The Global Dialogue on AI Governance is the first process of its kind where every country holds a seat. That gives it the standing to speak for the people who have had the least influence over how AI is built and deployed. In an earlier piece, we looked at whether the consultations are living up to that promise in practice. This piece shifts the focus from who is in the room to what the agenda asks them to discuss.
The Dialogue’s four thematic clusters cover a lot of ground, and on paper the issues raised below can be filed under one or another of them. The trouble is how they are read. Questions about access and implications tend to be treated as gaps to be narrowed or effects to be noted, rather than as questions about who holds power in the AI economy, who owns the means of building AI, who supplies the labour and the data, and who absorbs the costs. These are not secondary to the clusters, but are ground the clusters stand on.
Practically, this means much of the governance conversation currently concentrates on visible harms, such as biased model outputs or prematurely deployed systems. That work is necessary, and we do not mean to wave it away. But these harms are produced by a structure, one in which the resources needed to build AI sit with a few actors, while the people and places that supply the industry’s raw materials have the least say over it. If we keep treating the harms one at a time, we are left repairing outputs while the conditions that generate them keep producing new ones.
Who Owns the Infrastructure
The compute, the foundation models, and the benchmark datasets used to judge whether a model is any good sit with a small number of companies and jurisdictions, almost all of them in the Global North. Countries without a stake in that infrastructure end up importing systems built elsewhere. This concentration shapes everything downstream. When a handful of actors control the infrastructure, the models, and even the tests used to evaluate the models, the terms on which everyone else takes part are set before the conversation begins. For the Global Majority, this is the starting condition the Dialogue has to reckon with.
The consequence of concentration is dependency. Global Majority countries largely import foundation models trained on data drawn from elsewhere and optimised for contexts that are not their own, then deploy them in high-stakes settings — welfare, healthcare, identity, law enforcement — without the local capacity to evaluate them first. Bias enters not only through the models themselves but through the long history of communities being under-represented or mis-represented in the data they are built on. The result is systems that fail most where the power to challenge them is the least.
It is worth being clear about what does and does not address this. The ability to move data and systems between providers without being locked in is genuinely useful, and the Dialogue should support work on open standards. But interoperability on its own is not the same as having a say. Being able to switch between systems that were all designed elsewhere, on terms set elsewhere, is a thin form of agency. The harder question is whether Global Majority countries and communities can shape what gets built and how it is evaluated in the first place.
A few measures would begin to rebalance this. The most immediate is practical access to compute, so that researchers and institutions across the Global Majority are not left dependent on the pricing and goodwill of a few providers. Alongside it, open evaluation benchmarks built with and for under-resourced languages and contexts, instead of benchmarks that treat English-language, Global North performance as the universal standard. Then, further upstream, regional capacity to evaluate AI systems before they are deployed rather than after the harm is done — capacity hosted by institutions in the regions themselves — and public financing for a distributed network of research centres across the Global Majority, so that the ability to build and study AI is not permanently concentrated elsewhere. These are concrete, fundable, and within the Dialogue’s power to champion.
Whose Rules Count
There is a governance layer to this concentration as well. When a few jurisdictions own the infrastructure, their regulatory frameworks tend to become the default reference point too. The Dialogue should not treat Global North instruments such as the EU AI Act or US executive orders as the baseline from which everyone else’s approaches are measured as deviations. A law from Brazil, a regulation from Indonesia, or a policy from Kenya, each deserves to be treated as a contribution to global AI governance in its own right, not as a regional variation on a standard set somewhere else.
A useful role the Dialogue can play here is to help build shared reference points across these different traditions — minimum common understandings and baseline safeguards that let governance models work together without being forced into uniformity. Countries do not need identical rules. They need enough common ground to cooperate, and a forum with universal membership is well placed to help establish it.
What Gets Extracted
Extraction in this economy takes several forms, and each falls hardest on the Global Majority. AI systems do not run themselves. Behind them is a large and mostly invisible workforce labelling data, moderating content, and doing the piecework that makes models usable. Much of this work happens in the Global Majority under conditions that governance frameworks have barely touched, such as algorithmic wage-setting, performance surveillance, abrupt termination, and no real route to appeal or to bargain collectively, least of all across borders. Women make up a large share of this workforce and carry extra harms on top, from wage discrimination to harassment to exclusion from what little governance the platforms have.
The fix here is less novel than it sounds. In large part it means extending existing labour rights and protections to data and platform workers, rather than inventing separate frameworks for them, and adding coverage for what is genuinely new about this work: algorithmic management, and employment that crosses borders.
Two further kinds of extraction belong on the agenda, because both land so unevenly on the Global Majority. The first is the extraction of Indigenous and traditional knowledge. This knowledge is increasingly scraped into training datasets without consent, recognition, or compensation, and large language models tend to flatten and misrepresent knowledge systems they were never designed to hold. These same models perform poorly in under-resourced languages, pushing already marginalised languages further to the margins and failing the people who speak them. Women who hold traditional knowledge are often hit hardest, watching their expertise absorbed with neither acknowledgement nor return. What helps here is consent and stewardship — disclosure of where training data comes from, real consent mechanisms for indigenous and traditional knowledge, and support for community-led data stewardship, where communities keep a say over how knowledge about them is collected and used.
The second is environmental extraction. AI runs on data centres that consume large volumes of water and electricity, on hardware cycles that generate growing volumes of e-waste, and on raw materials whose extraction is concentrated in the Global Majority. These burdens are gendered too. When water access suffers, the extra care work tends to land on women, who are at the same time mostly shut out of the more formal parts of the recycling and materials economy. Assessing environmental impact across the full lifecycle of a system, and holding back from deployment where the impacts are uncertain, would give the Dialogue something concrete to work with rather than leaving all of this to a general line about sustainability.
Taken together, these costs raise the question of who bears responsibility for them. The Global Digital Justice Forum makes the case in its submission for a principle of common but differentiated responsibilities, borrowed from environmental governance, where those who have gained most from an extractive model of development carry a larger share of the duty to address the harms.
Who Is Accountable
Underpinning all of this is one governance question: when AI systems cause harm along this chain, who is accountable, and how? Voluntary safety commitments have not produced a convincing answer. What is needed is accountability that runs the full length of the multi-actor value chains through which AI is built and deployed. This calls for binding public accountability and corporate liability.
Accountability of that kind depends on being able to see how systems actually perform, which is where disclosure comes in. A safety claim from a developer should arrive with the conditions it was tested under, the limitations the developer already knows about, and failure rates broken down across the dimensions where harm tends to concentrate: language, gender, disability, region, skin tone, whatever is relevant to the system in question. Disclosure, however, is not the destination; it is the ground accountability stands on. You cannot hold a system to account if you cannot see how it was tested or where it fails.
What this means for the Dialogue
It is worth being clear about what a meaningful idea of digital sovereignty would mean here. It does not mean every country building its own frontier model, and it does not mean merely being free to switch providers. It means not being permanently cast as data supplier and harm absorber in a chain you had no hand in designing. The call for fair and frugal AI — locally rooted, mindful of ecological limits, shaped through participation — points in a complementary direction.
The thread running through all of this is data justice. Data is not a free-floating resource waiting to be extracted, but something produced by people and communities with a legitimate stake in how it is governed. Read that way, the scattered concerns above coalesce into a coherent agenda: one built on regional capacity to evaluate and build AI, public financing so capacity does not stay concentrated, labour protections that treat data workers as workers, and accountability that runs the full length of the chain.
None of this requires the Dialogue to abandon its existing clusters, only to read them with power and extraction in view. The Dialogue’s universal membership gives it both the standing and the mandate to read them in exactly this way.
TGI will be in Geneva for the Dialogue, making the case that governing AI well means governing its foundations and not only its symptoms. We will report back on what we hear.
