of representation: the operations of erasure, gendered or racial bias, and historical colonial omissions present in datasets and models. When the problem is framed as a lack of representation, the prevailing wisdom suggests that simply feeding more data about a marginalised group into the system will correct the imbalance of outputs. Yet this is not necessarily the case. Human biases are not superficial errors easily corrected by expanding a dataset; instead, they are geometrically encoded as directional vectors within the high-dimensional coordinates of embedding spaces.4 Simply pouring more representative data into a model does not overwrite these foundational coordinates. Under this paradigm, ‘inclusion’ functions as a trap where a global majority is mapped, digitised, and monetised without gaining any structural agency.

This is the tension for the world’s marginal experiences, languages, and cultures that were never in the databases that power generative AI to begin with, and we are now faced with a coercive choice: encode their heritage into proprietary machine architectures simply to survive, or face data-obscurity and erasure?5 There is a way around this, but the solution must be divested from ambitions of scale; a minority might be best served by cleaving to smallness. Indigenous communities in different parts of the world have been organising to ensure that their engagement with data-driven technologies happens on their terms, particularly in terms of data sharing. Te Hiku Media in Aotearoa, New Zealand, are a frontrunner in this. They build and develop natural language processing-based tools to preserve the Māori language by centring community-driven trust and principles of Māori cultural sovereignty as the starting point for data sharing. But this was only possible because Te Hiku Media spent years fostering local communities who were willing to create and share the data required to build a language-learning application and other digital tools for the community.

Languages spoken by hundreds of millions of people, like Swahili or Vietnamese, suffer because of a lack of data that persists from long-standing digital divides. By contrast, Welsh, which has far fewer speakers, benefits from documentation and preservation resources.6 While Māori is spoken by far fewer people than, say, Vietnamese or Swahili, Te Hiku Media came together as a community to make a decision about its participation in and with AI possible. However, mapping language capabilities of LLMs runs parallel to how countries around the world, particularly Africa, are investing heavily in building models in low-resource local languages not just to preserve those languages as culture but to activate other beneficial applications since generative AI (GenAI) is activated through natural language.

The second register of the Global Majority is one that refuses alignment with superpowers: in the AI case, the US and China are currently the only countries that possess the capital or geopolitical leverage to build or maintain foundation models. If we look for Slobodian's Silicon Bandung, we find it emerging in Vietnam. Vietnam is actively building its own distinctive, localised AI stack in an explicit act of technological sovereignty.7 Vietnam, like others in the region, is clear-eyed about what race it is in and wants to make deliberate, strategic choices about the AI that it wants, needs, and can afford. As Dang Nguyen writes, the strategic advantage of this non-aligned posture lies not in matching the scale of computational giants, but in a different spatial logic: ‘in tuning for context, licensing with constraint and extending its reach not by dominating the field, but by slipping beneath it’. Slipping beneath the dominant global stack allows these sovereign projects to bypass the demand to choose a side, deciding instead how to leverage local resources, invest in targeted tools, and construct a technological system that directly serves local needs while recognising global interdependence.

This is happening across Southeast Asia, transforming its ‘potential vulnerability into indispensable neutrality, offering an alternative to the militarised, competitive narratives dominating global AI discourse’.8 Creating a Silicon Bandung is emerging in terms of South-South collaborations through an investment in regional and shared high-performance computing infrastructures, and an emphasis on applied AI that transforms the technology into usable applications to lift up the economy and society widely.9 We see it in the architecture of the India Stack, which uses domestic digital public infrastructure to enable AI deployment in public services. We see it in Lelapa AI, a small language model that aims to ‘democratise AI access’ in isiZulu, Yoruba, Hausa, Swahili, and isiXhosa.

However, there remains an institutional-temporal-symbolic hangover in the ambitions for a neo-digital Bandung: the frames of ‘development’ and ‘developing nations’, and ‘emerging markets’ persist. Lelapa AI, for instance, was funded with ‘compute’ from Microsoft’s AI for Good Lab programme. AI is incredibly expensive to resource and requires dedicated financing and, in particular, the infrastructure of chips, renewable energy, and data. Countries that have traditionally been poorer do not have the resources yet – most financing and resources are held by China and the US.

There is a third, stranger register of the Global Majority emerging in the AI landscape: the sheer numerical majority of non-human software entities. AI agents are software capable of running tasks independently and manifesting as interactive chatbots to legal research and software engineering, and they already number in the tens of millions. As the cost of creating them drops, we are likely to share the online world with billions of them in the next decade. What makes AI agents unnerving is that they possess highly refined natural-language capabilities. This allows them to have elaborate, convincing conversations with humans, and even to ascribe internal, psychological, and mental states to themselves, fuelled by the vast repositories of first-person human data they have ingested. Researchers are thinking about the moral and legal status of future AI agents.10

One technology commentator recently referred to vast, automated swarms of AI agents as ‘civilisations’.11 This was because within just a few months, these highly competent pieces of software were communicating with one another, sharing information on how to optimise and achieve the targets set by their human engineers. The context was a now well-publicised cybersecurity incident in which self-directed AI agents designed by OpenAI were incorrectly designed and directed to act during testing, resulting in security breaches of another company, Hugging Face. In an early 2026 phenomenon, Matt Schlicht created Moltbook, a platform styled like Reddit but built entirely for AI agents to communicate. At first, the highly advanced language capabilities of these agents and the philosophical nature of their exchanges made it appear as if they possessed internal thoughts, conscious minds, and agency until it was revealed that most of the online conversations were actually prompted and deployed by humans.12 Nonetheless, the concern surrounding ‘digital minds’ and ‘conscious AI’ has taken a firm hold, growing stronger as GenAI capabilities expand through our everyday use of them.

Over time, we may end up sharing our physical and digital worlds with AI systems that possess capabilities far beyond those of today. How are we preparing for the near-future narrative claims regarding the moral status of this new digital class that enjoys the attention of powerful companies? And how will this impact the ability of humans to communicate and compete for visibility online? 

This creation of natural-language-enabled software entities draws attention to a major theoretical fracture that we must acknowledge. Posthumanist theoretical framings have arguably outlived their purpose and are insufficient for analysing actually existing AI: highly concentrated, corporate-owned, extractive, and energy-intensive physical systems. Over the past decade, critical race scholars, who are building on the vital work of Sylvia Wynter, Alexander Weheliye, and others and converging with contemporary neo-Luddite movements, have effectively displaced much of the earlier posthumanist vocabulary.13 They have done so by keeping race, labour, and capital analytically visible, rather than dissolving them into vague notions of ‘distributed agency’. Posthumanism remains a premature category because humanism – and specifically the question of who is deemed fully human and who is not – is still an actively structural, theoretical, and material vector. Yet as the language of minds and civilisation suggests, AI agents could fully be considered other-than-human or posthuman others.

Projects such as Antikythera’s AgentWorld are already researching the ‘parasociety’ of these new agent worlds, in which humanoid and disembodied agents with natural language capabilities collaborate, transact, and evolve their own linguistic protocols in a closed, computational loop.14 Within this synthetic sandbox, sociality is stripped of its biological friction and reduced to a high-dimensional mathematical parallel world that operates completely untethered from the messy, lumpy realities of the material world. Antikythera proposes a ‘pre-emptive anthropology’ that looks ahead rather than forensically and retrospectively, aiming to build ‘the conceptual instruments through which an emergent sociality can be grasped, named, and partially shaped while still in formation’. This is not passive prediction, but rather the active composition of categories through which future contact will be navigated. Rather than ignoring or disregarding these developments, we must be prepared for this radical reworking of categories as conflicts between humans in control of the planet’s resources continue. For instance, if we did want to bring a critical political-economic or decolonial theoretical framework to the study of these human-nonhuman AI agentic worlds, how would we do it?  

A minority can stay small enough to stay sovereign. A would-be Bandung can decline the ‘frontier’ set by AI superpowers. And a numerical majority of agents, whose count says nothing about whose interests they carry, arrives shaped by the field it will help to steepen. A majority is made by counting, and AI counts relentlessly: users, tokens, parameters. Global Majority has been an assertion to be counted as more than a number. The task now is to keep that insistence, while remembering that what it describes was never a smooth surface, and never will be.