The Inheritance and the Algorithm

01

Introduction

Almost every serious model of AI-enabled government now on offer was conceived inside an advanced state. The British government's GOV.UK assistant, Estonia's machine-readable administration, Singapore's government engineering corps and the transnational vision of the Agentic State share a premise their authors rarely need to state: authority is already settled. Somebody holds it by a recognised title, exercises it through rule-bound offices and leaves a record that society can inspect and contest. In such states AI-enabled government is an upgrade to a machine that already works, and the debate about it is a debate about how much of the machine to hand over.

The states in which most of humanity lives received no such settlement. Most of them were placed upon their societies rather than grown from them. Colonial administrations assembled them for extraction and order and handed them on, largely intact, at independence; mandates, client relationships, conditional lending and post-war reconstruction then added further layers written by others. Generations of scholarship, from Subaltern Studies to the comparative politics of state capacity, have documented the result: a persistent gap between state and society, and a state experienced as alien, obeyed where it can compel and bypassed where it cannot.

This paper asks a question that follows directly from that record and that the literature on AI in government has not yet posed: why does AI-enabled government designed for advanced states fail in fragile ones, and under what conditions could it succeed? The question concerns governing with AI, meaning AI built into the work of government and, in the end, into public office itself. It does not concern the governance of AI, the regulation of models and their risks, although the two questions meet at several points, most sharply where the supply of models is itself an instrument of other states' policy.

The argument advances four claims. First, the failure is constitutional before it is technical. Imported designs assume an impersonal authority that the externally authored state never built. They therefore enter offices whose formal chart and actual flow of power diverge, and there they are captured, bypassed or ruptured. The paper calls this prospect the second transposition: after the export of institutions comes the export of the reasoning that will run inside them.

Second, this is the latest instance of a long historical pattern. Every decisive technology, from the ocean-going ship to the steam engine and the telegraph, has handed its first holders an advantage and the temptation to turn it into rule over others. Where societies authored their own adoption of the new technology, the outcome reversed. The variable that decided between the two outcomes was never access to the technology; it was authorship of the institutions that used it.

Third, the theoretical toolkit best suited to diagnosing this condition survives the age of AI as diagnosis but fails as programme. That toolkit is the philosophy of difference from Foucault, Derrida and Deleuze, anchored in Gramsci and carried into postcolonial thought by Said, Spivak and the Subaltern Studies collective. It explains why the transposed state alienates, and it now explains how training corpora, algorithmic governmentality and procurement-led mimicry reproduce that alienation. It offers no account of how an accountable office is to be constituted, least of all an office that a machine might hold.

Fourth, AI is a double edge. It offers fragile states a historic chance to supply what external state-building withheld: consistent, recorded and impersonal authority, available in citizens' own languages, at a speed that took advanced states centuries. It also threatens to weave them into an algorithmic order authored, owned and revocable elsewhere. Which edge prevails depends on authorship: whether design begins by decoding local authority and ends by constituting local offices, or begins with a data platform.

The paper makes three contributions. It joins two literatures that have rarely met, the historical sociology of the transposed state and the emerging study of AI in public administration, and it identifies the variable that the second has so far treated as constant, namely the settledness of authority. It specifies a mechanism at the level of the public office, with propositions whose observable implications can be checked against deployments now under way. And it moves from critique to design, proposing a method by which AI could enter government as part of the constitution of authority rather than as an accelerant of its absence.

A note on method is in order. This is a theory-building paper. It argues a position from historical evidence, comparative scholarship and public cases, and it reports no field results. Pakistan serves as the reference case because its institutional history displays the externally authored condition with unusual clarity and because it is the case the author knows from inside; the claims are framed so that they can be tested elsewhere. The argument draws on the author's wider work on controlled chaos, the Anarchon state and the Synaptic State (Warraich 2024; forthcoming; unpublished).

The paper proceeds as follows. Section 2 sets out the historical pattern that links technological advantage to the external authorship of states. Section 3 defines the externally authored state and reconstructs its inheritance, with Pakistan as the reference case. Section 4 presents the toolkit of difference and its postcolonial translation, and Section 5 tests that toolkit against the AI condition. Section 6 examines the current paradigms of AI-enabled government and their shared blind spot. Section 7 states the mechanism of failure, its propositions, rival explanations and counter-evidence. Section 8 develops the double edge. Section 9 sets out the constructive response developed in the Synaptic State programme, and Section 10 draws implications for frontier AI developers. Section 11 states the limits of the argument and a research agenda, and Section 12 concludes.

02

Every advance is also a temptation

Every decisive technology has arrived with two possibilities. It can raise the capacity of everyone it reaches, or it can give its first holders an advantage so large that they are tempted to convert it into rule over others. The historical record shows that the second possibility has usually prevailed, and that it has done so less through conquest alone than through the authorship of other societies' institutions.

When Iberian navigators combined ocean sailing with ship-borne artillery at the end of the fifteenth century, the advantage became a claim over most of the world almost at once. By the Treaty of Tordesillas in 1494, Spain and Portugal divided the lands beyond Europe between them along a meridian drawn through the Atlantic. No one who lived in those lands was consulted, and the line itself became the first template of external authorship: a division of jurisdiction decided in one hemisphere and imposed on another.

The Industrial Revolution repeated the pattern on a scale the world had not seen. Daniel Headrick showed how a small set of nineteenth-century technologies became the tools of empire (Headrick 1981). Steam gunboats carried European power up rivers that had protected continental interiors for centuries. Quinine prophylaxis let Europeans survive where malaria had long defended the people who lived there. Breech-loading rifles and then the machine gun made small forces decisive against large ones, and telegraph cables and railways bound conquered territories to distant capitals. By 1914, on Edward Said's count, Europe held roughly 85 per cent of the earth as colonies, protectorates, dependencies, dominions and commonwealths (Said 1993). Britain alone governed about a quarter of the world's land and people at the height of its empire. A popular verse of 1898 caught the confidence of the age:

Whatever happens, we have got
The Maxim Gun, and they have not.

Hilaire Belloc, The Modern Traveller (1898)

The pattern did not end with formal empire; it changed form. After 1918 the mandate system placed former Ottoman and German territories under the tutelage of the victors, inside borders drawn in European ministries. The Cold War turned oil, aircraft, nuclear weapons and intelligence technology into instruments for keeping client governments in place. From the 1980s, loans arrived with conditions attached, and whole templates of administration were exported with them; James Ferguson's study of a development project in Lesotho showed how such templates could expand bureaucratic power while presenting themselves as technical and apolitical (Ferguson 1990). After 2001, states were rebuilt from outside in Afghanistan and Iraq. The Afghan republic, two decades in the building, dissolved within weeks in August 2021.

Each advance also produced a doctrine that made the advantage appear as a gift: the civilising mission, then tutelage and modernisation, then development and good governance. Each doctrine contained real goods, and it would be a poor history that denied them. Each also left behind states whose institutions had been written by others, and whose societies had not been party to the writing.

The record contains a second lesson, and it matters as much as the first. Where a society took the new technology into its own hands and authored its own adoption, the outcome reversed. Meiji Japan studied Western industry, law and administration after 1868, rebuilt its state on its own terms and was never colonised. South Korea, Taiwan and Singapore later used imported technology to build some of the most capable states in the world. Their case is not a simple one. Korea inherited an unusually penetrative bureaucracy from Japanese colonial rule, and Atul Kohli has argued that this inheritance shaped the developmental state that followed (Kohli 2004); Taiwan's colonial experience was in important respects similar. Yet that qualification strengthens the point rather than weakening it. What made the inherited machinery productive was the domestic political authorship that later took possession of it. India's construction of its own digital public infrastructure in the present century belongs to the same family of cases. The decisive variable was never access to the technology. It was who authored the institutions that used it.

Artificial intelligence is the most general advance of all. Earlier technologies extended muscle, distance or speed; AI extends judgement, the faculty through which states classify, decide and allocate. It can shape how a ministry reads a case, drafts a rule or allocates a budget without a single foreign official setting foot in the country. The capability is concentrated in a handful of companies, in very few countries, running on chips and cloud capacity that fewer still control. Access is already political. In June 2026, United States export controls led one leading developer to suspend worldwide access to its most capable models for about three weeks, until the controls were lifted (Anthropic 2026c).

The temptation of every earlier advance is therefore present again, in its most intimate form: influence without occupation, exercised through the reasoning inside other states' institutions. The builders of this technology nonetheless face a choice that their predecessors never made explicit, and some of them have already shown that they will refuse uses they judge dangerous, even at commercial cost. The question this paper pursues is how AI enters the governments of the world's fragile states, and whether it repeats the pattern of the last five centuries or reverses it.

03

The externally authored state

Most of the states that appear on today's fragility lists share a history that the lists do not record: their institutions were authored, in large part, by others. The fragility indices of the OECD and the World Bank measure outcomes such as conflict, institutional weakness and exposure to shocks. They do not measure the origin of the institutions whose weakness they record, and the analytical category used in this paper is designed to capture that origin.

Definition and scope

An externally authored state is a state whose core institutions were designed, imposed or decisively reshaped by more powerful outside actors, rather than grown from a settlement within its own society. The category is wider than the colonial one, and it runs across every period of modern history. It includes the colonies and protectorates of the European empires and the mandates and borders of the post-Ottoman settlement. It includes the client states of the Cold War, states remade by structural adjustment and donor conditionality, and states rebuilt after war by international coalitions. Bosnia and Herzegovina's constitution, for instance, was negotiated at a United States air base in Ohio in 1995 and adopted as an annex to the Dayton peace agreement. The forms of external authorship differ; the resulting condition is shared.

The concept is related to, but distinct from, three established ones. It is not identical to fragility, which describes performance rather than origin; some externally authored states are capable, and some fragile states authored their own institutions and lost them to war. It is close to Robert Jackson's quasi-states, which possess juridical sovereignty recognised by the international system without the empirical statehood to sustain it (Jackson 1990), but it locates the problem in the authorship of institutions rather than in the gap between recognition and capacity. And it generalises the postcolonial state of Alavi, Chatterjee and Mamdani beyond the colonial period, so that the same analytical lens applies to Kabul in 2004 as to Lahore in 1947.

Authorship in this sense is a matter of degree. A state may have authored some institutions and inherited others; its courts may be external in origin while its party system is indigenous. The concept is therefore best treated as a continuous property of institutions, measured office by office, rather than as a binary label for whole countries. That choice matters for the mechanism developed in Section 7, which operates at the level of the public office.

The missing settlement

The modern Western state did not arrive as a design. Weber's legal-rational authority, with office separated from person, rule applied without regard to status and decision fixed in the file, was the residue of centuries of contest over taxation, representation, religion and war (Weber 1978). Its impersonality rested on a social settlement rather than on paperwork. Those who obeyed came to recognise the office as theirs, or at least as bound by rules they could invoke against it. Michael Mann's distinction between despotic power, the power of state elites to act without negotiation with civil society, and infrastructural power, the capacity of the state to penetrate society and implement decisions through it, captures the result: the settled state combines limited despotic power with high infrastructural power (Mann 1984).

External authorship exported the offices and the paperwork; it could not export the settlement. The colonial state in South Asia was built to collect revenue and keep order, and its architecture followed those purposes. Its district officer combined revenue collection, magistracy and oversight of police, while below him a layer of intermediaries carried the state into the village. In Punjab the revenue settlements created the office of the village headman, the lambardar, which survives today. Authority looked impersonal at the top and was personal, brokered and negotiated below. Mahmood Mamdani's account of the bifurcated colonial state in Africa describes the same doubling: civil law for the citizen of the towns and customary authority for the subject of the countryside, a form he called decentralised despotism (Mamdani 1996). Peter Ekeh had earlier described the moral consequence, a split between a civic public from which citizens extract benefits without obligation and a primordial public to which they owe moral duty (Ekeh 1975).

The deepest diagnosis of this arrangement came from within South Asia. Ranajit Guha described colonial rule as dominance without hegemony: it governed by coercion supplemented by collaboration and never won the consent of the governed (Guha 1997). Hamza Alavi argued that the postcolonial state in Pakistan and Bangladesh inherited an "overdeveloped" apparatus, whose bureaucratic-military oligarchy was more powerful than any social class it ruled and relatively autonomous of all of them (Alavi 1972). In Mann's terms, such a state combines high despotic power with low infrastructural power. Joel Migdal later generalised the pattern as strong societies and weak states, in which social control is fragmented among local power-holders with whom the state must bargain rather than command (Migdal 1988).

The economic and comparative literature confirms that these inheritances persist. Acemoglu, Johnson and Robinson linked contemporary institutional quality to the type of institutions colonisers established (Acemoglu, Johnson and Robinson 2001). Banerjee and Iyer found that Indian districts where the British vested land revenue in landlords still showed lower agricultural investment and public goods provision a century and more later than districts where revenue was collected from cultivators (Banerjee and Iyer 2005). Dell traced the effects of a colonial forced-labour system in Peru across four centuries (Dell 2010), and Lange showed that the extent of direct as opposed to indirect British rule shaped the capacity of post-colonial states (Lange 2009). Pierre Englebert's study of African states found that legitimacy, and with it developmental performance, was higher where post-colonial state institutions were congruent with pre-colonial ones (Englebert 2000). The weight of evidence is that institutional origin is not a historical curiosity; it continues to shape what offices can do.

Pakistan as reference case

Pakistan displays the pattern clearly. Independence changed the flag before it changed the form. The country's first constitution was the Government of India Act 1935, adapted as an interim constitution and in force until 1956. The colonial Frontier Crimes Regulation of 1901 governed the former tribal areas until their merger with Khyber Pakhtunkhwa in 2018. In Punjab the revenue headmen created by colonial land settlement still hold office. Constitutional order was broken by military rule in 1958, 1969, 1977 and 1999, a cycle Ayesha Jalal traces to the political economy of defence in the new state (Jalal 1990). The author's doctoral research reads this history as controlled chaos: instability not as the failure of the state but as a medium through which sovereign power reproduces itself, a condition later theorised as the Anarchon state (Warraich 2024; forthcoming). For present purposes the point is narrower. In Pakistan the formal chart of authority and its actual flow have diverged, persistently and by design, for most of the state's existence.

Mimicry and political society

Two concepts explain why this divergence matters for any technology of government. The first is mimicry. Homi Bhabha described the colonial subject's mimicry as "almost the same, but not quite" (Bhabha 1994). Institutional theory observed organisations copying the forms of the successful in order to secure legitimacy (DiMaggio and Powell 1983). Pritchett, Woolcock and Andrews joined the two in their account of isomorphic mimicry, in which states adopt the forms of capable institutions without their function and so, in their phrase, come to look like states (Pritchett, Woolcock and Andrews 2013). The second is Partha Chatterjee's distinction between civil society and political society (Chatterjee 2004). Civil society is the small domain of rights-bearing citizens who meet the state through law. Political society is the vast domain of populations who meet it through negotiation, exception and intermediaries.

The result is the condition this paper calls the inheritance: the formal shell of impersonal authority laid over a society governed in practice through relational, negotiated and often personal power. What the externally authored state lacks is not offices but impersonal authority embedded in society, the combination of autonomy and embeddedness that Peter Evans found in developmental states that worked (Evans 1995). In such a state the formal chart of authority and the actual flow of authority diverge, and any technology of government that assumes they coincide will meet the second, not the first.

04

The toolkit of difference: from Paris to the subaltern

The most powerful critique of the transposed state did not originate in political science. It came from a philosophy that refused to treat difference as deviation from a norm. For Derrida, meaning is produced by différance, by difference and deferral within a system of signs, with no stable origin or centre from which other terms derive (Derrida 1982). Deleuze set out to think difference in itself rather than as a gap between identities, so that the copy is no longer a failed version of a model (Deleuze 1994). Read politically, both arguments dissolve the claim that one society's institutions are the standard and every other society's an approximation.

Foucault supplied the account of power that made this political. Knowledge does not merely describe its objects; discourse constitutes them, and the power to classify is the power to govern (Foucault 1977). His later lectures named governmentality, the conduct of conduct through population, statistics and administrative technique, as the modern art of rule (Foucault 2007). The panopticon, the prison in which the inmate behaves as if always watched, became his figure for disciplinary power that works through visibility rather than force.

Gramsci, writing earlier and from prison, gave the tradition its political sociology. Rule is hegemony, coercion armoured by consent, and consent is produced in the "common sense" (senso comune) that intellectuals organise for a ruling bloc (Gramsci 1971). The subaltern are those excluded from that bloc and from the history it writes. A passive revolution is change engineered from above that preserves the core of the old order.

The postcolonial translation of these ideas was the most important intellectual event in the study of the non-Western state. Said's Orientalism applied Foucault's discourse analysis to the West's knowledge of the East and showed that knowledge to be an instrument of rule (Said 1978). The Subaltern Studies collective, founded by Ranajit Guha, rewrote South Asian history from below; its founding statement argued that both colonial and nationalist historiography had been written by elites and for elites (Guha 1982). Chatterjee showed that anticolonial nationalism had been a "derivative discourse", adopting the colonial state's categories even as it fought the colonisers (Chatterjee 1986). Spivak, who had translated Derrida's Of Grammatology into English, asked whether the subaltern could speak at all within the archive that represented her (Spivak 1988). Chakrabarty concluded that Europe's categories are "indispensable and inadequate" for thinking political modernity in India, and called for Europe to be provincialised (Chakrabarty 2000).

The tradition was never only South Asian. Fanon warned that a national bourgeoisie would inherit the colonial state and use it as the coloniser had (Fanon 1963). Quijano traced a coloniality of power that outlived formal empire in Latin America (Quijano 2000). Mbembe described the postcolony as a regime of power improvised from the colonial inheritance (Mbembe 2001).

From this body of work one lesson matters most for the present argument: difference is constitutive, not residual. The externally authored state is not an early or failed edition of the European state. It is a different formation, produced by the transposition itself, and the universal categories used to design it are themselves provincial. Any project that measures such a state against an advanced-state template, and treats the gap as a deficit to be closed with better technique, repeats the founding error of the colonial administrator.

05

Does the toolkit hold in the age of AI?

The postcolonial toolkit was forged against the archive, the census, the district office and the nationalist party. It is fair to ask whether it can read a world of foundation models, cloud contracts and autonomous agents. A growing body of work has begun to ask the same question. Mohamed, Png and Isaac proposed decolonial theory as a form of sociotechnical foresight for AI (Mohamed, Png and Isaac 2020). Birhane described the "algorithmic colonization" of Africa by systems built elsewhere and imported under the banner of progress (Birhane 2020). Kwet analysed a digital colonialism in which ownership of the infrastructure confers economic and political control (Kwet 2019), and Couldry and Mejias argued that the appropriation of human life through data is a new stage of colonial extraction (Couldry and Mejias 2019). This literature establishes that the critique travels. The answer offered here is more divided: three of the toolkit's instruments become sharper in the AI condition, and three of its silences become disabling.

What survives as diagnosis

The archive becomes the corpus. Said's Orientalism was an archive: a body of texts through which the West knew, and therefore ruled, the East. Its successor is the training corpus, the body of text through which a model knows the world, and the distribution of that corpus is not neutral. In 2026 Urdu accounted for about 0.03 per cent of the web pages gathered by Common Crawl, against roughly 0.7 per cent each for Arabic and Persian and about 40 per cent for English (World Bank 2026). Spivak's question returns in statistical form. The subaltern is silenced not by a censor but by a sampling ratio, and what a model cannot read it cannot represent, except through the categories of those it can.

Common sense becomes a model. Gramsci located hegemony in common sense, the taken-for-granted judgements that make one order seem natural. A general-purpose model is, among other things, a machine for producing common sense at scale. It answers the official's question, drafts the memorandum and summarises the case. When that common sense is learned overwhelmingly from the text of the advanced world, the model acts as an intellectual of a hegemony that the society it serves did not choose. It is an imported intellectual rather than an organic one, and it is persuasive precisely because it is fluent.

Governmentality becomes algorithmic, and mimicry acquires a new instrument. Rouvroy and Berns described algorithmic governmentality: government through data profiles that act on behaviour while bypassing the subject's own reflection (Rouvroy and Berns 2013). Deleuze had anticipated the shift from disciplinary enclosures to societies of control that modulate behaviour continuously (Deleuze 1992). The panopticon no longer needs a tower; it needs a data pipeline. Mimicry, too, gains a new instrument. A state can now procure the appearance of AI-enabled government, whether a chatbot, a dashboard or even an artificial "minister", without constituting the authority to act on what these systems produce. Albania's appointment in September 2025 of an AI system, Diella, as the minister responsible for public procurement, announced as a means to make public tenders free of corruption, shows how quickly the symbol of office can run ahead of the institution (Semafor 2025).

What fails as programme

The toolkit has no positive theory of institutions. Its genre is critique. It deconstructs authority brilliantly and designs it rarely. Chatterjee describes political society with precision but does not propose its offices; Chakrabarty provincialises Europe without specifying the counter-state. The decolonial AI literature shares the limitation: it identifies extraction, bias and dependence with great clarity, and it recommends reflexivity, participation and solidarity, but it rarely specifies the institutional form through which a society would own and answer for an AI-enabled decision. In an age when the architecture of government is being rewritten by firms and their clients, a theory that can only unmask designs leaves the field to those who make them.

It has no account of non-human authority. The tradition's subjects are human: coloniser and colonised, elite and subaltern, citizen and population. AI introduces something the tradition never theorised: a non-human bearer of decision competence, which drafts, recommends and may increasingly decide. Who authors a decision that a model drafts and an official signs? Where does responsibility lie when the model is replaced, or when its supplier withdraws access? The toolkit cannot answer, because its concepts of agency, voice and representation presuppose a speaking human subject.

Its suspicion can end in refusal, and refusal is no longer neutral. The tradition's justified distrust of the state and of Western technique can harden into a stance of refusal. In the AI condition refusal does not preserve difference; it surrenders the field. Pakistan already accounts for the largest number of conversations with Anthropic's Claude in its World Bank region (World Bank 2026). Officials and citizens will use foreign models with or without a public design, so the real choice is not between AI and no AI. It is between AI that is authored locally and AI that arrives unauthored.

The toolkit therefore holds, but only as half of what is needed. Its diagnostic instruments, namely difference, hegemony, the archive, governmentality and mimicry, should be kept and sharpened. What must be added is what it never supplied: a constructive theory of authored office, capable of including non-human incumbents without surrendering human authorship. The move required is from deconstruction to authorship, and Section 9 sets out an attempt to make it.

06

The current paradigms and their shared blind spot

Three paradigms now organise most thinking and practice about AI-enabled government outside the advanced states. Each has real merit, and each treats the state as given and AI as the input. In the externally authored state, however, the state itself is the problem to be constituted. A fourth body of work, the public administration literature on digital discretion, sharpens the point by showing where authority moves when systems are automated.

The Agentic State

The most ambitious public framework is The Agentic State, a vision paper by Luukas Ilves, Manuel Kilian and colleagues, launched at the Tallinn Digital Summit in October 2025 with more than twenty contributors (Ilves et al. 2025). It maps how agentic AI could transform service design, internal workflows, policy-making, compliance, crisis response and procurement. Follow-on work decomposes government into some seventy recurring functions, and an observatory tracks several hundred initiatives (Ilves 2026). Its central image is a state in which agents handle complexity and scale, while human officials concentrate on strategy, political accountability and decisions that require judgement.

The framework deserves credit for taking agency seriously and for refusing to treat AI as a chatbot. It even devotes attention to low- and middle-income countries, which it suggests can leapfrog by opening interfaces to agents or by starting with voice-first services. Yet it is shallow in exactly the dimensions that decide outcomes in the externally authored state, for four reasons.

It models the state as functions, not as authority. A function presupposes an authority that assigns it, funds it and answers for it. In the externally authored state that authority is contested, brokered and often located elsewhere than the chart indicates. Decomposing government into seventy functions describes a production system; it does not describe who rules.

Its accountability rests on the human layer that the externally authored state lacks. The vision keeps political accountability with human officials. Where the human chain of accountability is itself the deficit, preserving it leaves the deepest problem untouched.

It treats difference as a stage. Leapfrogging is the language of modernisation theory: the fragile state is behind on the same road and can skip ahead. The argument of this paper is that it travels a different road, shaped by external authorship.

It is silent on what a decoding of power must cover. The framework says little about informal and customary authority, religion, ethnicity and language, regional geopolitics or dependence on foreign suppliers. Its reference cases, among them Estonia, Finland, Italy, Portugal, the United Arab Emirates and a mobilised wartime Ukraine, are high-capacity or exceptionally centralised states.

Sectoral deployment by frontier developers

The second paradigm is practical rather than theoretical: frontier developers and their partners signing frameworks with governments and deploying in health, education and administration. Anthropic's three-year memorandum of understanding with Rwanda gives government developer teams access to its models alongside health and education programmes (Anthropic 2026a). Its Joint Declaration with Kenya creates an umbrella under which ministries design their own pilots, with Kenyan law governing any use of data (iAfrica 2026). Its partnership with the Gates Foundation commits USD 200 million over four years, the largest part of it to health in low- and middle-income countries (Anthropic 2026b). Microsoft and G42 announced a USD 1 billion package for digital infrastructure in Kenya (Microsoft 2024), and Google signed a memorandum with the Government of Pakistan that includes AI-powered analytics for the Federal Board of Revenue (Radio Pakistan 2026). These are serious commitments, and some are carefully drafted; the Kenyan data clause is a model of its kind.

Their common design, however, is a transfer of capability: access, credits, compute and skills. It assumes that the receiving institution can absorb the capability and answer for its use. Observers have already asked who is accountable once Rwanda's government teams build on a foreign model (Sofi 2026), and Kenya's declaration had announced no funding, tools or timeline in the weeks after its signature (iAfrica 2026). Neither point is a criticism of intent. Both mark the place where the institution, not the technology, is missing.

The legibility machine

The third paradigm is rarely named but everywhere assumed: AI as the perfect instrument of what James Scott called seeing like a state (Scott 1998). High-modernist schemes failed, Scott argued, because they made society legible from above and destroyed the practical local knowledge, mētis, on which it actually ran. In the colonial world the census and the cadastre were the original legibility machines; Nicholas Dirks showed that the colonial census helped fix caste as a rigid administrative category (Dirks 2001). AI multiplies this power. Classification of citizens by ethnicity, sect, risk or eligibility can harden identities that were fluid, and it can do so at a speed no colonial census could reach.

Where discretion moves

The public administration literature adds a fourth element that the three paradigms overlook. Lipsky showed that policy is made in practice by street-level bureaucrats exercising discretion at the counter (Lipsky 1980). Bovens and Zouridis then observed that information technology was turning street-level bureaucracies into system-level bureaucracies, in which discretion migrates from the official at the counter to the designers of the system and the authors of its decision rules (Bovens and Zouridis 2002). Young, Bullock and Lecy extended the analysis to AI, describing an artificial discretion whose effects on administrative quality depend on the task and the institutional setting (Young, Bullock and Lecy 2019). Legal scholars in the United States have shown that automated administration raises acute problems of due process and legitimacy even where authority is settled (Citron 2008; Calo and Citron 2021).

Applied to the externally authored state, this literature yields a precise consequence. When an imported AI system enters an office, discretion moves to the level of system design, and the designers are usually outside the state. The second transposition is therefore not a metaphor. It names a measurable migration of discretion from local offices to foreign system authors, through procurement contracts rather than through conquest.

All four bodies of work share one blind spot. They begin from the advanced state's assumption that authority is settled and ask how AI can serve it, or how it alters the distribution of discretion within it. None begins from the externally authored state's actual question: how authority can be constituted, made impersonal and owned by the society it governs, with AI as part of the answer rather than as an accelerant of the old gap.

07

The mechanism of failure

Unit of analysis

The argument is pitched at one level: the public office that performs a defined function in a defined territory over a defined period. The office is where authority, record and remedy meet, and it is where external authorship is most visible; the district office remains the place where the inherited state meets its population. National strategies and model benchmarks are the wrong units. A national strategy can succeed on paper while every office ignores it, and a model can top every benchmark while the office around it remains unaccountable. Choosing the office as the unit also makes the argument testable, because offices within one state vary in how far their formal and actual authority diverge.

The mechanism

An AI system that enters a public office acts through that office's actual authority structure, not its formal chart. Where the two coincide, as they broadly do in advanced states, the system amplifies a working institution. Where they diverge, as they do across much of the externally authored state, one of three outcomes tends to follow.

Capture. The system's outputs are taken over by whoever holds real power in the office: the patron, the broker or the faction. They become a new instrument of the old order, or in the author's terms a new medium of the Anarchon condition, in which instability and discretion reproduce power rather than threaten it.

Bypass. Officials and citizens route around the system. Dashboards are filled and ignored, and recommendations are noted and not acted upon. The pilot report records success while the service stays the same, which is isomorphic mimicry with a new instrument.

Rupture. The system enforces the formal rule against the negotiated order that actually keeps public peace: a rule about land, eligibility or debt applied without the accommodations that made it bearable. Legitimacy collapses, and the system is withdrawn or becomes a grievance.

These outcomes can be derived from the comparative study of informal institutions. Helmke and Levitsky classified informal institutions by whether their goals converge with those of formal institutions and whether the formal institutions are effective, yielding complementary, accommodating, competing and substitutive types (Helmke and Levitsky 2004). An AI system strengthens whichever rules it is built to enforce, and imported systems are built to enforce the formal chart. Where informal rules accommodate the formal ones, bending them without openly violating them, the system tends to be captured, because those who already manage the accommodation also manage the system. Where formal rules are ineffective and informal rules substitute for them, the system tends to be bypassed, because no one depends on its outputs. Where informal rules compete with formal ones, the system tends to produce rupture, because it enforces the rule that the local settlement had suspended. Only where informal rules are complementary, which is to say where formal and actual authority already coincide, does the system amplify a working office.

Success therefore requires two conditions together. The system's assumed authority must match the office's actual authority, and the design must create a path by which that authority becomes more impersonal, recorded and contestable over time. The first condition alone produces a smarter instrument of the old order; the second alone produces rupture.

Five mismatches

The gap between assumed and actual authority appears in five recurrent mismatches between what a design assumes and what an office faces.

The geopolitical layer

A sixth condition operates above the office. The stack on which AI-enabled government runs, from chips and cloud to frontier models and the rules that govern them, is concentrated in a few advanced states and a handful of firms. Access is therefore a matter of other states' policy, as the June 2026 suspension showed (Anthropic 2026c). Benjamin Bratton's account of planetary-scale computation as The Stack describes an architecture in which sovereignty is exercised through layers of infrastructure rather than through territory alone (Bratton 2015). Regional geopolitics adds a second layer of dependence. For a state such as Pakistan, choosing an AI supplier is also a choice of alignment among American firms, Chinese infrastructure and Gulf capital. A design that ignores this layer can make an office hostage to decisions taken in another capital.

Propositions

The mechanism implies five propositions, each with observable implications that distinguish it from the rival explanations considered below.

Proposition 1 (divergence). The greater the divergence between an office's formal and actual authority, the lower the probability that an imported AI system will change the office's outputs in the direction its designers intended. Observable implication: identical systems deployed in offices of the same state will perform differently, and the difference will track measures of divergence such as the share of decisions taken outside the formal chain, the incidence of informal intermediaries and the completeness of the office's own records, more closely than it tracks funding or connectivity.

Proposition 2 (type of informal institution). The form of failure depends on the type of informal institution that governs the office. Observable implication: capture will predominate where accommodating informal rules prevail, bypass where substitutive rules prevail and rupture where competing rules prevail.

Proposition 3 (authorship). Systems whose design began with a documented mapping of the office's actual authority, and whose decision rules were authored or substantially revised within the state, will show lower rates of capture, bypass and rupture than systems imported as templates, controlling for funding and infrastructure. Observable implication: locally redesigned deployments will outperform imported ones in high-divergence offices, while the two will perform similarly in low-divergence offices.

Proposition 4 (migration of discretion). Imported systems move effective discretion from local officials to external system authors. Observable implication: after deployment, the decisions that local officials can alter without vendor involvement will contract, and disputes about outcomes will increasingly be resolved by reference to the system's configuration rather than to local rule or precedent.

Proposition 5 (dependence). Offices whose AI capability cannot survive a change of supplier will show greater susceptibility to external shocks of access, pricing or policy. Observable implication: when access is interrupted, as in June 2026, offices without substitution rights will suspend functions, while offices with constituted records and substitution rights will continue them at reduced capability.

Rival explanations

Four rival explanations deserve a fair hearing.

Resources and capacity. Projects fail because they are underfunded and understaffed. This is often true, but well-funded systems also fail. Richard Heeks estimated two decades ago that most e-government projects in developing countries failed totally or partially, and he traced failure to design–reality gaps between a system's assumptions and local conditions (Heeks 2002; 2003). The present argument builds on Heeks but locates the decisive gap in authority, not only in infrastructure, skills or process. Proposition 1 discriminates between the two explanations: if resources explain failure, outcomes should track funding; if authority explains it, they should track divergence.

Infrastructure. Electricity and connectivity are necessary, as the World Bank stresses for Pakistan's region (World Bank 2026). They are not sufficient. An office with power and bandwidth can still capture or bypass a system, and the propositions predict that it will do so where divergence is high.

Elite capture and corruption. These are real and important, but in this account they are a product of the authority gap rather than a rival to it. Capture is one of the three outcomes the mechanism predicts, and Proposition 2 specifies the institutional conditions under which it, rather than bypass or rupture, should occur.

Management method. Andrews, Pritchett and Woolcock propose problem-driven iterative adaptation as the cure for imported designs (Andrews, Pritchett and Woolcock 2017). This paper agrees on locality and iteration, but holds that method is neutral about authority, and that authority is precisely what AI changes. Iteration within an office whose actual authority is unmapped can converge on capture as easily as on reform.

Counter-evidence and scope

The strongest counter-evidence comes from states with colonial or externally shaped histories that nonetheless built powerful digital government: India's digital public infrastructure, Singapore, Rwanda and, in a different register, Estonia after Soviet rule. These cases narrow the claim rather than refute it. Each built its systems under strong domestic authorship, designed and owned by its own state, rather than importing a template. Inheritance is not destiny; the argument concerns imported design, not the capacity of fragile states. The Singaporean and Rwandan cases add a second caution. Capability built under highly centralised authority can coexist with limited contestability, which is why the design proposed in Section 9 joins capability to record and remedy.

The scope conditions follow. The argument applies where three conditions hold together: the office's core institutions are externally authored; its formal and actual authority diverge to a significant degree; and the AI system is imported as a template rather than designed from a mapping of local authority. It makes no prediction for offices in which any of these conditions is absent. The claim would be falsified if imported designs were shown to succeed systematically, without local redesign, in offices where formal and actual authority diverge widely, or if variation in outcomes across offices were explained by funding and connectivity rather than by divergence.

08

The double edge: opportunity, fear and the algorithmic panopticon

The asymmetry of fear

Advanced and fragile states fear AI for opposite reasons, and much confusion in global debate comes from treating their fears as one. The advanced state fears AI because it has an authorship to lose. Its democratic institutions, professional bureaucracy and settled rule of law are achievements that AI could hollow out, through automated decisions no one owns, through concentration of power in firms and through the displacement of the human judgement on which accountability rests. American scholarship has documented these risks inside a settled state: automated systems that profile and punish the poor (Eubanks 2018), decisions taken without the notice and hearing that due process requires (Citron 2008) and an administrative state whose legitimacy is strained by automation (Calo and Citron 2021). The advanced state's debates are therefore about preservation: alignment, safety, regulation and the human in the loop.

The fragile state's deficit is of a different kind. It concerns an authorship never fully gained. Its fear is less of AI than of AI's holders: the danger that, at the moment a new capability appears, it will again be woven into an order designed elsewhere, as it was woven into the imperial order in the age of steam and telegraph. The same technology that the advanced state fears may dissolve its authorship could, for the fragile state, either complete an authorship it never had or foreclose it for another century.

The historic opportunity

For the first time, the capabilities that made the advanced state's authority impersonal can be constituted rather than slowly accumulated. Consistent application of rules, irrespective of who asks, is what patronage corrodes. Institutional memory that survives the rotation of officers is what the externally authored state lacks most. A complete record of decision and reason is what makes remedy possible. Access to the state in a citizen's own language is what the colonial state never offered. Reasoning at a scale and speed that no thin bureaucracy can supply is the newest of these capabilities.

AI can supply each of them, provided they are lodged in offices that society recognises as its own. That is the historic opportunity: fragile states could measure up to advanced states on the merit of what AI makes possible, without having to repeat their history. In Mann's terms, AI can raise infrastructural power without raising despotic power, but only if the office through which it acts is constituted to be recorded, contestable and answerable. The deficiency left by external authorship, the missing impersonal authority, is exactly what a well-constituted AI-enabled office could supply. In the author's Synaptic State argument, the public office becomes the bearer of continuity, and the model becomes a replaceable incumbent of that office.

The algorithmic panopticon

The opposite edge is just as real. Foucault's panopticon worked through visibility within an enclosure. A planetary system of models, platforms and data flows can render whole societies visible to institutions they do not control. Scholars have named the pattern surveillance capitalism (Zuboff 2019), data colonialism (Couldry and Mejias 2019) and digital colonialism (Kwet 2019).

For the fragile state, the danger is a second transposition more intimate than the first. The colonial state transposed institutions; the algorithmic order would transpose the reasoning inside them. An official's draft, a citizen's eligibility and a district's priorities would be shaped by models trained on other societies' text, hosted under other states' law and revocable by other governments' decisions. Dependence would be cognitive as well as technical: a generation of officials trained to reason through a foreign system will find it harder to reason without it. Without design for local authorship, such an order could fix the fragile state's subordinate place for another century, on terms harder to see, and therefore harder to contest, than those of formal empire.

The question is no longer whether AI will reach the world's fragile states. It already has. The question is who will author the states it arrives in.

Four futures

Two variables decide which edge prevails: how deeply AI is integrated into government, and how far its design is authored locally. Crossing them yields four futures. Low integration under external authorship preserves inherited stasis, the existing gap with a few pilots attached. Deep integration under external authorship produces algorithmic dependence, the second transposition completed. Local authorship without integration produces sovereign enclosure, which preserves difference at the price of capability. Only deep integration under local authorship reaches authored integration, the condition this paper calls the Synaptic State.

09

From decoding to coding: the Synaptic State response

If the failure is constitutional, the response must be constitutional too. The Synaptic State programme, developed in the author's unpublished manuscripts, starts from a principle the postcolonial toolkit implies but never states as method: decode before you code. No institutional design is drawn until the black box of a state's political power has been opened. Difference is treated as the first input to design, not as an obstacle to be engineered away.

Decoding: six layers

The decoding dossier for each state answers six sets of questions. The first concerns political power and its configuration: who actually holds power, whether formal or informal, civil or military, religious, tribal or economic, how that power is acquired, exercised, maintained, reproduced and renewed, and where the real veto points lie. The second concerns geopolitical placement at three scales: locally, the relations of centre and periphery, federal and provincial tiers and contested territories; regionally, neighbours, rivalries, blocs and corridors; and globally, great-power alignment, aid and debt, and dependence on technology suppliers. The third maps the external influence tracks, meaning the colonial and post-colonial legacies and the donors, lenders, security partners and suppliers that shaped each institution and still hold leverage over it. The fourth reconstructs the history of governance: the path of state formation, regime cycles, reforms attempted and the reasons they failed, and the record of earlier digitisation projects. The fifth records aspirations, meaning what the state declares it wants to become, what its elites and public want, and how these differ. The sixth is a synthesis of how power actually flows, where legitimacy resides, what an AI-enabled office must respect, strengthen or avoid, and which dependencies it must not deepen.

The dossier is analytical and rests on documented evidence. It is not covert collection on individuals, and its sensitive findings remain with the commissioning authority. In the terms of Section 7, its purpose is to measure divergence office by office and to identify the type of informal institution with which each office must reckon.

Coding: the constituted office

From the decoding the design is coded as a whole package. Public purpose charters fix what an office is for, which decisions it may take and which are reserved to human authority. Office designs place AI-enabled offices only where the decoding shows they can hold legitimacy. Interfaces with informal and customary authority, where that authority is lawful and legitimate, allow the office to work with existing power rather than against it, converting competing informal rules into accommodating or complementary ones where that is possible. Supplier safeguards and substitution rights prevent the office from deepening old dependencies or creating new ones. A sequence of pilot domains is matched to the state's own aspirations and political calendar.

The unit of all this is the constituted public office. It has an identity, a mandate, resources, a memory, a record of decisions and reasons, a remedy for those it affects, and rules of succession. Its identity persists when its incumbent changes, whether the incumbent is a person or a model. This answers two problems at once. It supplies the postcolonial toolkit's missing theory of non-human authority, because authority belongs to the office, which society constitutes and can revise, and never to the model. It also answers the problem of dependence, because a model that is a replaceable incumbent allows an office to survive a change of supplier, a revoked licence or a failed system.

Samuel Huntington proposed four criteria of political institutionalisation: adaptability, complexity, autonomy and coherence (Huntington 1968). They provide a yardstick for the constituted office. An office is adaptable if it survives the replacement of its incumbent and the revision of its mandate; complex if its functions are differentiated and recorded; autonomous if its decisions follow its charter rather than the interests of a patron or a supplier; and coherent if those it affects can predict and contest its decisions. Each criterion can be observed, which allows the programme's claims to be checked against the propositions of Section 7.

Proof before power

No office receives authority by assertion. It first runs in shadow operation, producing decisions that are compared with actual decisions but bind no one. It is measured against credible alternatives, including a strengthened human office. Only then does it receive a bounded transfer of defined powers, which can be narrowed or reversed at each review. The approach excludes mass surveillance, social scoring and any electoral use. Seventy states would mean seventy designs; cooperation among states shares methods and evidence, never a single blueprint.

This is the constructive step the postcolonial tradition did not take. It keeps that tradition's insight that difference is constitutive and turns it into a design discipline. It moves from deconstructing the transposed state to authoring the offices of a state that society can recognise as its own, with AI inside them rather than above them. The programme's name for the destination is the Synaptic State: a political order that joins revisable public knowledge to substantive artificial offices under institutions of public authorship.

10

Implications for frontier AI developers

Frontier AI developers have become actors in state formation, whether or not they intend it. When a developer signs a framework with a ministry and its model drafts that ministry's decisions, it helps shape the authority of the state it serves. That is a responsibility. It is also an opportunity to do what no earlier holder of a decisive technology has done: to use the advantage to strengthen the states it reaches rather than to subordinate them. The argument of this paper yields five principles for that role.

The first is to treat difference as a design input: decode the authority structure of an office before deploying into it, and publish the method even when the findings remain confidential. The second is to invest in institutions, not only in data systems. Credits, compute and developer training transfer capability; offices, records and remedies transfer accountability, and only the second makes the first safe to use. The third is to make substitution possible, designing so that an office survives a change of model, supplier or access, and so that the developer's presence strengthens the state's authorship rather than its dependence. The fourth is to close the archive gap by evaluating and improving models in the languages the public actually speaks, and by publishing those evaluations openly for any model to use, including a country's own. The fifth is to hold the red lines: no surveillance, profiling or electoral use, and no automated decisions on rights before an office has earned them in shadow operation.

Among frontier developers, Anthropic is unusually well placed to act on these principles, for reasons of record rather than of rhetoric. In early 2026 it refused to remove restrictions on the use of its models for mass domestic surveillance and fully autonomous weapons, in a dispute that put a contract with the United States Department of Defense at risk (Axios 2026). Its Beneficial Deployments work already signs government frameworks in the developing world. Its Kenyan declaration places data under Kenyan law and leaves each ministry to design its own programme. As a public benefit corporation it has bound itself to purposes beyond revenue.

What its current template lacks is the step this paper argues for. Its frameworks are led by sectors: health, education, skills and developer tools. In an externally authored state, capability on its own flows to whoever holds actual power. The decisive question is whether the office using a model is constituted, recorded and answerable; whether a health ministry can use the model at all is secondary. Political investment in local institutions, rather than in data management systems alone, is the differentiation that fragile states need from advanced-state templates. It is also where a developer committed to safety can do the most good for the largest number of people. These principles serve the builders as well as the governed. Most of the world's future users live in these states; deployments that strengthen legitimate institutions endure, while deployments captured by local power become liabilities of trust and safety. If trusted developers do not help these states author their AI-enabled institutions, less careful suppliers will.

This role does not belong to the developer alone. It requires institutional designers who can decode a state's power and code its offices, working between the developer and the government and answerable to the government. Form & Flux was founded in Islamabad to undertake that intermediate work. Nothing in this paper implies that Anthropic, or any government, has endorsed the approach. It is offered as an argument and as an invitation.

11

Limitations and a research agenda

The argument has limits that should be stated plainly. It is a theory-building paper, and its propositions have not yet been tested. Its evidence is drawn from history, comparative scholarship and public reports of deployments that are, in most cases, too recent to have produced outcomes. Pakistan is the reference case, and although the concept of the externally authored state is defined to travel, the paper has not shown that it travels equally well to Latin America, the post-Soviet space or the Gulf. The concept itself requires measurement: divergence between formal and actual authority is easier to describe than to quantify, and any test of Proposition 1 will depend on indicators, such as the share of decisions taken outside the formal chain or the completeness of office records, whose validity must be established.

Three further limits deserve mention. The first is that the mechanism treats AI systems as enforcers of formal rules. Some systems may instead be designed to learn and reproduce informal practice, which could convert rupture into capture rather than prevent it. The second is that the argument concerns imported design, and the line between importation and local redesign is a continuum rather than a boundary. The third is that the author is both a scholar of the externally authored state and the founder of a firm that proposes to work within it. That position gives access to evidence that outside observers lack, and it also creates an interest that readers are entitled to weigh. The propositions are stated in falsifiable form so that others can test them independently.

The research agenda follows from these limits. A first step is an office-level comparative study that pairs similar AI deployments across offices with differing degrees of divergence within the same state, which would test Propositions 1 to 3 while holding national conditions constant. A second is the systematic recording of shadow-operation data, comparing machine-drafted and actual decisions, which would provide the first direct evidence on whether constituted offices outperform imported ones. A third is a cross-regional extension of the concept to states outside South Asia and Africa. A fourth is the study of supplier dependence, using episodes such as the June 2026 suspension as natural experiments for Proposition 5.

12

Conclusion

AI-enabled government built for advanced states fails in fragile ones because it rests on something those states were never allowed to build: authority that is impersonal, recorded and recognised by the society it governs. The empires and their successors exported the forms of that authority without its substance, and the gap between state and society has outlived empire by generations. Imported AI designs assume the substance, meet only the forms, and are captured, bypassed or ruptured.

The tradition that runs from Gramsci and the philosophers of difference to Said and the Subaltern Studies collective remains the best instrument for seeing this. In the age of AI its concepts grow sharper: the archive becomes the corpus, common sense becomes the model, governmentality becomes algorithmic and mimicry becomes procurement. Yet the tradition was built to unmask authority, not to constitute it, and it has no account of an office that a machine might hold. It survives as diagnosis and fails as programme.

The stakes are high because AI cuts both ways. Every decisive advance in history has carried the temptation to subject others, and AI carries it in its most intimate form because it reaches the reasoning inside institutions. It also carries the opposite possibility, more fully than any technology before it. It could supply, within a generation, the impersonal authority that external state-building withheld, and it could let externally authored states meet advanced states on equal terms. The difference between those futures lies not in the technology but in authorship: who decodes the state's power, who codes its offices and who can revise them.

The argument therefore ends where the Synaptic State begins, with the constituted public office as the unit through which AI should enter government, and with local authorship as the condition of its legitimacy. For frontier developers the lesson is that the most valuable thing they can help build in a fragile state is not a platform but an institution. For fragile states the lesson is older. A state that is to belong to its society must be authored by it, now with new instruments and, for the first time, with a chance of doing so on equal terms. The builders of frontier AI did not create the inheritance. They can decide whether the algorithm repeats it.