“The next frontier of digital inclusion is not access, but agency.”
For years, much of the discussion around artificial intelligence has been framed in terms of access. Who has connectivity, who can use a model, which institutions have computing capacity, and which populations receive training. That framework remains relevant, but it is beginning to fall short.
The United Nations Development Programme (UNDP) put it explicitly this week when discussing Africa’s digital future: “The next frontier of digital inclusion is not access, but agency.” The measure of inclusion therefore shifts from connection itself to the ability to turn it into learning, income, innovation and meaningful participation in the economy. [1]
The word agency may appear to describe an individual capacity. This week’s cases point to something more complex. The capacity to act depends on material infrastructures, institutions, accumulated knowledge, classifications and relationships stable enough to sustain action.
A person needs more than training. A public administration needs more than a tool. Two states competing over a technology need more than technical capability to keep that competition within manageable bounds.
AI is turning agency into an infrastructure problem.
From being connected to being able to act
UNDP’s argument begins with a paradox. Africa has the youngest population in the world and an enormous potential pool of technological talent, yet training people in AI, data or cybersecurity does not guarantee that this knowledge can be converted into economic capacity.
Jobs, practical experience, mentoring, professional networks, finance and access to the tools required for work are also needed. UNDP summarises the relationship simply: talent needs tools, and tools need talent. [1]
Agency therefore shifts from the individual towards the system of relationships that makes action possible.
A second UNDP report makes this materially visible. The organisation has deployed local AI compute capacity in Kenya, Malawi, Rwanda, South Africa, Togo and Zambia. Published on 17 September, the report examines what is required to move from investment in advanced infrastructure to sustained national capacity able to generate public value. [2]
The physical presence of machines does not, by itself, create technological sovereignty.
The assessment of the six deployments concludes that effective use requires accessible data, people able to operate the systems, clear institutional responsibilities, governance and cybersecurity arrangements, sustainable financing, and defined pathways from priority applications to operational services. Across the six countries, 21 applications were identified; Rwanda, meanwhile, has defined 42 priorities across eight sectors. [2]
Infrastructure becomes capacity when an institution exists that can maintain it, direct it and decide what it is for.
An important distinction therefore emerges between localisation and agency. A GPU can sit within a country’s borders without that country controlling the models it runs, possessing the right data, being able to maintain the infrastructure, or capturing a significant share of the economic value it produces.
UNDP had already framed this tension in the language of sovereignty: who benefits from African data, who builds digital public goods, and who controls the infrastructures used in health, agriculture, finance or education. [3]
An anthropological reading draws attention to the relationships that turn localised infrastructure into genuinely owned capacity.
Agency is not possessed in isolation. It is assembled.
When experience becomes a use case
Something similar happens inside public administrations.
Government AI Navigator, launched on 8 July by Apolitical with support from Google.org, began with 300 AI projects developed by public administrations in more than 50 countries. The platform makes it possible to see what was attempted, what worked, what failed and who was behind each experience. [4]
Its starting point is also a lack of institutional agency. According to the research accompanying the project, around 40% of public servants spend a week or more gathering comparative evidence before starting new work. Meanwhile, different administrations may be trying to solve similar problems without knowing that someone elsewhere has already experimented with them. [4]
Apolitical responds by accumulating experience in order to reduce that repetition.
Turning experience into portable knowledge, however, requires a less visible operation.
A project developed in Helsinki, Nairobi, Singapore or a small Spanish public administration exists within a particular legal framework, budget structure, set of labour relations, data systems, technical capabilities and bureaucratic cultures. Converting that experience into a “use case” requires deciding which elements are essential and which can be discarded.
The case stops being only an experience. It becomes a classification.
Budget, technology used, outcomes and metrics can be converted relatively easily into comparable fields. Internal resistance, labour negotiations, informal relationships, public expectations or the political contingencies that enabled — or prevented — a project from working are much harder to transport.
More than a century ago, Max Weber described one of the central operations of legal-rational bureaucracy: turning particular situations into administrable matters through defined competences, relatively impersonal rules, documentation and files. Public administration gains the capacity to act because it can make different cases comparable and handle them through relatively stable procedures. [14]
Government AI Navigator takes that logic to another scale. For an experience developed in one context to travel into another, it must first be transformed into comparable, retrievable information.
This is also where the problem studied by Geoffrey Bowker and Susan Leigh Star in scientific and medical classification infrastructures appears. Comparability allows knowledge to travel, but it requires decisions about what counts and what is left out. [15]
Government AI Navigator does more than transmit knowledge between bureaucracies. It also helps stabilise what counts as relevant knowledge about public-sector AI.
A decisive issue follows: what survives when a situated practice becomes a use case.
The same logic of infrastructural assembly reaches a different scale when the actors are no longer national bureaucracies, but technological powers in competition.
Washington and the construction of a relationship around AI
Between 23 and 25 September, Xi Jinping made a state visit to the United States. For this analysis, we deliberately use only Chinese institutional and media coverage of the visit. The coherence of that corpus forms part of the object being examined. It describes the encounter while also revealing the narrative apparatus through which the Chinese state chooses to make it legible.
Protocol and political staging exist on both sides. Here we are observing specifically how the official Chinese corpus organises and interprets them.
Artificial intelligence occupies an explicit place within that narrative.
In his talks with Donald Trump, Xi presented China and the United States as two countries with strong AI capabilities that can compete and cooperate. The Chinese formulation calls for continued dialogue on risks and benefits, the prevention of misuse, and keeping the development of the technology under human control, with a people-centred approach oriented towards human progress. [5]
For now, this language expresses the position attributed to Xi by Chinese coverage; it does not yet constitute a shared bilateral norm.
The vocabulary also predates the visit. Chinese diplomacy has for years used the idea of human control in international debates on autonomous systems and military applications of AI. An official 2024 document called for relevant weapons systems to remain under human control and placed ultimate responsibility with people. [9]
Washington therefore carries an existing Chinese governance grammar into a different bilateral relationship.
The visit produced something more concrete. China and the United States agreed to establish a bilateral dialogue on artificial intelligence, with its next meeting due in November, and a bilateral communication channel for AI-related incidents. In parallel, the two militaries agreed to move towards a memorandum aimed at strengthening communication and crisis prevention. [6]
AI thus enters the repertoire of issues for which two major powers consider permanent communication mechanisms necessary.
Chinese coverage places that mechanism within a broader architecture for producing stability.
Ritual, reciprocity and trust
Xi’s arrival at Joint Base Andrews was marked by military honours, the national anthems of both countries and a 21-gun salute. The following day, the White House ceremony included military performances, inspection of honour guards and a flypast by combat aircraft. [7] [8]
The same narrative contains other forms of exchange. Xi recalled the shared memory of China and the United States as allies during the Second World War. He announced that China would invite 100,000 young Americans to take part in exchanges and study programmes over the next five years. He presented the pandas Ping Ping and Fu Shuang, bound for Zoo Atlanta, as “envoys of friendship”. [8]
Pandas, students, the memory of a shared war, military ceremony and an AI incident channel appear to belong to different worlds. Official Chinese coverage nevertheless places them within the same work of relationship-building.
Marcel Mauss is useful here without turning these diplomatic exchanges into a literal application of The Gift. Giving, receiving and reciprocating produce relationships as well as transfers of objects. [16]
Another sociological tradition broadens the reading. Writing about rural Chinese society, Fei Xiaotong used the term chaxugeju (差序格局), often rendered as a “differential mode of association”, to describe relationships extending outwards from each person in concentric circles, with obligations and proximity varying according to the tie. [17] Contemporary research on Chinese bureaucracy continues to show the importance of informal relationships, guanxi, and their interaction with formal structures. [10]
Using chaxugeju as a lens — rather than as a causal explanation of contemporary Chinese diplomacy — helps us attend to the official corpus’s insistence on cultivating relationship, proximity and reciprocity alongside formal mechanisms.
Pandas, youth exchanges, shared memory and protocol do not replace the incident channel. They help build the relational environment within which that channel is meant to operate.
What counts as an incident?
This is where one of the most consequential outcomes of the encounter becomes visible.
China and the United States have agreed to a communication channel for “AI-related incidents”, yet the Chinese sources published so far do not define what an incident is. [6]
Before an incident can be communicated, an event has to be placed within that category.
A data leak caused by an agent?
An AI-assisted cyber operation?
Unexpected system behaviour?
A disinformation campaign?
An autonomous action crossing infrastructures located in both countries?
An emergent effect that becomes visible only months after deployment?
For now, these are analytical possibilities, not categories announced by China or the United States.
Determining what counts as an incident will also mean deciding what information must be shared, what evidence is sufficient, which body has authority to evaluate the event, who bears responsibility and what response is legitimised.
The difficulty grows because an AI system does not necessarily resemble the objects around which earlier communication mechanisms between major powers were built. A ship, an aircraft or a missile usually has a relatively identifiable location, operator and chain of responsibility.
Those mechanisms also have a genealogy. The 1963 Moscow-Washington direct communications link was created to reduce the risk of war and provide immediate communication during a crisis. [12] The 1972 US-Soviet Incidents at Sea Agreement sought to reduce conflict arising from accidents, miscalculation or communication failures between military forces. [13]
Chinese diplomacy itself invites the present competition to be read in that register. Days before the visit, Ambassador Xie Feng wrote in People’s Daily that AI competition should not raise a new “Silicon Curtain” or turn into an AI version of the “Star Wars” programme. [11]
Those metaphors do not, by themselves, justify describing the present as a new Cold War. They do show that Chinese diplomatic discourse itself draws on that historical repertoire to articulate the risks of technological competition at the same moment that it proposes new mechanisms for managing them.
Current objects place strain on that inherited grammar.
An incident involving commercial agents, globally available models, private infrastructures or emergent behaviour may be much harder to attribute to a state, a company, a person or the operation of the system itself.
The answer remains open. The two states have judged it necessary to create an institution before publishing a precise definition of its object.
In Weberian terms, creating a channel means bureaucratising uncertainty. An ambiguous event must be made into a recognisable case, enter a procedure, be assigned to a competent authority and produce an expected form of response.
Bowker and Star help us understand how the category comes into being. Weber helps us see how that category can become administrative capacity.
An incident channel needs more than a line of communication. It needs shared categories for recognising events, bureaucracies capable of processing them and a relationship stable enough for the act of reporting not to be interpreted immediately as a new aggression.
In November, we will begin to see what kind of classification emerges.
This is less a semantic matter than it may appear. Defining an incident also defines which events enter the field of political responsibility.
An infrastructure that loops back on itself
African compute capacity, bureaucratic use cases and the new China-US AI incident channel appear to be very different objects. Yet all three show that access to a technology is not the same as retaining the capacity to act on it.
The six African countries studied by UNDP are trying to turn locally installed compute into their own institutional capacity within a global ecosystem in which chips, cloud services, models, finance and standards depend on transnational chains they do not fully control. [2]
The circuit therefore returns to its starting point.
Will future international mechanisms on AI risks and incidents expand the ability of third countries to govern AI systems?
Or will they establish the categories and conditions within which that capacity must be exercised?
The answer is still being made.
Agency needs machines, but it also needs institutions. It needs accumulated knowledge, but also classifications that allow that knowledge to circulate. It needs safety mechanisms, but also relationships that allow those mechanisms to be used when something fails.
In November, the new China-US AI dialogue will provide a first concrete place from which to observe what the two sides call an incident, who can declare one, what evidence that classification requires, and what obligations it begins to produce.
Because governing a technology also requires building a social relationship around it.
Sources
[1] UNDP. From Connectivity to Capability to Competitiveness. 25 September 2026. https://www.undp.org/africa/news/connectivity-capability-competitiveness
[2] UNDP. Local AI compute infrastructure in Africa: Lessons from six country-hosted deployments on governance and effective use. 17 September 2026. https://www.undp.org/africa/publications/local-ai-compute-infrastructure-africa-lessons-six-country-hosted-deployments-governance-and-effective-use
[3] UNDP. Africa’s AI Moment: Build the Infrastructure, Own the Future. 18 December 2025. https://www.undp.org/africa/blog/africas-ai-moment-build-infrastructure-own-future
[4] Apolitical. New tool helps governments stop starting from scratch with AI. 8 July 2026. https://apolitical.co/en/pages/gov-ai-navigator-launch-news
[5] Ministry of Foreign Affairs of the People’s Republic of China. President Xi Jinping Holds Talks with U.S. President Donald J. Trump. 25 September 2026. https://www.fmprc.gov.cn/esp/wjdt/wshd/202609/t20260925_12031248.html
[6] Ministry of Foreign Affairs of the People’s Republic of China. 中美达成八点成果共识 [China and the United States reach eight outcomes and understandings]. 26 September 2026. https://www.mfa.gov.cn/zyxw/202609/t20260926_12031616.shtml
[7] Xinhua. 习近平主席抵达华盛顿 80秒看美方高规格迎接. 24 September 2026. https://www.xinhuanet.com/20260924/14020ed2f8e443158ce46cda72f483e7/c.html
[8] Ministry of Foreign Affairs of the People’s Republic of China. President Xi Jinping Attends the Arrival Ceremony at the White House Hosted by U.S. President Donald J. Trump. 25 September 2026. https://www.fmprc.gov.cn/mfa_eng/xw/zyxw/202609/t20260925_12031176.html
[9] Ministry of Foreign Affairs of the People’s Republic of China. Position Paper of the People’s Republic of China on Regulating Military Applications of Artificial Intelligence (AI). 2024. https://www.fmprc.gov.cn/eng/zy/wjzc/202405/t20240531_11367523.html
[10] Zhou, Xueguang; Sui, Yuze. What Can the Research on Chinese Bureaucracy Do for Organization Theory? Management and Organization Review, 2025. https://www.cambridge.org/core/journals/management-and-organization-review/article/what-can-the-research-on-chinese-bureaucracy-do-for-organization-theory/F7CD283D1CAD14E38A075C45AF8B1733
[11] Xie Feng. 推动中美建设性战略稳定关系落地落实 [Advancing a constructive China-US relationship of strategic stability]. People’s Daily, 23 September 2026. https://paper.people.com.cn/rmrb/pc/content/202609/23/content_30182619.html
[12] U.S. Department of State, Office of the Historian. Editorial Note: Direct Communications Link between Washington and Moscow. 1963. https://history.state.gov/historicaldocuments/frus1961-63v05/d333
[13] U.S. Department of State. Incidents at Sea Agreement. 25 May 1972. https://2009-2017.state.gov/t/isn/4791.htm
[14] Weber, Max. Economy and Society. Theoretical reference for legal-rational bureaucracy.
[15] Bowker, Geoffrey C.; Star, Susan Leigh. Sorting Things Out: Classification and Its Consequences. MIT Press.
[16] Mauss, Marcel. Essai sur le don / The Gift. Theoretical reference on reciprocity and obligation.
[17] Fei Xiaotong. From the Soil: The Foundations of Chinese Society. University of California Press, 1992 [1948].