Yujiro Masaki is a researcher at the External Affairs Institute and an entrepreneur specializing in artificial intelligence, digital policy, and international cooperation. He studied for an M.S. in European Studies at UCLouvain and previously served at Japan’s Ministry of Foreign Affairs (Middle East Bureau). He is the Founder and CEO of Jexen. Views are the author’s.
1. Introduction
Europe is writing the world’s most elaborate rules for artificial intelligence while many of the administrations that will enforce those rules still treat the technology as a clearance problem. That is not a small inconsistency. It is the condition under which the rules will be written and applied.
A modern administration is already a large information-processing industry. Officials read statutes, hunt precedents, draft notes, align wording with other services, answer parliaments, and handle inquiries. Diplomacy looks like meetings and speeches. Most of the work is the selection and writing of information. Generative AI fits this work in a narrow sense: it is good at search, summary, comparison, translation, and first drafts. The bottleneck is not a shortage of text. It is institutions that treat every tool as a clearance risk rather than as official memory.
A government that never uses these systems will still regulate them — and it will regulate a cartoon. It will overweight harms that photograph well and underweight procurement, data architecture, and the design of human review. It will write rules that incumbents absorb and that startups cannot. The result is not safety. It is a quieter form of industrial selection.
This paper is not an argument for shrinking the civil service, nor for handing public power to a model. The point is to raise each official’s leverage, keep judgement with humans, and put the state in the market as a competent user. Europe will not design a better AI market than the governments that police it can operate.
The observations that follow are institutional, not personal case files. I spent seven years in a foreign ministry and now work on the other side of the same document culture. I will not discuss classified material or named colleagues. The grammar will be familiar in Brussels: document-ism, clearance chains, inter-service coordination, and a speed unmatched to the seriousness of the file. The audience is the Commission, the Council secretariat, and the digital and interior ministries of the Member States.
2. Administration as information processing
Public administration is described as authority, law, or service delivery. Those descriptions are true and incomplete. Day to day, it is a factory of unstructured text. Laws sit on implementing acts, circulars, guidance, and parliamentary language. A policy that looks like one decision is a stack of notes rewritten until every responsible service can live with the sentence.
This is why the usual division of labour — firms innovate, governments write rules — does not fit generative AI. The actor with the largest document load, the heaviest duty to cite precedent, and the sharpest obligation to explain itself is the state.
The affinity is specific. Large language models do not require clean tables. They operate on messy prose and arguments scattered across files. That is closer to a ministry’s archive than to a warehouse inventory. Officials already spend much of their time doing what models do cheaply: reading, searching, comparing, summarising, spotting contradictions, and aligning yesterday’s wording with today’s. Judgement is not that work. Judgement is what remains after that work is done.
The fixed costs are large: finding the last note, matching legal bases, rewriting a paragraph so another directorate will not object, repeating the same explanation to a citizen, a firm, or a parliamentarian. A reply written when an official is exhausted and a reply written on organised material may carry the same signature. They are not the same act. The cut should be written into job descriptions from the first day: the machine proposes; the human checks the source and keeps the signature.
Europe already has the raw material. EUR-Lex, the Official Journal, parliamentary records, and impact assessments are, in principle, a retrieval environment of unusual quality. In practice much of that material is still a deposit of files. Buying a powerful model and pointing it at cardboard boxes is not modernisation. It is an expensive photograph of the archive.
3. Why capable bureaucracies still stall
From the outside, bureaucracies look slow because the people inside them dislike change. From the inside the picture is different. Many officials are careful, literate, and hard-working. They understand the constraints with precision. Their locally rational choice is often not to move. Individual competence and organisational stall can coexist. That is the fact any AI programme has to face.
Document-ism is the first constraint, and it is not a vice. An oral understanding is not an official decision until it becomes a text that can be retrieved, audited, and defended. That is how a democracy keeps continuity and makes corruption harder. The side effect is obvious. The moment a sentence is written it becomes a trace of responsibility. A good document states what is true. A defensive document states what will injure nobody. The second kind is easier to produce, because it spreads risk. Generative AI will produce either. If the incentive is self-protection, the machine will industrialise self-protection.
Clearance is the second. A file travels through the services that have a stake in it. The final signature may belong to one person; the intermediate vetoes belong to many. The design prevents personal rule. It also means that nobody can decide alone. One objection returns the file. Refusal is cheap; initiative is expensive. For a routine consultation that is a safety device. For a change in method it is a brake. European readers do not need a Japanese word for this. Interservice consultation, the silence procedure, and the journey of a working-party paper through the Council are the same geometry.
Coordination is the third. Few files live inside one ministry. Vertical organisation is not malice. It is the legal allocation of mandate, budget, staff, and parliamentary accountability. Coordination is therefore less persuasion than translation. The same facts must be rewritten in the vocabulary of security, of money, and of statute. Each translation re-enters another clearance chain. A gifted coordinator shortens the loop. The number of loops does not fall.
Personnel systems add a fourth drag. Promotion still rewards the clean file more reliably than the useful experiment. Rotation, which exists for good reasons, breaks the tacit map of why a sentence was written. Official memory then lives in a few heads, a handover note, and an old shared drive. The OECD’s Digital Government Outlook 2026 finds the same gap: almost all OECD countries now use AI somewhere in government; only a minority measure whether the tools changed the work.
None of this is an argument that ministries should move at the speed of a five-person firm. Public power binds rights, spends other people’s money, and makes promises to other states. The defect is different. Procedure does not scale with seriousness. A sensitive security judgement and a cosmetic change to a form can attract the same density of clearance. Officials see the mismatch, and they know that skipping a step is called a solo run if it works and a procedural breach if it fails. A rational official does not skip. The difference with firms is not talent. It is the allocation of failure. The answer is to be fast when the file is light and slow when the file is heavy.
4. What “using AI” in government actually means
“Using AI” can mean a citizen chatbot, a facial-recognition trial, or a minister photographed next to a screen. Those images are a poor guide. For a ministry, use means something drier: the official environment in which staff retrieve, draft, translate, and compare, under classification rules, with a record of what was asked and what was accepted.
Several uses are already obvious, and none requires the machine to decide. The first is retrieval across the official pile. Keyword search finds only the words the official already knows. A grounded system can surface the circular, the Council conclusions, the parliamentary answer, and the impact assessment that treated the same point in different language. The second is alignment: past answers and legal bases assembled so that a human sees contradiction before it becomes a public mistake. The third is translation and multilingual compression — a daily cost in Brussels still treated as a craft mystery. The Commission’s eTranslation service, now with generative features, shows that the Union has been in this business longer than the present debate admits. The fourth is the comparison of options before a file converges. A model can be told to keep the discarded options visible. That divergence is useful before clearance. After clearance it is noise.
What use does not mean is equally important. It does not mean an unsupervised answer to a citizen on a benefit or a police matter, or a model that “learns who gets approved.” An administrative decision that cannot be explained is a defective act. It also does not mean a headcount cut announced before a change in method. If the posts go first and the tool arrives later, the remaining people simply carry more of the old procedure. That is not leverage. It is the automation of overload.
Two public experiments are now on the record and should be read without romance.
The European Commission has put an in-house generative tool, GPT@EC, on every official’s desk, with an API into other systems, an internal register aligned with the AI Act, and an interservice steering group. A first open-source release has been announced for 2026. Digital statements are now attached to legislative proposals; an internal system drafts them and policy officers keep the last word. That is the right sequence: a controlled environment, a record, and a human signature.
Japan’s Digital Agency has built a government generative-AI environment formally called GENAI and pronounced Gennai. A Digital Agency trial began in May 2025. A government-wide pilot, launched in May 2026 and running into March 2027, was opened to about 100,000 officials at the end of May, with a target of some 180,000 across the ministries, and full operation planned for fiscal year 2027. The published functions are unspectacular and therefore serious: chat, drafting, minutes, translation, summary, search of parliamentary answers, and legal research, with input limits by classification. Part of the stack was released as open-source software in April 2026 so that local governments would not buy the same chatbot at seventeen different prices. Japan’s AI Promotion Act of May 2025 and the Artificial Intelligence Basic Plan of 23 December 2025 both tell the government to begin with itself.
I do not over-praise either project. Distributing a tool and changing the work are different events. If the clearance path is unchanged and appraisal still rewards the absence of error, staff will choose the safe use, which is almost no use. Seat licences are not an outcome. The test is whether the same task produces a better-sourced document in less time. The same test applies to the Union’s Apply AI communication of October 2025 and the GenAI4EU pilots for administrations.
5. The cartoon regulator
The cartoon is easy to recognise. It has three features.
First, it overweights visible harms and underweights invisible machinery. Deepfakes, fraud, election interference, and the exploitation of children are real. They are also easy to show in a hearing. Procurement templates, ownership of official data, classification of prompts, and the months a startup spends in security review are not photogenic. A regulator who has never used the tools will write at length about the first list and skip the second. The second list is where European firms live or die.
Second, it takes vocabulary and mistakes it for capacity. The Brussels effect, as Anu Bradford described it, is real: a large market plus a strict standard plus the cost of splitting a product can export a rule. The General Data Protection Regulation exported a vocabulary — consent, erasure, the fine. It did not produce European platform companies of equivalent scale. The AI Act is already repeating the linguistic success. The world now talks of unacceptable risk, high-risk systems, general-purpose AI, and transparency. That is a diplomatic achievement. It is not an industrial one. A jurisdiction that owns the words and imports the models has not designed a market. It has annotated someone else’s.
Third, it writes procedures that a compliance department can digest and a ten-person company cannot. Each requirement can be defended alone: a conformity file, a training-data summary, an incident report, human oversight, a fundamental-rights impact assessment under Article 27 for public deployers. Bundled, they arrive as legal invoices before the first customer. The founder then leaves for a looser market or sells early to an incumbent, and the authority can say that the market “chose” concentration.
I am not arguing for silence. Fraud, non-consensual sexual imagery, the sexual exploitation of children, and the organised falsification of an election should be banned narrowly, quickly, and in a form prosecutors can use. Existing criminal and sectoral law will do more of that work than a new horizontal article. High-harm uses — credit, employment, policing, medical devices, critical infrastructure — need evaluation, logging, and a named human. The cartoon begins when everything else is treated as if it were those cases, and when the people writing the tests have never failed a retrieval on their own files.
Use is also industrial policy. The first customer sets the specification. If that customer is vague, slow to review, and unable to admit a defect, domestic developers do not grow up. If that customer starts small, reuses what works, and publishes what broke, they do. Government as a lead user is the instrument Brussels already owns. The AI Office will read model documentation better if its own house has already seen what a hallucinated citation looks like in a real file.
A government that does not use the systems it regulates also loses the diplomatic argument. Officials who can describe an incident and a repair take the vocabulary. Officials who offer only principles translate other people’s incidents. The Hiroshima AI Process, launched under Japan’s G7 presidency in 2023, was a reasonable attempt to set a floor for advanced systems. A floor written by governments that do not operate the technology becomes a reading list. Citizens already see a split screen: their procedures are slow; their behaviour is minutely supervised. If the same administration that cannot retrieve its last parliamentary answer then tells a startup how to log a model, the public will hear “rules for thee.”
6. Human judgement, machine memory
Judgement is the choice of a value and the acceptance of a consequence. Elected politicians and officials with a legal mandate carry that consequence. A model cannot. The core of public power — the decision that binds a person, spends money, or commits the state — must not be transferred to a system that cannot appear before a committee.
The useful division of labour is old-fashioned. The machine searches, cross-reads, summarises, flags contradiction, sets options side by side, produces a first translation, and compresses open sources. The human establishes the fact, chooses the value, makes the international commitment, explains the decision, and owns the failure. That boundary should be written down before the licences are bought. Bureaucracy, at its best, has always tried to separate processing from decision. Generative AI can accelerate the first. If it is used to accelerate the second, democratic thinness follows.
Two failures sit on either side of this line. Prohibition looks clean. Staff terminals are locked; the official system is “safe.” The work does not vanish. People open a personal account at home and paste the document they cannot finish. The risk is not removed. It is moved into the dark. The OECD’s 2026 paper on generative-AI experimentation in government is blunt: unofficial use is already ahead of official governance. Laissez-faire looks modern. Everything becomes a draft; sources are not checked; the manager relaxes because the prose is smooth. Speed rises. The organisation’s ability to verify rots.
The space between those failures is a defined use. Which classification of document may meet which model on which network? Of the output, what is a draft and what is a final act? Who reads the logs, for how long, and for what purpose? Until those three questions have answers, “adoption” is an exhortation, and the best officials will not move. They are not conservative. They will not accept a task whose responsibility they cannot see.
Official memory is the prize. Rotation and inter-service work destroy context. If an official can reconstruct, in the first week in post, the last ten years of decisions, evaluations, audit findings, and repeated parliamentary questions in that portfolio, the cost of rotation falls. Memory must include failure. A system trained only on official triumph will repeat official optimism. Evaluation reports and cancelled schemes have to be in the corpus on purpose.
Security and organisational learning remain in tension. A foreign service that pours sensitive material into a consumer chatbot has sprung a leak. The opposite error is to refuse any official environment and then lecture firms about hygiene. The adult position is a government network that can process open information at scale, with a human matching it against what must remain closed.
7. Recommendations
These recommendations are addressed first to the Union institutions, not only to Member States. The Commission is the author and a principal enforcer of the AI Act. The Council secretariat is the mill through which national positions become text. The Parliament’s administration and the European External Action Service live on multilingual documents under time pressure. If those houses cannot use the technology, guidance to national digital ministries will not repair the market.
Treat internal use as part of regulatory competence. The AI Office, national market-surveillance authorities, and the services that draft implementing acts should include people who have used official tools on official files. A sandbox that only industry enters teaches the authority to read applications. A sandbox that the authority also enters teaches it to read products.
Measure leverage, not licences. Track time to a sourced first draft, citation errors caught before signature, and the share of outputs accepted, edited, or rejected. If those numbers do not move, stop celebrating access figures. Japan’s Gennai programme will fail or succeed on the same test.
Build official memory before buying another model. The limiting factor is usually the corpus: legislation, guidance, parliamentary answers, evaluations, and audit reports, machine-readable, with classification the retrieval layer will honour. The Publications Office, national gazettes, and ministry registries are industrial policy in this narrow sense.
Rewrite procurement so that a competent small vendor can sell. Specifications should name the job — grounded retrieval, logging, classification barriers, human confirmation — not a brand. Where the Commission or a large Member State has paid for a stack, release what can be released, as Japan did with part of Gennai and as the Commission has said it will do with GPT@EC. Twenty-seven ministries buying twenty-seven chatbots is an older digital failure repeated.
Change appraisal before changing the homepage. Using a system and showing sources is the standard. Taking longer because one refused the tool is not a virtue. Using the tool and not checking is a disciplinary matter. Training that teaches only prompt style will fail. So will a culture in which juniors use the system and the signing official does not.
Keep the human on the decision. High-risk public uses — law enforcement, benefits, migration, employment, credit — require a named official who can explain the act. Article 27 assessments record that the habit was attempted; they do not replace it. Do not announce headcount cuts as the business case. Ageing societies need more output per official. That is an argument for leverage, not for a smaller service running the old procedure.
Use the Union’s own multilingual load as the proving ground. Council working papers, COREPER notes, and parliamentary questions are a better test than a demonstration to visitors. If a system cannot keep “notes” distinct from “welcomes,” or hold a legal base steady across 24 languages, it is not ready for a file. Export practice, not only norms. The Hiroshima vocabulary and the AI Act’s categories will travel further if European administrations can describe their own repairs.
8. Conclusion
Bureaucracy and generative AI fit each other in one respect the public debate keeps missing. Both live on unstructured text. The administration already is the industry. The tool lowers the cost of memory. It does not abolish the official.
Capable services stall not because they lack clever people, but because a grammar of documents, clearance, coordination, and appraisal punishes speed and forgets context. AI will not rewrite that grammar by being installed. It can make the grammar less expensive to operate, if official memory is treated as an asset rather than a leak.
A government that refuses the experiment will still regulate a cartoon of visible harm, with too little to say about procurement and data, and with procedures the large can price in and the small cannot. Europe has seen this pattern. Vocabulary travelled. Industry did not.
The alternative is not a smaller civil service and it is not a model with a stamp. It is a more leveraged official, a human signature, and a state that knows the market because it is in it. The Commission, the Council secretariat, and the ministries that will police high-risk systems need to become competent users. Until they are, they should be slow to tell the rest of the economy how the work is done.
Europe will not design a better AI market than the governments that police it can operate.
Selected references
- Anu Bradford, The Brussels Effect: How the European Union Rules the World (Oxford University Press, 2020).
- Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act).
- European Commission, “Apply AI Strategy,” COM(2025) 723 final, 8 October 2025; “The Commission internal use of Artificial Intelligence,” factsheet, October 2025 (GPT@EC; AI@EC).
- Digital Agency of Japan, “Launch of Large-Scale Pilot Project of Government AI ‘GENNAI’,” 6 March 2026; update 28 May 2026; “Release of Government AI ‘GENAI’ as Open Source Software,” 24 April 2026.
- Cabinet of Japan, Artificial Intelligence Basic Plan, 23 December 2025, pursuant to Act No. 53 of 2025.
- G7, “Leaders’ Statement on the Hiroshima AI Process,” 30 October 2023; Hiroshima AI Process Comprehensive Policy Framework, December 2023.
- OECD, Digital Government Outlook 2026; Generative AI experimentation in government, July 2026.