The AI-Native State: What It Is and Where the Boundary Lies

The label came before the definition. The earliest official use I found is the theme of the Abu Dhabi Government pavilion at GITEX Global in October 2024: “Towards an AI-Native Government”. In January 2025 the emirate announced an AED 13 billion digital strategy with targets for 2027: 100% of government operations on a sovereign cloud, 100% of processes automated, and more than 200 AI systems. In April 2026 the UAE cabinet set a goal of moving half of government sectors and services onto AI agents within two years. China's State Council, in its “AI+” program (August 2025), set a target for the whole economy and society: penetration of new-generation smart devices and agents above 70% by 2027 and above 90% by 2030, without defining what the share is a share of.010101I once wrote that “digital economy” had turned into near-noise and that the word “digital” would soon be dropped, the way nobody says “electric economy”. The AI-native label repeats the story: it arrived before the definition.Knizhny kub · digitalization as noise01I once wrote that “digital economy” had turned into near-noise and that the word “digital” would soon be dropped, the way nobody says “electric economy”. The AI-native label repeats the story: it arrived before the definition.Knizhny kub · digitalization as noise

These targets describe how much will be spent and how many processes are to be automated. They do not tell us whether people will receive a service faster, face fewer wrongful refusals, or be able to get an error corrected. Even what counts as moving a service to AI is left undefined. The OECD warned in February 2026 that growing agent adoption should not be mistaken for technological maturity.

In the emirate's materials I found, “AI-native government” is described by its attributes: AI “at the core”, digital identity, a sovereign cloud, proactive services. No measurable criteria have been published, so the goal of becoming “the world's first AI-native government” by 2027 cannot be checked against anything. A DGE official places autonomous handling of most routine services on a horizon “beyond 2035” in a personal essay on the platform of the vendor Kyndryl, not in an adopted document. The UAE's service design guidelines (August 2026) require a clear outcome and the citizen's right to object to or correct the result, but that is a design principle, not a measurement.

In my view, the closest anyone has come to describing how the work is organized is the programmatic paper “The Agentic State” (October 2025). In it, autonomous systems absorb complexity and scale, while civil servants concentrate on strategic direction, political accountability and decisions that call for judgment. But this too is a picture of an end state without measurable criteria. The authors offer a useful lens in the question “who decides what, with what authority, and how reversibly”.

My working definition carries the formula of the first longread over to government: processes, accountability and learning change together. An AI-native state turns a lawfully authorized decision into an accepted, contestable and reconstructible result, and learns from its consequences.

I distinguish two levels of change. Government changes how it delivers services and who answers for their outcomes. At the country level, the question is what data, computing capacity, people and rules are available to support that work. Businesses and researchers also use this shared foundation.

In the first longread I distinguish three states: AI assists a person; AI is embedded in a workflow; the organization operates as AI-native. In government, these mean assisting an official, redesigning a specific service, and changing how the whole agency works, including how people can contest its decisions and get errors corrected. These states describe individual services and agencies. Within one country, AI may merely assist staff in one service while already changing how another works. Assigning a single state to the whole country would therefore be misleading.

A ministry is a poor unit of change: departmental boundaries cut the citizen's end-to-end journey into pieces. Company logic has been carried over to the state before: New Public Management reforms broke up agencies and brought in competition and incentives. In 2006 Dunleavy and co-authors argued that in several leading countries these reforms had stalled or gone into reverse, and that the digital era pushes the other way, towards reassembling functions around citizens' needs. That is why the company framework is worth carrying over as a method of testing, not as the slogan “run government like a business”.

The Agentic State team, with the World Economic Forum and Capgemini, proposes starting from 70 functions of government rather than from ministries. In September 2026 California presented AskCA, a prototype organized around life events such as looking for a job; the state has published no usage data.020202I wrote up Eugene Sergueev's talk on Flo Health: teams built around channels sped up growth but fragmented the experience, so the next step was to cut teams around the user's life scenarios with an inverse Conway maneuver. It also brings Organizational Decision Records, so a year later someone still remembers why the teams were cut this way.Knizhny kub · the inverse Conway maneuver02I wrote up Eugene Sergueev's talk on Flo Health: teams built around channels sped up growth but fragmented the experience, so the next step was to cut teams around the user's life scenarios with an inverse Conway maneuver. It also brings Organizational Decision Records, so a year later someone still remembers why the teams were cut this way.Knizhny kub · the inverse Conway maneuver

In government the owner is defined not by a manager but by law: it is an official with competence, that is, with the range of matters assigned to the body. Competence is often split between agencies, and the person who holds it changes with the political cycle. When ownership is split, the transition breaks in three places:

• —Identity checks and the delivery of assistance have different responsible parties. In Jharkhand in 2016–2018, ration distribution was tied to mandatory Aadhaar biometric authentication. According to an assessment by Muralidharan and co-authors, authentication itself did not cut leakage, and critics regard the drop that followed the “reconciliation” of distribution as temporary. Meanwhile 1.5–2 million eligible recipients out of about 15 million were at some point left without rations; the authors attribute the harm primarily to how the state ran the transition. This case did not involve AI. For my framework, the distinction is between the operator of the identity-checking system and the agency responsible for providing rations: connecting biometrics does not itself ensure that a person receives the assistance they are entitled to.
• —The product is not accepted by its operational owner. In April 2026 Germany's Federal Court of Audit showed through examples that ministries' data labs brought some “data products” to readiness but did not hand them over. One agency refused to take on a dashboard even when its ministry ordered it to, citing a shortage of staff and expertise.
• —The model and the service sit at different levels of government. China's 2025 guidelines on large models in the public sector instruct counties “in principle” to use models from the level above rather than build their own, so that “model islands” do not emerge. By design, the owner of a county service does not control the model running inside it; there are no independent data on implementation.

In Table 1 each element of the first longread gets a verdict: carries over, changes meaning, or breaks.

Three things carry over intact: autonomy per action, work design, and the “AI assists a person” state. One thing breaks: the “organization operates as AI-native” state applied to government as a whole. Everything else changes meaning, and the main shift is the second consumer of the result: the person the decision is about.

A company learns that it has failed from customer churn. The state learns from a complaint, a court ruling or an auditor's report. Albert Hirschman distinguished two responses to decline in firms, organizations and states: exit to another supplier, and voice, the demand for change from within. Exit from public goods, he wrote, is difficult. I put it more bluntly: from most government services there is no exit at all. Even the authors of “The Agentic State” call the duty to serve everyone a structural difference between the state and a firm. So the feedback loop of government, the “learning” in the definition, runs through voice: notice, appeal, the ombudsman, the courts, audit and elections.

The loop breaks at different links, and the break lasts for years.

• —Robodebt, Australia: they knew and did not stop. In November 2014 the department received legal advice that calculating benefit debts by income averaging was not consistent with the law. The Royal Commission found that by early 2017 the unfairness and probable unlawfulness of the scheme had become obvious, yet most of the reviews took place after that point. Raising debts on averaged income alone stopped only in November 2019.
• —Försäkringskassan, Sweden: seven years from warning to withdrawal. In 2018 the inspectorate found that the algorithm that selects claims for checks under the benefit for parents caring for a sick child did not ensure equal treatment. In November 2024 journalists showed that it more often selected women, migrants and their children, people on low incomes and people without higher education. The model was taken out of service around August 2025, two months after the regulator opened its inquiry; the agency gave no reasons. The regulator closed the case without a finding on legality, and risk-based selection continues with other profiles.
• —BOSCO, Spain: a court ruling left unenforced. In May 2019 the Civio foundation showed that the program denies the social electricity discount to pensioners who are entitled to it. In September 2025 the Supreme Court ordered the government to release the source code. As of May 2026 the code had not been handed over, and no handover had been reported by 17 October.
• —Poland: profiling without a law. From 2014 a system sorted the unemployed into three support profiles. In 2018 the Constitutional Tribunal ruled that the method of profiling had to be set by statute, and from 14 June 2019 profiling was abolished.

Fast reversals happen too. In 2020 the model used by the exams regulator Ofqual lowered 39.1% of the A-level grades assessed by schools. Results came out on 13 August, and on the 17th, after the U-turn in Scotland, protests and the first legal challenges, Ofqual and the government gave students the higher of the two grades. Warnings about the model had been voiced weeks before the results. The reversal could be fast because the harm was mass, simultaneous and visible, and a fallback, the school-assessed grades, already existed.

This gives my first rule: autonomy must not outpace the capacity of voice. If the system makes decisions faster than government can hear and work through complaints, the error is replicated before anyone learns of it.030303I read Ron Westrum's original paper, the one Accelerate cites: it lists six ways an organization reacts to an unwelcome signal, from shooting the messenger and isolating them to a local fix, a global fix and an inquiry into root causes. Westrum himself warns that the link between culture type and performance is correlation, not causation.Knizhny kub · Westrum's typology03I read Ron Westrum's original paper, the one Accelerate cites: it lists six ways an organization reacts to an unwelcome signal, from shooting the messenger and isolating them to a local fix, a global fix and an inquiry into root causes. Westrum himself warns that the link between culture type and performance is correlation, not causation.Knizhny kub · Westrum's typology

Voice can be raised and still go unheard. In Amsterdam in March 2022, a council of 15 benefit recipients and their advocates demanded that the experiment with Smart Check, a model that selects applications for checks, be stopped as disproportionate: by the council's estimate, only about 3% of applications turn out to be fraudulent. The city council was merely informed that the pilot was starting and was never asked to vote on it. In November 2023 the pilot was halted on the grounds that the absence of unintended discrimination could not be guaranteed.

As in the first longread, the evidence belongs to different levels. Self-reports and the “percentage through AI” are one level; comparison with a matched group, a control sample and an audit are another. In the pairs collected here, the measured result turns out narrower and more modest than the claimed one.040404I wrote up McKinsey's State of AI 2026: 80% of people who use AI themselves say it made them more productive, yet only 37% see any EBIT contribution at all. These are self-assessments, not a financial audit, which is why I suggested measuring the chain “individual effect → workflow → function economics → EBIT”.Knizhny kub · McKinsey State of AI 202604I wrote up McKinsey's State of AI 2026: 80% of people who use AI themselves say it made them more productive, yet only 37% see any EBIT contribution at all. These are self-assessments, not a financial audit, which is why I suggested measuring the chain “individual effect → workflow → function economics → EBIT”.Knizhny kub · McKinsey State of AI 2026

• —TAMM, Abu Dhabi. DGE does not define its “95%”, and in June 2025 the department itself tied that share to requests to TAMM's AI assistant. For the first half of 2026 it reports separately that 97.11% of conversations were handled by AI and that 187,400 of 55.5 million transactions involved AI. So the “95%” describes the chat, not the share of transactions the system completed.
• —Estonia. The ministry's list of AI systems, kratid.ee, has more than 130 entries from 53 organizations, but it has not been updated for two to three years and, in experts' view, includes retired systems and systems that are not AI. Working from questionnaires returned by 48 agencies, Estonia's National Audit Office counted fewer than 30 AI systems in active use; most, in its assessment, are still being tested and do not yet produce noticeable savings. These are two counts on different bases, not a funnel and not a measure of results.
• —Microsoft 365 Copilot, United Kingdom. In a cross-government experiment, participants reported saving 26 minutes a day. The Department for Work and Pensions (DWP) compared license holders' answers with a matched group and got 19 minutes, and 10–12 without adjusting for interest in AI and experience with it. This is still self-assessment, licenses were not handed out at random, and the authors themselves allow that the figure may be inflated. The Department for Business and Trade (DBT) found no robust evidence that time saved turns into higher productivity, but it had not set out to test that. These three figures do not belong in one row: the samples, tasks and measures all differ.

Agencies also measure for themselves. In small pilots in 2024, the UK Home Office excluded 9% of interview summaries as inaccurate or incomplete, and the share of well-founded decisions in an internal sample check was similar in the test and control groups. DWP publishes an annual assessment of its model for selecting Universal Credit advance claims for checks: the model is 2.5 times as accurate as random sampling.

Measured failures are more visible where courts, auditors, journalists and civil society organizations are at work; the cases collected say nothing about how often failures occur. In the EUROSAI parallel audit of AI by twelve supreme audit institutions, 56% of responding governments (five of nine, judging by the shares) reported having no mechanism to evaluate the impact and effectiveness of AI projects. The absence of a check is a property of the system, neither a success nor a proven failure. Where checks do exist, the effect can be modest: France's Court of Audit estimated France Travail's gains from AI over 2017–2025 at about €120 million under favorable assumptions, slightly more than costs of about €108 million.

Before allowing a system to act on its own, I propose checking four conditions: verifiability of the result, reversibility of the action, consequences of an error, and authority under law. Assess them for a specific action: drafting a reply, awarding a benefit and refusing it each call for a separate decision about autonomy.

The first three conditions come from the company framework. In my view, they also help explain the limits on decisions about rights. White House memorandum M-25-21 treats AI as “high-impact” when its output is the principal basis for a decision with legal, binding or otherwise significant effects on rights or safety, regardless of human oversight. Annex III of the EU's AI Act classes as high-risk the assessment of eligibility for benefits, the examination of asylum and visa applications, and assistance to courts, but the obligations for such systems do not yet apply.

In government a fourth condition joins the three: authority under law, meaning a legal norm that permits the system this action and an official who answers for it. Authority already appears in the lens of “The Agentic State”; I only tie it to law. A label does not create authority. In 2025 Albania's prime minister presented the chatbot Diella as “minister for AI” in public procurement, yet the procurement AI system was still not ready by August 2026. In October the Constitutional Court dismissed a request from MPs without examining the merits, but its summary for the press stated that a technological tool cannot create or replace the powers of state bodies.050505I wrote up Stephen Brozovich's AWS talk: Singapore's agentic AI governance framework and AgentCore converge on the same questions — who is this agent, who authorized it, what is it allowed to do, and can it be audited. What I liked was governance as running infrastructure: not a policy document but a check of access and action on every request, outside the model's loop.Knizhny kub · AWS and the Singapore framework05I wrote up Stephen Brozovich's AWS talk: Singapore's agentic AI governance framework and AgentCore converge on the same questions — who is this agent, who authorized it, what is it allowed to do, and can it be audited. What I liked was governance as running infrastructure: not a policy document but a check of access and action on every request, outside the model's loop.Knizhny kub · AWS and the Singapore framework

Reversibility in government is asymmetric: an automatic “yes”, by my assessment, is cheaper to correct than an automatic “no”. Below are designs in which the system can speed up an approval but not refuse, provide a lead but not grounds for an arrest, interpret an intent but not approve a payment. That they are harmless has not been proven.

• —The system used by IRCC, Canada's immigration department, can find a routine temporary visa application eligible but, in the department's words, “never refuses or recommends refusing”. Routine files, by its figures, move 87% faster, while the rest wait longer; there is no independent check of outcomes.
• —In Detroit, the rule in force since 2019 that face recognition is “only a lead” did not prevent the third known wrongful arrest. Independent evidence and court oversight were required only by the 2024 settlement, after the harm had been done.
• —On 10 September 2026 the non-executive chairman of India's NPCI, the operator of the UPI payment system, set out a principle: AI may recommend, but authentication and final settlement must follow deterministic, auditable rules. As the press paraphrased it, the agent interprets the intent but does not approve the payment. UPI has no separate rules for agents.

At the other pole sits the automatic “no”. In Michigan's MiDAS system, by the agency's figures, 85% of the 2013–2015 benefit fraud determinations made without a human turned out to be wrong, and 44% of those made with one. The groups are not comparable, but even with a human involved, nearly half the accusations were false.

A “human in the loop” is no guarantee in itself: what matters is how that involvement is designed.

• —At DWP, the officer deciding on a flagged advance claim does not see the risk score and does not know that the model referred the case. This is the department's description; it publishes no data on how reviewers behave.
• —In IPIC, a UK tool for immigration enforcement, an officer has to justify only disagreement with a recommendation. In Privacy International's view, this encourages rubber-stamping.
• —In Spain, according to a 2014 study, police left the risk level calculated by VioGén unchanged in 95% of cases (they may raise it but not lower it). No more recent data have been published, although VioGén 2 has been in use since 2025.

In the longread on the boundaries of agent autonomy, the reliability threshold at which an agent can be trusted with an action rises with the loss from an error. When the error falls on a person, that loss includes redress years later and lost trust, so the threshold for a refusal, a debt or a sanction is higher. This is my interpretation, not a measurement.

This gives my second rule: start automation with actions in the person's favor and with reversible actions. The UK medical device regulator drew a similar line in a clarification of existing rules (July 2026): transcription, summaries and drafts for a clinician to check are not a medical device, while issuing prescriptions without a clinician's check and supporting diagnosis or treatment are.

In a company, memory is a production asset. In government it is also a legal record, without which a decision cannot be appealed or checked. The authors of “The Agentic State” write that more due process is expected of the state than of the private sector, and that agentic systems break the familiar accountability of civil servants. Decisions number in the thousands while oversight runs in weekly cycles; it is hard to establish whose action led to the outcome; responsibility is spread across agencies. Reconstructing a decision from documents and interviews with officials, the authors warn, “may become technically impossible”.

This record has three parts.

• —An inventory of systems: what is running at all. The Board of Audit of Japan (a September 2025 report with data as of March 2024) found that 137 of the 356 important information systems it examined had no Digital Agency identifier. At least one kind of audit log is collected for only 32% of them, against 92% of systems that have an identifier.
• —A decision record: the data, the model version, the rule and the officer's action. The AI Act requires automatically generated logs of a high-risk system to be kept for at least six months, but for Annex III systems only from December 2027. Brazil's National Council of Justice already requires courts to log their use of generative AI automatically for monitoring and audit.
• —The law as a version of the rule in code. In Australia, from April 2022 the law required a person's circumstances to be taken into account before a payment was cancelled, yet the TCF system kept cancelling payments automatically for about two more years. That is how 964 people lost their payments, and the Ombudsman found these cancellations unlawful. The error went unnoticed for 17 months and was found by outside lawyers working on something else, and the department, in the Ombudsman's assessment, kept the old procedure for another 10 months.

Part of the record can disappear. According to an Amnesty International report, Denmark's benefits agency UDK replied that model evaluation metrics were “not documented” during development. Information on high-risk cases is continually overwritten, and the agency holds no demographic data, which, as Amnesty itself notes, also protects privacy. UDK declined a joint audit but supplied aggregate control figures for 2023. For an outside researcher, checking for bias after the fact is hard; for a body with a right of access, it is not impossible. The agency disputes the report and admits no wrongdoing.

The law also limits memory itself. India's Supreme Court in Puttaswamy limited UIDAI's retention of central Aadhaar authentication records to six months. Banks and other requesting entities keep their own logs for two years and archive them for five more: the center forgets, the bank remembers. My conclusion: there is no need to remember everything, but knowledge of the systems involved in coercive decisions must reach at least the auditor and the court, and the impact assessment must be published before launch, as Canada's directive on automated decision-making requires.

In the cases examined, government metrics count access. According to the US General Services Administration (GSA), the OneGov agreements “expanded access to AI tools” for about 3.5 million federal employees, and GSA publishes no data on active use.

For the second consumer of the result (section 2), what gets counted is decisions that stand on appeal, errors by group, reversals, time to correction and the people left without the service. Decisions that stand can be counted only if there is an appeal. In the US Department of Homeland Security's inventory of AI use cases (June 2026), components themselves marked appeal as “not applicable” for 18 of 35 deployed high-impact systems, and for two more it is ruled out by law or operational constraints. These include not only radiation monitors that handle no personal data, but also face recognition and passenger risk assessment.060606I wrote up Noise by Kahneman, Sibony and Sunstein: you can measure the scatter of decisions without knowing the right answer — give the same cases to several independent experts and compute the spread. And a caveat from the same book: an algorithm removes noise but does not guarantee fairness; with bad data it amplifies bias consistently.Knizhny kub · Kahneman's Noise06I wrote up Noise by Kahneman, Sibony and Sunstein: you can measure the scatter of decisions without knowing the right answer — give the same cases to several independent experts and compute the spread. And a caveat from the same book: an algorithm removes noise but does not guarantee fairness; with bad data it amplifies bias consistently.Knizhny kub · Kahneman's Noise

• —Amsterdam. According to an analysis by Lighthouse Reports, the first version of the Smart Check model wrongly flagged applicants with non-Western nationalities almost twice as often. After the data were reweighted, the skew in the pilot shifted against Dutch nationals and women.
• —DUO, the Netherlands. In March 2024 the government apologized for indirect discrimination in checks on student grants. According to follow-up research by the Algorithm Audit foundation, 86% of the students who filed an objection had non-European migrant roots.
• —DWP, United Kingdom. The advances model from section 4 refers foreign nationals for checks 3.35 times as often for the same rate of confirmed cases. Of nine protected characteristics, the department was able to test one for 2024/25 and two for 2025/26.

Since the state has to serve everyone, the tail has to be counted too. A Singpass account, according to the operator, is held by 97% of Singapore citizens and permanent residents aged 15 and over. The operator does not say how many holders use it without help, so the tail may be larger than 3%.

According to the National Taxpayer Advocate, in the 2026 filing season (through 18 April) the US Internal Revenue Service (IRS) had about 27% fewer staff than a year earlier, and operators answered 20% fewer calls on almost the same volume. This was partly the IRS's own choice: staff were moved from the phones to working through correspondence, which, in the Advocate's assessment, helped many taxpayers. Almost 99% of returns were processed: what sagged was phone assistance, not the processing pipeline and not assistance as a whole.

Metric definitions matter too. The head of the US Social Security Administration (SSA) put the average answer time for May 2026 at under five minutes. But that metric counts the wait of callers who chose a callback as zero, and the callback itself, according to the inspector general, took on average about 1 h 49 min to arrive in fiscal 2025, and about an hour by September. The inspector general found the published figures accurate by their own definition, but in summer 2025 the callback wait was removed from the public data. On its Accounts Management lines, the IRS lowered its level-of-service target from 85% to 70% and changed the measure, yet the fall in the actual level from 87% to 73% stayed in the reports. My practical conclusion: fix definitions in advance and do not change them along with the process; if you do change them, publish both series.

In my view, coverage targets reward the breadth of adoption rather than the appropriateness of delegation, and the second rule is turned on its head. The best-documented case is the work of the US Department of Government Efficiency (DOGE); it is a particular case, not a general fact about “AI cuts”.070707In December 2024 I wrote up the plan for DOGE: the targets came as numbers from the start — cut the federal workforce by 75%, shrink 400-plus agencies to fewer than 100, an estimated $2 trillion effect. Even then it was clear the commission had no direct executive powers.Knizhny kub · DOGE07In December 2024 I wrote up the plan for DOGE: the targets came as numbers from the start — cut the federal workforce by 75%, shrink 400-plus agencies to fewer than 100, an estimated $2 trillion effect. Even then it was clear the commission had no direct executive powers.Knizhny kub · DOGE

• —VA. According to ProPublica, a DOGE engineer wrote a tool on his second day on the job that flagged more than 2,000 Department of Veterans Affairs contracts as “munchable”, that is, fit to be cut. By June 2025 at least two dozen of the flagged contracts had been cancelled, but according to the VA, staff made the decisions, and some of the cancellations began before the tool existed.
• —NEH. A DOGE staffer ran short descriptions of National Endowment for the Humanities grants through ChatGPT, asking it to answer in under 120 characters whether the project related to DEI, that is, diversity, equity and inclusion. The answers became the justifications in termination lists; in April 2025 more than 1,400 grants were terminated. On 7 May 2026 a federal court held the terminations unconstitutional: the decisions were made by DOGE, which had no legal authority to make them; the government has appealed. Some of the correspondence, held in a messaging app with auto-delete, was lost, but the emails and spreadsheets were enough for the court.
• —Measuring the whole program, not just the AI. The US Government Accountability Office (GAO) could neither confirm nor refute the $110 billion in savings DOGE claimed.

The least verifiable model answers ended up where actions are irreversible: at NEH, according to the court, DOGE adopted them wholesale, while at the VA the flags merely overlapped with the cancellations. At the same time, oversight was shrinking. By the Partnership for Public Service's count, staffing in departmental inspector general offices was on average 19% lower at the end of May 2026 than in December 2024, and the president fired 11 of 13 of those inspectors general. The federal CIO, in an interview on 18 August 2025, called AI the main way to ease the staff shortage, but described no plan to measure the results.

In other countries a breach of the second rule has not been proven. But reports again count deployments while leaving the effects on service recipients unclear. I am not assessing political systems, only whether each of them has a way to correct a mistake. In China, by a count of the Zhiding think tank as relayed by Jamestown's China Brief, 72 provincial and city governments had “fully deployed” DeepSeek within about seven weeks of R1's release. That is the think tank's own label, not a verified status. The 2025 guidelines from section 2 told agencies to prevent “digital formalism”, including compulsory use that serves no purpose. In Korea, AI textbooks were planned as mandatory, but in January 2025 they were turned into a voluntary trial; they launched in March, and in August parliament stripped them of textbook status. Auditors measured only usage: on average, 60% of students in participating schools never logged into the system.

Governments are not the only ones that build IT badly. In the sample studied by Flyvbjerg and co-authors, government IT projects, mostly US federal ones, cost on average 2.1 times as much as planned, and private ones 1.4 times. But the medians are almost equal (1.0 and 0.9), and the shape of the heavy tails does not differ statistically between the two datasets. The full cost of an error in decisions about people arrives years later. According to a 2021 court judgment, under Robodebt the government unlawfully raised debts of at least A$1.763 billion against about 433,000 people. The Dutch redress program for the childcare benefits affair has set aside €11.8 billion over several years, for the consequences of the whole scandal, not just the algorithm.

So far the subject has been the machinery of government. Its six checks from the first longread depend on country conditions that no single agency controls. An agency cannot create value if people have no channel of access. It cannot establish accountability without independent courts and auditors. It cannot keep portability if its chips depend on another state's permission.

At this level I apply my Stack Ownership Boundary framework: what a country keeps as its own is not models and capacity but the right to act. The framework has two axes: rate of change and uniqueness. Rent what changes fast and is not very unique: frontier models, commercial capacity, communication channels. Adapt how outside technologies are used in your services: a gateway to several models, language fine-tuning of someone else's foundation model, subsidized access to someone else's capacity. Own what is stable and unique: whatever determines who may act and on what grounds. In the framework these are the “contracts that define the right to act”: identity and mandates (records of agents' authority), registries and data rights, acceptance criteria and evaluation sets, in-house reviewers, the appeal procedure and outcome statistics. Sovereignty in this framework is the cost of replacing a model, a cloud or a vendor. The evidence is a replacement carried out or rehearsed, not a declaration. As with the six checks, scores for the layers do not add up to a “sovereignty score”: owning model weights does not make up for the absence of appeal.080808I looked at why Cambricon and Ant Group joined the PyTorch Foundation: my reading is that they are buying not control of the framework but a seat at the table where the cost of leaving CUDA gets lowered. I suggested watching boring engineering signals, not logos — shared CI, current PyTorch without special builds.Knizhny kub · PyTorch as a portability layer08I looked at why Cambricon and Ant Group joined the PyTorch Foundation: my reading is that they are buying not control of the framework but a seat at the table where the cost of leaving CUDA gets lowered. I suggested watching boring engineering signals, not logos — shared CI, current PyTorch without special builds.Knizhny kub · PyTorch as a portability layer

A “sovereign model” has two incompatible meanings, and both are a legitimate political choice.

• —Own weights from scratch. In January 2026 Korea's science ministry dropped the Naver Cloud consortium from its sovereign model program, citing the requirement to train models in-house: fine-tuned foreign models do not count as sovereign. The Dutch GPT-NL is being trained from scratch on data with a lawful supply chain, and the data owners receive a share of the revenue.
• —A replaceable foundation model plus your own data, evaluation and hosting. Taiwan's TAIDE moved from Llama 2 to Llama 3 and then to Gemma 3; fine-tuning on Llama 3 after its release took four days, according to the program. This shows that one's own data and fine-tuning pipeline carry over between foundation models, but it does not prove that a running government service has switched models without loss. The dependence, meanwhile, passes from one American licensor to another.

According to a selective review by Stanford HAI, the flagship “sovereign” offerings of the five largest American companies give more control over infrastructure and data, but often restructure dependence on the vendor rather than remove it.

Gateways to several models are a precondition for replacement, not proof of it. The Albert API of DINUM, France's digital agency, serves several open and proprietary models through a single access point; according to the agency, more than 70 government projects use it.

Exit rules are written into contracts and law. White House memorandum M-25-22 requires AI contracts to include data and model portability where applicable, and, where a contract is not renewed, directs agencies to agree a data transfer plan where possible. The EU Data Act will, from 12 January 2027, bar cloud providers from charging customers, public bodies included, for switching to another provider.

Shared systems for identity, payments and data exchange are the most mature layer, and they above all should stay in the country's own hands. Connecting agents to them is still only in preparation, as add-ons on top of existing authorization rules, without creating separate payment systems. According to anonymous sources cited by Reuters, NPCI is building a registry to verify AI agents that pay through UPI; liability for their errors has not yet been defined. The Monetary Authority of Singapore and the industry have proposed a voluntary framework for pilots: a checkpoint that verifies and records an agent's action before it is executed. In the longread on the right to pay I put it the same way: a mandate records the authority, and the party is still a person.

Almost all the figures for these systems are published by the operators themselves. They describe reach and usage, rather than service outcomes. Aadhaar has issued about 1.45 billion numbers, a cumulative counter rather than a count of living holders. In December 2025, 148 million people made or received at least one transaction through Brazil's Pix, about 86% of the population over 14. A shared payment system can widen reach: in 2023, 72% of adults in the CadÚnico social registry paid with Pix at least once.

But independent evidence mostly describes people who lost access to a service because of a system's requirements or failures, as in Jharkhand. Courts and auditors set limits years after mass rollout. The audit of Aadhaar for 2014–2019 was tabled in parliament in 2022: by November 2019 almost 475,000 duplicate numbers had been cancelled. A court suspended the rollout of Kenya's Maisha Namba digital ID in December 2023 because no data protection impact assessment had been done; the suspension was lifted in February 2024. As of August 2024 the assessment had been prepared in response to a lawsuit but not published; according to the government, the data protection commissioner confirmed that it met the requirements.

Gaps between the announced and the built exist in the layer a country owns too. Every EU member state must provide an EUDI digital identity wallet by about 24 December 2026. According to independent trackers, as of 14 October no wallet had been certified, OpenID Europe counts only Denmark's as fully working, and 24 of the 27 states, by Euronews's assessment, will miss the deadline.

Data become AI infrastructure only once the right to use them has been settled, and here countries have gone separate ways. Japan, through a broad copyright exception, allows models to be trained on works. Singapore prohibits contracts from overriding the exception for computational data analysis. The EU gives rights holders a machine-readable opt-out. The UK acknowledged in March 2026 that an exception with an opt-out is “no longer the government's preferred way forward”: only 3% of responses to its consultation supported it.

The main constraints on compute are external. In Ireland, data centers consumed 23% of the country's metered electricity in 2025, and the regulator has required new sites to build their own generation or storage for their full requested capacity. By the IEA's 2025 estimate, grid constraints could delay about 20% of planned data centers. Advanced chips come with another state's permission: the US Commerce Department allowed the UAE's G42 and Saudi Arabia's Humain to buy up to 35,000 GB300 chips each under strict conditions, and G42's access rests on a rule that expires in April 2027.

Here too the announced and the built diverge. Stargate UAE, at 1 GW, was announced in May 2025. In Epoch AI's independent directory the site is under construction: a November 2025 image shows 0 MW of IT power, and Epoch expects the first 200 MW by December 2026. Saudi Arabia's Humain targets 1.9 GW by 2030, while the country's entire data-center capacity in Q1 2026 was about 467 MW; it announced the launch of its first cluster in August 2026 without giving a figure in megawatts. The Edinburgh supercomputer was cancelled in 2024 and brought back in 2025. Some capacity has actually been built: Isambard-AI in Bristol has been running since July 2025.090909I wrote up MIT Technology Review's piece on China's data-center glut: up to 80% of new AI capacity stood idle, though the state had put $6.1 billion into clusters and private companies another $28 billion. Sites went to remote regions for cheap power and land, while demand moved from training to inference.Knizhny kub · China's idle data centers09I wrote up MIT Technology Review's piece on China's data-center glut: up to 80% of new AI capacity stood idle, though the state had put $6.1 billion into clusters and private companies another $28 billion. Sites went to remote regions for cheap power and land, while demand moved from training to inference.Knizhny kub · China's idle data centers

People cannot be rented, and this layer moves slowest. According to the UK's National Audit Office (NAO), in autumn 2023, 70% of the 87 public bodies surveyed named difficulty recruiting and retaining AI specialists as a barrier to adoption, and about a third of the civil service's digital specialists were contractors, agency or temporary workers. In the US, by an independent analyst's calculation from data of the personnel agency OPM, 181 IT specialists were hired in September 2025 for 7,456 who left; that was a month of mass departures from the service, not a trend.

School is where future reviewers grow up, and here the autonomy rule flips: the student's effort is the product. In a randomized experiment at a single Turkish school, open access to GPT-4 raised scores on practice problems by 48% but lowered results on an unassisted exam by 17%. A tutor version with guardrails did no harm, but brought no gain on the exam either. In a World Bank trial in Nigeria, an AI tutor built on the same GPT-4 (Microsoft Copilot) and guided by a teacher raised students' English results by 0.23 standard deviations. This is a working paper: six weeks of after-school sessions, with participants drawn by lot from volunteers. The question shifts from “what to delegate” to “which step must remain human”.

Law with appeal cannot be rented either, and most of the binding rules are still ahead. The AI Act's requirements for high-risk systems will apply only from 2 December 2027. Of the 21 signatories of the Council of Europe Framework Convention on AI, only the European Union has ratified it, and the convention is not in force. Korea's law requires only that one “endeavor” to carry out an impact assessment, and Japan's AI promotion act contains no sanctions.

AI diffusion in the economy is measured by different gauges, from vendor telemetry to consultants' surveys and official statistics, and they do not fit on one scale. According to Microsoft's telemetry, in June 2026, 73.3% of UAE residents aged 15–64 used generative AI, but these are a vendor's data. Humlum and Vestergaard (an NBER working paper), using Danish administrative registers, found a zero effect of chatbots on earnings and hours: effects larger than 2% two years after ChatGPT's launch are ruled out. An instrumental-variable estimate for European firms (a BIS working paper, January 2026) puts the short-run gain in the level of labor productivity at about 4%; the gain is larger where firms invest in software, data and training at the same time. Readiness indices mostly measure inputs. By my own calculation on 2023 data for 164 countries, about 78% of the variation in the IMF's readiness index statistically coincides with variation in income per head; that is an association, not a cause. The state does not own the diffusion layer, but it must own the outcome statistics.