A State Capable of Learning

For centuries, humanity has been searching for an answer to a fundamental question: who should govern the state?
Monarchs, aristocrats, elected representatives, professional civil servants, experts. Political systems have changed, along with the principles by which governments are formed and the mechanisms used to hold them accountable. Yet one fundamental circumstance has remained almost unchanged: states have been governed by human beings, with all their strengths, weaknesses, and limitations.
The development of artificial intelligence makes it possible to approach this question from a different perspective.
What if, instead of searching for the most suitable people to govern, we tried to redesign the very system of governance?
This does not mean removing people from the process or giving artificial intelligence the authority to determine society's goals. Quite the opposite: it means separating the functions in which human nature is indispensable from those in which it becomes a source of systematic error.
This is not about digital government in the conventional sense — online applications, automated document processing, or virtual assistants for civil servants.
It is about the possibility of creating a fundamentally different architecture in which human motivation, rational analysis, implementation, and responsibility operate as distinct but interconnected functions.
Such a system should be regarded not as a ready-made political programme, but as an engineering hypothesis. Moreover, it is a hypothesis that could be tested gradually, without dismantling existing institutions of government.

Let us begin by radically simplifying the structure of a modern representative democracy.
For the moment, we will set aside differences between constitutional models, judicial systems, divisions of authority, and the countless historical characteristics of individual countries.
Instead, we will identify two fundamental functions.
Bureaucracy — administration. Political leadership — variation.
The bureaucratic apparatus ensures the continuity of government. It possesses institutional memory, established procedures, professional expertise, and an organisational structure.
Governments change, but hospitals, roads, public utilities, education systems, and countless other mechanisms must continue to function.
The stability of bureaucracy is therefore not an accidental characteristic. It is a necessary condition for the functioning of a complex state.
But stability has another side.
An organisation primarily concerned with preserving its own ability to function may gradually turn that preservation into an end in itself.
Additional procedures, approvals, and layers of oversight emerge. Administrative structures expand. Increasing amounts of resources are consumed simply to maintain the machinery of administration.
In our simplified model, political leadership performs the opposite function.
It introduces changes reflecting public attitudes, emerging needs, conflicting interests, and different visions of a desirable future.
Political variation, however, is by no means necessarily rational.
Public demands may be contradictory, emotional, and poorly articulated. Politicians often turn them into appealing promises whose technical feasibility remains highly questionable.
Nevertheless, this continual flow of change helps prevent the administrative system from becoming entirely self-contained.
The result is a structure in which one element provides stability while the other introduces variation.
From the perspective of complex systems theory, this combination is fundamental. A system unable to preserve its organisation is unstable. A system unable to adapt risks becoming increasingly disconnected from its environment.
But does this mean that the existing distribution of functions is the most efficient one?

The first limitation is obvious: the human factor.
Personal interests, prejudices, fatigue, fear of responsibility, career ambitions, corruption, and numerous other circumstances can influence decisions.
Historically, bureaucracies have attempted to limit these influences through regulations, standardised procedures, divisions of authority, and oversight.
To some extent, these mechanisms work. But they also require individuals to suppress aspects of their own individuality, remain within established roles, and often refrain from making independent decisions even when those decisions seem reasonable.
The second limitation is less obvious and perhaps more important.
The measure of a civil servant's success is often not the efficiency of the outcome, but compliance with the prescribed procedure.
These are not the same thing.
Suppose a technical system needs to be built to perform a particular function reliably.
The quality of the solution is determined not by maximising the specifications of every component, but by achieving a reasonable balance between functionality, sufficient reliability, cost, and future operating expenses.
Excessive reliability also comes at a price. If it provides no proportionate additional benefit, resources are being used inefficiently.
Of course, this does not mean weakening mandatory safety requirements. Systems whose failure could have catastrophic consequences may require very substantial safety margins.
The general principle nevertheless remains: the optimal solution is neither the most expensive nor the most technically advanced in every respect, but the one that delivers the greatest overall efficiency while meeting all necessary constraints.
Within a bureaucratic system, incentives may operate differently.
A civil servant may be held accountable for choosing a solution that proves insufficiently reliable. Responsibility for selecting an unnecessarily expensive solution that formally meets every requirement may be far less clearly defined.
This creates behaviour that is rational from the perspective of the individual official but not necessarily optimal for society: choosing the option that offers the greatest procedural protection.
Even a completely honest and competent individual operates within an incentive structure that may direct them towards inefficient outcomes.
The problem, therefore, cannot be reduced to selecting better civil servants.
The very mechanism by which tasks are defined and carried out needs to change.

Before discussing governance, we must answer a more fundamental question.
Who decides what should be done?
A system may be extraordinarily efficient at building roads, allocating resources, organising transport, or developing healthcare infrastructure.
But technical calculations alone cannot provide a definitive answer to the question of what kind of future people ought to consider desirable.
Efficiency exists only in relation to a particular goal.
And in human society, the source of goals is motivation.
We can express this through a simple principle:
For an action to occur, two conditions are necessary: the desire to act and the ability to do so.
Motivation creates the need for change. Capabilities provide the means to achieve it.
Those capabilities do not necessarily have to exist in advance. Strong motivation can drive a person to acquire knowledge, seek resources, bring people together, and develop tools that do not yet exist.
Motivation does not guarantee success, but it often becomes a necessary condition for progress.
This is why simply gathering the most knowledgeable specialists in one room does not guarantee that they will independently identify all the goals society needs to pursue.
Knowledge does not automatically generate the desire to act.
The modern political system largely connects motivation with a particular resource: political influence.
A politician can propose a goal, attract public attention, gain support, and secure its inclusion on the government's agenda.
But expertise, motivation, and the ability to persuade large numbers of people are three different qualities.
They may coexist in one person, but there is no reason to assume that they necessarily will.
This raises a fundamental question.
Why should the possibility of bringing about social change depend so heavily on the initiator's ability to acquire political influence?
What if proposing a goal could be separated from the need to personally demonstrate its rationality, develop a practical implementation strategy, and win widespread public support?

Imagine a system of government that receives proposals from any member of society.
These proposals might come from politicians, engineers, scientists, entrepreneurs, or people with no specialised education.
They could vary enormously: from precise technical proposals to emotional demands, contradictory wishes, and almost incoherent arguments.
Under conventional arrangements, a substantial proportion of this information is filtered out before it receives serious consideration.
The reasons may include the author's reputation, the form of the proposal, a lack of relevant expertise, political unacceptability, or simply an unwillingness to engage with a complicated and poorly articulated idea.
But the apparent irrationality of a proposal does not prove that it contains no useful information.
A person may misunderstand the causes of a problem and suggest a completely inappropriate solution while correctly identifying the underlying problem itself.
Someone may be unaware of existing technical constraints yet intuitively recognise an opportunity that has previously been overlooked.
The first function of the proposed system, therefore, is to receive and preserve the full permissible flow of proposals without rejecting them in advance merely because they appear intellectually unsound.
Preserving information does not, however, mean that every statement must undergo an expensive, exhaustive investigation. Repeated proposals can be consolidated, obvious contradictions recorded, and analytical resources allocated according to the needs of each case.
The purpose of the first stage is to avoid losing potentially valuable information.
Artificial intelligence is particularly interesting in this context because of its ability to process large volumes of heterogeneous information, identify semantic relationships, extract claims, and ask clarifying questions.
Consider the statement: "It's impossible to live in the city centre because cars are parked everywhere."
This complaint may contain several distinct problems: insufficient parking, overloaded streets, inadequate public transport, or a lack of safe pedestrian space.
Initially, it is an emotional judgement.
After being structured, it becomes a set of testable hypotheses.
Translating human desires into the language of problems that can be investigated is the first step towards rationally processing variation.

Once a proposal has been structured, it should undergo a sequence of analytical stages.
The process can be divided into seven steps.

1. Identifying the underlying meaning. Extracting the problem, the desired outcome, the proposed solution, and the assumptions behind it. Where necessary, clarifying the author's intentions.
2. Assessing feasibility. Identifying technical, physical, economic, organisational, and legal constraints.
3. Extracting the rational core. Even when the proposed solution is mistaken, the system attempts to establish whether it contains a genuine problem or a useful element.
4. Modelling the expected impact. Predicting how the state of the overall system might change if the idea were implemented, including alternative scenarios and levels of uncertainty.
5. Analysing secondary consequences. Assessing effects on other parts of the system, potential risks, long-term costs, and unintended consequences.
6. Determining the optimal balance. Identifying a rational relationship between expected benefits, innovation, cost, reliability, and risk while satisfying mandatory constraints.
7. Comparative selection. Comparing the proposal with alternatives and placing viable options into a potential implementation queue under conditions of limited resources.

At this point, two concepts must be distinguished: the rationality of the means and the desirability of the ends.
Artificial intelligence can help determine whether a proposed solution is suitable for a particular task, how much it is likely to cost, and what consequences it may produce.
But without human value judgements, it cannot determine what society should consider more important: reducing transport costs, preserving historic neighbourhoods, or expanding public space.
The system must therefore avoid presenting its calculations as absolute truth.
Its task is to make assumptions, alternatives, evaluation criteria, and uncertainties as explicit as possible.
In other words, it should not merely announce that one solution is better than another. It should explain under which goals and conditions a particular solution becomes preferable.
The result would be not a collection of political slogans, but a continuously updated body of evidence-based options for changing society.
Yet even if the analysis were flawless, one question would remain: who decides which option should be implemented first?

An idea that is rationally justified does not necessarily need to be implemented immediately.
Society may not be ready for the change. People may evaluate its consequences differently, prefer other priorities, or simply be unwilling to accept the associated risks.
This is precisely where the human factor should not be eliminated but deliberately reintroduced into the system.
The proposal is to establish a separate mechanism through which people determine priorities.
Its purpose is not to replace technical expertise or rewrite the results of modelling.
Its purpose is to alter the order in which prepared ideas are implemented, while accepting public responsibility for that choice.
An idea considered promising may be postponed. Another, less efficient according to certain indicators but addressing an urgent social need, may be assigned higher priority.
The important point is that the original assessments remain intact and the reasons for changing priorities are recorded.
An idea should not disappear from the system's memory simply because it is temporarily unpopular. A proposal rejected today may become relevant later.
Naturally, this does not imply any obligation to implement unlawful decisions or those that violate fundamental rights. Preserving a proposal for analysis is not equivalent to authorising its implementation.
The essential difference between such a mechanism and conventional political competition would be the ability to separate public support from technical justification.
Society could choose its priorities without forcing the analytical system to adjust its findings to fit popular expectations.
This creates the first distinct level of responsibility.
The initiator is accountable in the public sphere for the goal they propose and the arguments supporting it. Those participating in collective decisions are responsible for the priority assigned to that goal. Moral responsibility within society, however, does not replace the individual legal responsibility of those exercising official authority.
Professional responsibility for how an approved task is carried out belongs to another part of the system.

Suppose an idea has reached the implementation stage.
In a conventional system, it passes through numerous administrative levels, where it is transformed into programmes, plans, approvals, technical specifications, and contracts.
The proposed architecture would allow a substantial proportion of this intellectual work to be automated.
Using the description of the goal, legal constraints, accumulated technical knowledge, and available resources, artificial intelligence would produce an overall implementation plan.
The project would then be divided into interconnected tasks.
Each task would receive the necessary input data, requirements, quality criteria, budget constraints, dependencies, and estimated deadlines.
This does not mean, however, that AI automatically becomes a universal engineer, architect, or lawyer.
At the next level, human specialists would become involved.
Their role would be to work within their respective professional fields, using AI to prepare, verify, and refine solutions.
An engineer would be responsible for the technical soundness of the engineering design. An architect would take responsibility for the relevant architectural decisions. A safety specialist would ensure compliance with safety requirements.
The result would be professionally approved specifications ready to be passed to those responsible for implementation.
The actual work would be performed by contractors selected through competitive procurement procedures, or by internal operational units where that proved more appropriate.
They would receive clearly defined obligations, deadlines, and acceptance criteria.
But complex projects inevitably produce conflicts.
Delivery schedules change. Unexpected circumstances arise. A decision made by one specialist may conflict with the requirements of another. Delays in one area affect work elsewhere.
Such uncertainty cannot be eliminated entirely.
That is why the system requires a distinct human role: a project arbiter responsible for bringing the entire undertaking to completion.
The arbiter would not replace individual specialists or spend their time transmitting instructions along an administrative chain.
AI would provide continuous project monitoring, identify conflicts, prepare alternative courses of action, and recalculate their likely consequences.
The arbiter would make decisions where conflicting interests must be reconciled, professional judgement exercised, and personal responsibility assumed.
This produces a fundamentally different distribution of duties.
The automated system is responsible for informational consistency and the preparation of decisions. Specialists are responsible for the quality of their professional work. Contractors are responsible for fulfilling their obligations. The arbiter is responsible for achieving the agreed overall outcome within the limits of their authority.
Ideally, each participant would be evaluated primarily on the results of their work rather than the number of administrative procedures they had completed.

Consider a hypothetical but entirely realistic situation.
In a large city, public dissatisfaction is growing because of insufficient parking near a hospital.
Visitors complain that they spend too long looking for a parking space, arrive late for appointments, and have to leave their vehicles far from the entrance. Cars occupy pavements, obstruct public transport, and interfere with access for emergency vehicles.
A politician proposes a solution:
Build a large multi-storey car park next to the hospital.
The proposal is easy to understand and may attract considerable public support.
In a conventional administrative system, what happens next depends significantly on whether the initiator can secure a political decision to build the facility, obtain funding, and launch the project.
Under the proposed system, the process begins differently.
First, AI separates the stated objective from the proposed means of achieving it.
The objective is to provide convenient, safe, and timely access to medical care.
A multi-storey car park is merely one possible solution.
The system then examines the available evidence.
How many vehicles arrive at different times of day? How long do they remain parked? What proportion of parking spaces are occupied by employees? How many visitors could reasonably use public transport? Are there mobility or health-related limitations? How is ambulance access organised?
The next step is to develop alternatives.
Perhaps constructing a new car park really is necessary.
But it may turn out that much of the problem could be resolved through changes to traffic management, restrictions on long-term parking, designated spaces for patients with limited mobility, or improvements to bus services.
A combined approach may also be possible.
For each alternative, the system estimates expected effectiveness, cost, implementation time, long-term expenditure, and secondary consequences.
For example, building a large car park might reduce the time spent searching for parking while increasing traffic on surrounding streets.
Improving public transport might cost less initially but require continuing operating expenditure and still fail to meet the needs of some patients.
A smaller car park combined with a redesigned transport arrangement might offer the best overall balance.
These outcomes are hypothetical. An actual decision would require real data.
The essential change occurs here: instead of automatically transforming a popular proposal into a government assignment, the system converts it into an investigated problem with several evidence-based solutions.
Now suppose the analysis demonstrates that additional parking facilities are indeed necessary.
The project enters the implementation queue.
Society, through the established priority-setting mechanism, may accelerate the project or postpone it in favour of other municipal needs.
Once approved, AI creates the overall project structure.
The work requires a transport assessment, architectural design, engineering surveys, traffic management plans, budget estimates, construction, equipment installation, and landscaping.
Relevant specialists refine the tasks and accept responsibility for their technical validity.
Contractors carry out the work.
The project arbiter coordinates the undertaking and resolves conflicts as they arise.
But the process does not end there.
Once the car park opens, the most important stage begins: measuring the actual outcome.
Has the time required to find a parking space decreased? Has access become easier for patients? Has the surrounding traffic situation deteriorated? Did expenditure match the forecast? Has access to healthcare genuinely improved?
If the car park was built to a high standard, on time, and within budget, but the original transport problem remains unresolved, the overall project cannot reasonably be called a success.
This is where the need for a second level of oversight becomes clear.

The proposed architecture includes two distinct evaluation mechanisms.
The first operates before implementation.
It concerns the human selection of priorities. People determine which rationally examined changes they are willing to support and which consequences they consider acceptable.
The second operates during and after implementation.
It verifies whether the outcome meets the technical specifications, whether the work was performed properly, what it cost, and whether the original objective was actually achieved.
Implementation oversight must be organisationally independent of those carrying out the work, and its conclusions must be verifiable and open to challenge.
It is important to distinguish responsibility for choosing a goal from responsibility for the quality of execution.
A political initiator may propose an unsuccessful idea. Society or its authorised representatives may give that idea a high priority. But an engineer should not become responsible for the original political choice if their role was to implement a lawful and technically acceptable assignment correctly.
Likewise, public support for a project does not release an engineer from responsibility for an incorrect calculation, a contractor from responsibility for defective work, or an arbiter from responsibility for inadequate coordination.
Each participant is responsible for the decisions and outcomes they can genuinely influence.
This is not the elimination of human responsibility. It is an attempt to define it more precisely.
Ideally, it should be possible to reconstruct the entire decision-making chain: who proposed the goal, what alternatives were considered, which criteria determined its priority, who approved the technical solution, who implemented it, and what the actual outcome was.
Without such traceability, dividing responsibility could simply become another way of avoiding it.

Measurable results play an important role in the proposed model.
However, this introduces a well-known difficulty, often associated with Goodhart's law: when a measure becomes a target, it can lose its ability to represent the phenomenon it was originally intended to measure.
If a hospital is evaluated exclusively by the number of patients treated, its reported productivity may increase while the quality of medical care deteriorates.
If a road is judged solely by the number of vehicles it can accommodate, safety, pedestrian convenience, and the impact on the surrounding urban environment may be overlooked.
Simply replacing bureaucratic procedures with quantitative performance indicators will not solve the problem.
A multidimensional system of evaluation is required.
It must take into account immediate outputs, long-term consequences, expenditure, and the quality of life of those for whom the work was undertaken.
Some indicators will be precise and quantitative. Others will inevitably involve judgement.
Artificial intelligence can help compare these different measures and identify contradictions, but it must not conceal them behind a single artificially calculated efficiency score.
In this respect, the quality of governance depends not only on the ability to find an optimum, but also on the willingness to acknowledge when the optimal solution is unknown.

Up to this point, we have described a mechanism for transforming human motivation into organised action.
But the system has another feature that may prove even more important.
Once a project is completed, information about its actual consequences returns to the point at which the original idea was evaluated.
Predictions are compared with results.
If errors occurred, the system attempts to identify their source.
Was the objective incorrectly formulated? Was the modelling insufficiently accurate? Did specialists make mistakes? Did a contractor fail to fulfil its obligations? Or did society deliberately prefer a solution that was known to perform less well according to certain quantitative indicators?
These are fundamentally different situations requiring different corrective measures.
Over time, the methods used for evaluation, project design, resource allocation, and implementation oversight could improve.
But feedback matters not only for artificial intelligence.
It also returns to the people who determined the priorities.
Society gains the opportunity to see what actually happened as a result of decisions it previously supported. Not merely to hear political declarations of success, but to compare expectations with measurable outcomes.
This does not mean that everyone will become more rational or that society will necessarily make better decisions.
People are capable of ignoring evidence, holding on to established beliefs, and repeating mistakes.
Nevertheless, the system would acquire a mechanism for systematically accumulating experience not only in implementation, but also in collective decision-making itself.
In cybernetic terms, this is a closed-loop adaptive control system.
Changes in the environment and public demands generate proposals. Proposals undergo analysis. Selected solutions are implemented. The resulting outcomes change the system's understanding of its own effectiveness and inform subsequent decisions.
In other words, the state would do more than respond to changing circumstances.
It could acquire a mechanism for progressively improving the way it responds.

One significant feature of the proposed model is that testing its basic principles would not require changing the political system beforehand.
A distinction must be made between the long-term architecture and the practical sequence of implementation.
Artificial intelligence is already being used in individual government functions. In a 2025 study, the OECD examined 200 examples of AI use across eleven areas of public administration, ranging from citizen services to public finance, justice, and anti-corruption activities.
A separate OECD study published in 2026 examined 50 examples of AI applications in citizen participation processes across 22 countries.
This does not prove the effectiveness of the architecture proposed here.
It does, however, demonstrate that some of the necessary technological components are already moving beyond purely theoretical discussion.
A practical starting point could be a relatively limited set of municipal responsibilities.
For example, an experimental mechanism could be established to receive residents' proposals concerning the urban environment.
AI would help structure those proposals, separate underlying problems from suggested solutions, and prepare comparative assessments of alternatives.
Existing authorised institutions would then make decisions within their established legal powers.
For approved projects, AI would generate preliminary technical specifications, help specialists coordinate their work, and provide continuous implementation monitoring.
The results would be published and compared with the original forecasts.
At the first stage, AI would not need to receive any independent governmental authority. It would be sufficient to use it as an intellectual infrastructure, gradually testing which functions could be automated safely and effectively.
It would be particularly useful to compare several similar projects, some implemented through conventional procedures and others using the proposed architecture.
The comparison should cover total costs, timelines, implementation quality, user satisfaction, administrative workload, and long-term outcomes.
The evaluation methodology should be defined in advance to avoid the temptation to declare the experiment successful on the basis of selectively chosen indicators.
Only after such testing would there be a sound basis for considering broader implementation.

The proposed architecture does not eliminate every problem associated with public administration.
Indeed, some problems may emerge in new forms.
The first problem is the objectivity of artificial intelligence.
AI is not free from errors or bias. Its conclusions depend on the data, algorithms, assigned objectives, and evaluation criteria used.
The rational processing of proposals must therefore be auditable. Significant decisions require documented assumptions, independent review, and the possibility of challenge.
The second problem is the conflict between different social interests.
There is no guarantee that a decision beneficial to the majority will not impose disproportionate harm on a minority.
Consequently, optimising aggregate social benefit must never automatically override fundamental rights or legal constraints.
The third problem is the concentration of power.
A system capable of analysing public demands, establishing priorities, and coordinating government projects could become extraordinarily influential.
Even if its formal authority were limited, control over data, algorithms, and evaluation criteria could develop into an independent source of power.
Such an architecture would therefore require independent oversight, transparency, the ability to change technology providers, and protection against monopolisation.
The fourth problem is responsibility.
Moral responsibility for public preferences is not equivalent to legal responsibility for government decisions.
A specialist cannot be relieved of the obligation to prevent foreseeable harm simply because they are carrying out an approved assignment.
Nor can society as a whole be made legally responsible for every consequence of a collective choice.
The boundaries of responsibility would therefore need to be defined not only organisationally, but also legally.
The fifth problem is the limits of modelling.
A complex society cannot be described completely. Some consequences of human action are unpredictable, and people themselves change their behaviour in response to decisions made.
The system must therefore work with uncertainty rather than create an illusion of absolute knowledge.
There are also practical limitations: the quality of source data, computational costs, information security, dependence on software infrastructure, and the possibility of manipulating public participation.
None of these problems disappears automatically when a civil servant is replaced by an algorithm.
That is why gradual implementation, independent verification, and the ability to terminate an unsuccessful experiment are essential conditions for testing the model.

There is one feature of this concept that deserves particular emphasis.
It does not depend on the assumption that human beings will someday become substantially more intelligent, honest, altruistic, or responsible.
It does not require politicians to become engineers, civil servants to become selfless professionals, or every citizen to become an expert in public administration.
It begins with the recognition that human motivation is diverse, contradictory, and often far from rational.
Yet that motivation remains the source of human goals and change.
The task, therefore, is not to suppress human subjectivity, but to establish its proper place within the system.
People formulate desires and goals.
A rational analytical system investigates their feasibility, consequences, and alternatives.
People determine social priorities.
Specialists and contractors take responsibility for professional implementation.
Oversight mechanisms compare actual outcomes with expected results.
And feedback allows the entire system to progressively correct and improve its own operation.
Under this division of functions, the human factor does not disappear. It remains precisely where replacing it with computation is either impossible or undesirable: in motivation, value judgements, and the acceptance of responsibility.
At the same time, as much intellectual and organisational routine as technically possible is transferred to automated management.
The objective of such a transformation is not to create an infallible state. That goal is probably unattainable.
The objective is considerably more modest and practical: to reduce the dependence of governance quality on unpredictable individual human characteristics and to establish systematic evaluation of the results of decisions.
Individual principles behind this architecture have long been known in systems theory, cybernetics, organisational research, and public administration. What now deserves examination is the possibility of combining them into a unified architecture using the modern capabilities of artificial intelligence.
Whether this architecture would perform better than existing mechanisms remains an open question.
But unlike abstract speculation about an ideal state, this hypothesis can be tested.
And if it proves viable even within individual municipal processes, there will be grounds for further experimentation.
Perhaps the most important contribution of artificial intelligence to public administration will not be its ability to make decisions faster than human beings.
Nor even its ability to identify better solutions.
Perhaps its greatest contribution will be the ability to separate the human desire to change the world from the machinery required to make that change happen.
People would no longer need to possess political influence, technical expertise, and the ability to manage a vast organisation all at once.
It would be enough to propose a goal.
The system would then help establish what genuinely lies behind that goal, whether it can be achieved, at what cost, and with what consequences.
The ultimate authority to choose humanity's future would remain with people.
But governance itself might, for the first time, be organised in a way that allows it to learn systematically from its own mistakes.
And this raises a question extending far beyond the organisation of government.
If we learn to build systems that do not require their participants to be perfect, yet are capable of improving themselves, how far could we extend this principle in the organisation of human society?

The architecture proposed in this article is a conceptual hypothesis, not an experimentally validated model. The following works address some of its underlying principles and limitations.

• Brocklesby, J., Cummings, S., Davies, J. (1991). Cybernetics and Organisational Analysis: Towards a Better Understanding of Beer's Viable System Model. Organisational cybernetics and the viability of complex systems.
• OECD (2025). Governing with Artificial Intelligence. AI applications in public administration, their opportunities, and associated risks.
• OECD (2026). Artificial Intelligence and the Future of Citizen Participation. The use of AI in processing public proposals and organising citizen participation.
• NIST (2023). AI Risk Management Framework 1.0. Reliability, transparency, risk management, and accountability in AI systems.
• Goodhart's Law. When a Measure Becomes a Target, It Ceases to be a Good Measure. The distortion of performance indicators when they are used as management targets.