Governing AI is not only about regulating models. It means deciding who controls data, which institutions retain power, and which forms of knowledge count.

Three developments this August, in Southern Africa, the United States and Pakistan, allow us to watch that struggle while it is still open.

They do not concern the same technology, nor do they respond to the same problem. Yet all three raise a common question:

when artificial intelligence reorganises institutions, knowledge and social relations, who retains the power to decide?

Before artificial intelligence come the data

Between 12 and 14 August 2026, policymakers, regulators and technical specialists from countries in the Southern African Development Community met in Harare for the Southern Africa Regional Workshop on Data Governance in the Digital Age: Data Policy Harmonisation.

The meeting was hosted by the Government of Zimbabwe through its Ministry of Information and Communications Technology and POTRAZ, with UNESCO, Smart Africa and the World Bank among the organisations involved.

Its purpose went beyond making national legislation compatible. It sought to advance regional frameworks capable of supporting cross-border data flows, digital public goods and readiness for artificial intelligence.

The programme was explicitly anchored in UNESCO’s Data Governance Toolkit and its 4Ps framework: Purpose, Principles, People and Practices. It was also designed to feed into a co-developed regional roadmap for data-policy harmonisation within the SADC framework.

This can sound like an administrative discussion. It is not.

An AI system needs data before it can learn from a society. Those data do not emerge spontaneously: someone decides what to collect, how to classify it, who may use it and under what conditions it may cross a border.

Languages, territories, occupations, identities and behaviours must first be transformed into categories that an institution — and later a machine — can process.

Governing data also means governing the ways in which a society becomes legible.

Data sovereignty therefore does not necessarily mean closing digital borders. It means retaining institutional capacity to determine the conditions of circulation: who gets access, for what purpose, what must remain protected, which rights are retained by the people and communities from whom data are derived, and how much of the value generated remains in the territory.

The warning voiced in Harare was explicit. Without strong governance and coordinated policies, the region risks falling behind in AI readiness or becoming a “passive consumer of external technologies”.

The discussion also addressed highly concrete problems: privacy, algorithmic bias, cybersecurity and unauthorised cross-border data extraction.

Artificial intelligence thus appears at the end of a chain of political decisions, not at its beginning.

Making machines legible too

On 20 August 2026, OpenAI launched AI Futures, the public blog of its new Strategic Futures team, led by Dean Ball.

Its founding question is unusual for an artificial-intelligence company: how should a free society be restructured in order to preserve individual rights and agency while accommodating the emergence of transformative AI?

OpenAI itself describes these concerns as concentration-of-power risks and states that the inaugural article represents Ball’s views rather than necessarily an organisational position.

Ball begins with a historical observation: states and firms have depended on mass human cooperation to produce, administer, levy taxes and exercise power. That dependence distributed power profoundly unequally, but it also imposed limits.

Automation could change that balance.

Ball argues that autonomous systems could allow states to project force without relying to the same extent on soldiers or police officers, that parts of the bureaucracy could be automated, and that revenue generated by data-centre production could reduce reliance on human labour.

In such a scenario, people could lose material bargaining power in relation to those who control infrastructure, capital and computing capacity.

Elections and formal rights might even remain in place while people’s effective ability to influence the institutions governing their lives declines.

Another circumstance makes the text particularly interesting as an anthropological object. Ball joined OpenAI after working on AI policy at the White House and taking part in the development of America’s AI Action Plan.

His trajectory shows how permeable the boundaries have become between public administration, the technology industry and the production of frameworks for governing AI.

And the paradox is difficult to ignore: one of the organisations building infrastructures capable of concentrating new forms of power is simultaneously studying the political risks of that concentration.

This does not invalidate the analysis.

It also makes the text a source for studying how the industry itself imagines the future: what it means by freedom, which institutions it regards as central, which risks it recognises, and which solutions it can formulate from its own position.

One of those proposals is particularly suggestive: bounded legibility.

Ball proposes that when an AI system undertakes high-stakes actions affecting the physical wellbeing or property of third parties, those actions should be traceable to a responsible person or human-controlled organisation.

In his words, “those actions must be able to be tied back to a responsible human”.

But attribution should not become generalised surveillance. The proposal maintains that privacy and anonymity should remain possible in many other uses.

The idea does not solve model opacity or every problem of accountability. It establishes something narrower: a minimum threshold of traceability when automated action can cause harm beyond the screen.

The reversal is significant.

Modern states developed censuses, maps and registers to make their populations legible.

We are now beginning to need the chain of machine action to become legible as well.

Who gave the instruction? Which decisions followed? Which tools intervened? Who answers when an automated sequence produces consequences?

AI then ceases to be only a problem of technological control. An old social problem returns with new actors: how to attribute action, responsibility and power.

Problems first, technology afterwards

In Pakistan, artificial intelligence is arriving within a network of problems that were already connected.

According to UNESCO’s report, on 19 August 2026 the seven Chairs gathered at Beaconhouse National University in Lahore sought to strengthen national cooperation across research on climate, AI, heritage, water, health, education and culture.

The combination is revealing.

MNS University of Agriculture presented work on low-carbon agriculture and smallholder farming. COMSATS presented research on water and disaster-risk reduction. Information Technology University highlighted cyber-resilience, public-sector digital governance and literacy through its AI for All course.

NED University connected urban resilience with documentation of living heritage in Tharparkar. The National College of Arts presented work on historic conservation and craft knowledge.

The International Center for Chemical and Biological Sciences at the University of Karachi brought biomedical research, support for early-career researchers and training in molecular medicine into the conversation.

And BNU demonstrated how visual arts and design can transform complex climate and technological data into public narratives that are easier to understand.

Outside the university, none of these problems is truly separate.

Water is connected to agriculture. Agriculture to climate. Climate to territory and ways of life. Health to environmental conditions. Heritage to the knowledge through which a community learns how to inhabit those territories.

Technology comes afterwards.

That sequence reverses a familiar logic of the technology industry: first build a capability, then search for problems to which it can be applied.

Lahore suggests another possibility:

first come the problems, the forms of knowledge and the institutions. Only then do we decide which technology may be useful, and under what conditions.

The difference matters.

A system capable of combining information about water, soil and climate might, for example, help inform agricultural decisions. But the question would not only be whether the model can do it. It would also be who defines the problem, using whose data, for whose benefit, and with which local forms of knowledge participating.

This sequence does not guarantee justice or eliminate power relations.

But it preserves something fundamental: the capacity to set the agenda.

It also reminds us that building digital futures should not be the exclusive province of computer science or engineering. Agricultural research, public health, heritage, the arts, local communities and water specialists all possess knowledge that matters when deciding what should be digitised and how.

Interdisciplinarity therefore ceases to be an academic formula.

It becomes a dispute over which forms of knowledge have the right to participate in technological decision-making.

Govern before automating

Harare, OpenAI and Lahore occupy very different positions.

Southern Africa is trying to retain regional capacity to determine how its data circulate before becoming dependent on systems developed elsewhere.

OpenAI acknowledges that deep automation could alter the historical relationship between institutions and populations, and proposes mechanisms for making some machine actions attributable.

In Pakistan, universities are trying to keep territory, health, heritage, climate, knowledge and technology connected before deciding what deserves to be automated.

All three scenes point towards the same problem.

Artificial intelligence never arrives alone.

It arrives with a promise of legibility.

It turns societies, languages, territories and practices into data so that machines can act upon them. Yet as those machines acquire greater capacity to act, we also need to make their decisions, chains of responsibility and deploying institutions legible.

The political question is not only how much a machine will be able to decide.

Nor is it simply who will control it.

There is an earlier question:

who will have the power to decide what must become legible, what may remain opaque, and who retains the capacity to say no?

And there is another question we can bring back to our own context:

which data, forms of knowledge and infrastructures would we treat as public goods before allowing a machine to decide what to do with them?

Sources and references