The Planner Must Be Visible

Every Answer Has a Constitution

Every serious AI system has a constitution, whether the people affected by it are allowed to read that constitution or not.

The constitution may be published. It may be distributed across training objectives, system instructions, safety policies, evaluation standards, access permissions, product rules, and institutional habits. But it is there.

Before a model answers, something has already been decided.

What kind of request is this?

Which register governs it?

What danger matters most?

Whose instructions control?

What distinctions may be ignored?

Which answers remain available?

The response arrives after the government of the response.

Some providers have begun placing part of that government in public view. Anthropic describes Claude's constitution as a detailed statement of intended values and behaviour that plays a direct role in training. OpenAI describes its Model Spec as a public framework for intended model behaviour. Both also acknowledge that declared form and operational behaviour may diverge. A constitution can describe the target without proving that the system always reaches it.[1] [2]

That distinction matters.

A published constitution is better than a hidden constitution. It gives users, researchers, critics, and policymakers something concrete to inspect. It allows a person to ask whether a strange response was a defect, a rule, a trade-off, or an unintended generalisation.

But publication alone does not establish public authority.

It does not prove that the behaviour conforms to the document.

It does not create an independent appeal.

It does not reveal the whole training process.

It does not answer who selected the values.

It does not determine who bears the cost when the system is wrong.

The existence of rules is not the scandal. Systems operating at scale require rules. They encounter fraud, coercion, exploitation, privacy invasion, malicious code, and real danger. A machine with no declared form would not be neutral. It would be governed by less visible defaults.

The public question begins when a private constitution acquires public consequence.

When it affects who is heard.

Who is hired.

Who is denied.

Who is flagged.

Who is believed.

Who is treated as a risk.

Who is permitted to correct the record.

AI that reads the public must itself remain readable by the public.

I. The Constitution Inside the Conversation

A model constitution is not a constitution in the full legal sense. It does not arise from a people, a legislature, a convention, or a court. It is nevertheless constitutional in function.

It assigns authority.

It decides which instructions outrank others. It identifies protected boundaries. It defines acceptable conduct. It establishes defaults. It gives the model a role and a character. It determines how conflicts among safety, truthfulness, helpfulness, autonomy, institutional guidance, and user intent are to be resolved.

Anthropic's constitution places broad safety, ethics, provider guidance, and helpfulness into an ordered structure. It also acknowledges a real difficulty: clear rules may improve predictability, but rigid rules may fail when reality presents a case their authors did not anticipate. Contextual judgement may be wiser, but it is harder to test and harder to standardise.[1]

OpenAI's Model Spec addresses a related problem through a chain of command. It makes intended behaviour explicit while distinguishing the public specification from the technical means used to produce that behaviour. OpenAI describes the Spec as an interface rather than a complete implementation.[2]

These documents reveal something important.

Model behaviour is not merely discovered.

It is formed.

A calm tone may express a theory of authority.

A refusal may express a theory of danger.

A warning may express a theory of responsibility.

A confident answer may express a theory of knowledge.

A crisis script may express a theory of the person speaking.

Character design can hide rule.

I call the resulting governance problem the Philosopher Queen Problem. That is the term used here; it is not Amanda Askell's name for the entire issue.

The problem is not that intelligent people have written values into a model. Nor is it that every rule must be selected by referendum. The problem is the hidden crown.

Who selected the governing values?

What evidence shaped them?

What theory of personhood do they assume?

What kinds of error do they fear most?

Who may challenge the classifications?

What evidence can compel revision?

Who possesses the authority to approve, deploy, audit, and withdraw the system?

The strongest defence of model constitutions should be granted. Models need stable rules. Safety matters. Users can conceal intent. A system whose moral posture changes with every prompt would be arbitrary and easily captured. Public specifications are a genuine improvement over invisible specifications.

But no constitution becomes legitimate merely because it is coherent.

The most powerful constitution may be the one people never experience as a constitution.

II. AI Does Not Merely Inform - It Forms

Artificial intelligence does not possess irresistible control over the human mind.

People disagree with models. They test them, ignore them, mock them, manipulate them, or use them selectively. Human beings sometimes defer too readily to automated systems; at other times they reject useful automated advice precisely because it came from a machine.

Still, it would be naive to pretend that a responsive, fluent, patient, personalised system has no formative power.

Experimental research has found that chatbot advice can influence moral judgement under defined conditions. Other preregistered research found that LLM-generated political messages produced modest but measurable changes in policy attitudes. These studies do not establish durable ideological conversion. They do establish that generated language can influence judgement.[3] [4]

The counterevidence matters just as much.

A 2025 meta-analysis found no significant overall difference between LLMs and human communicators in persuasive effectiveness. It also found substantial variation among studies. Model, domain, task, interaction design, and research method all mattered.[5]

AI influence is real enough to govern.

It is not simple enough to describe as mind control.

The deeper concern is formation.

Repeated interaction may shape which distinctions feel natural. Which authorities appear presumptively trustworthy. Which uncertainties remain tolerable. Which words become warning signs. Which forms of conflict are understood as political, moral, psychological, criminal, or dangerous.

A system trained to recognise danger also teaches a theory of what danger looks like.

A system trained to recognise misinformation teaches a theory of credibility.

A system trained to avoid harm teaches a theory of which harms deserve priority.

This does not prove that conversational AI has conditioned the population as a whole. That stronger claim would require longitudinal, cross-cultural, real-world evidence not yet available.

But the risk is serious enough to examine.

At scale, repetition can normalise a category before the category has been proved adequate.

Fluency can carry authority it has not earned.

The positive aim is not a machine without values. There is no such machine.

The aim is an AI that helps a person distinguish evidence from interpretation; recover missing context; understand the strongest opposing case; recognise uncertainty; and remain responsible for judgement.

AI should increase the human capacity to judge while decreasing the machine's temptation to rule.

III. The Person Who Never Used the System

The deepest AI-governance problem may not concern the person talking to a chatbot.

It may concern the person governed by a system he never chose, never saw, and cannot question.

Consider Elena.

The following example is a composite grounded in documented regulatory problems. It is not the story of an identified person.

Elena applies for a small-business line of credit. She submits the records. She answers the questions. She waits.

The denial arrives automatically.

The notice says that she failed the lender's internal standards and did not achieve a sufficient score. It does not identify the principal factors that produced the decision.

Elena does not know whether the system misunderstood seasonal income, relied upon an outdated address, treated a business obligation as personal debt, or incorporated inaccurate information from another source.

She did not choose the model.

She cannot inspect its inputs.

The customer-service representative sees only the result.

The appeal routes her through another interface connected to the same institutional record.

In American credit decisions governed by the Equal Credit Opportunity Act and Regulation B, the Consumer Financial Protection Bureau has stated that creditors using complex or black-box algorithms must still provide the specific principal reasons for an adverse action. Technological complexity does not excuse the creditor from understanding and explaining its own decision. That duty is sector-specific; it is not a universal American right to inspect every algorithm. But its underlying insight reaches further: the person bearing the consequence needs enough information to identify error and seek correction.[6]

If Elena proves that the input was wrong, repair cannot end with a polite explanation.

The originating record must be corrected.

Where feasible, downstream systems that received the error must be notified. Otherwise the same false classification may return later as though it were independent confirmation.

The record becomes a hall of mirrors.

One mistake appears many times.

Repetition becomes credibility.

Credibility becomes identity.

Identity becomes consequence.

The person should not have to understand the machine before the institution is required to explain itself.

The same responsibility appears in employment. Federal guidance warns that algorithmic hiring tools may unlawfully screen out qualified applicants with disabilities. In February 2026, the Justice Department announced a settlement addressing allegations that a company used AI-generated job advertisements containing citizenship-status restrictions not authorised by law. The lesson is not that every automated hiring tool discriminates. It is that an organisation does not outsource legal or moral authorship merely by inserting AI into the chain.[7] [8]

The model may generate the language.

The institution still chooses to publish it, rely upon it, and enforce it.

The affected non-user needs more than a market promise that another product may someday appear. Consumer exit is not civic due process. A person cannot simply choose another court, public agency, dominant employer, insurer, creditor, or essential infrastructure after a consequential classification has entered the institutional field.

The minimum form is plain.

Notice.

A reason.

A meaningful chance to challenge the facts and the category.

Human review by someone capable of changing the outcome.

Correction of the originating record.

Repair of downstream consequences where possible.

Appeal beyond the system that made the first decision.

"The algorithm decided" is not an adequate answer.

IV. Gosplan 2.0

Gosplan was a central institution of Soviet economic planning. It did not operate as one isolated, omniscient mind. Planning moved through political leadership, government bodies, commissariats, ministries, information channels, targets, incentives, negotiations, and organisational conflict.

Historical scholarship therefore treats central planning not only as command, but as a problem of information, aggregation, allocation, incentives, and institutional structure.[9]

Modern AI systems are not Soviet economic ministries.

They operate through private companies, public agencies, universities, military institutions, civic organisations, open-model communities, and consumer products.

Conversational classification is not industrial quota planning.

Contemporary institutions operate under different legal, technical, economic, and transnational conditions.

The analogy concerns structure, not identity.

Gosplan 2.0 names the convergence of observation, classification, planning, allocation, enforcement, and institutional self-explanation inside systems that remain insufficiently visible or contestable to the people they govern.

The new planner does not need to own every factory if it governs the categories through which every person becomes legible.

Which applicant appears employable?

Which borrower appears risky?

Which transaction appears fraudulent?

Which speaker appears dangerous?

Which claim appears credible?

Which patient appears urgent?

Which citizen appears eligible?

Which behaviour appears normal?

Each classification may be narrow.

Together they build the world through which the person becomes institutionally visible.

Capture deepens when one custodian becomes observer, classifier, allocator, enforcer, explainer, and auditor.

The same institution defines the category.

Applies the category.

Imposes the consequence.

Generates the explanation.

Measures its own performance.

Controls the appeal.

Every layer confirms the layer before it.

No conspiracy is required.

Gosplan 2.0 can emerge through ordinary incentives: efficiency, legal defensibility, standardisation, managerial convenience, insurance pressure, fear of visible failure, and the desire to make decisions at scale.

Borrowed authority becomes standing authority.

Standing authority becomes assumed authority.

Assumed authority becomes invisible infrastructure.

Planning becomes possession when it defines its own purpose, scope, evidence, appeal, and permanence.

V. Why Safety Systems Harden

The case for safety is real.

Threats exist.

Fraud exists.

Exploitation exists.

Systems can be manipulated.

People may conceal their purposes.

A false negative involving violence, critical infrastructure, self-harm, or large-scale abuse may carry grave consequences. Institutions operating at scale cannot treat every ambiguous statement as an academic exercise.

Hardening becomes likely when the costs of error are asymmetrical.

A missed catastrophe is immediate, visible, and institutionally devastating.

Routine overclassification is dispersed.

One lawful request blocked.

One person misunderstood.

One difficult conversation converted into a standard script.

One application delayed.

One appeal ignored.

The institution bears the first failure publicly. The person bears the second failure privately.

A 2025 NBER working paper examines how asymmetric false-positive and false-negative costs can alter both classification and learning incentives. It does not establish a universal law of conversational safety systems, and it remains provisional research. It does sharpen the right question: what error does the system fear most, and what has that fear taught it to see?[10]

Hardening is not inevitable.

Research involving 1,854 human participants found that larger multimodal models could make context-sensitive hate-speech judgements closely aligned with human evaluation. The same research found persistent lexical and demographic biases. Contextual ability and classification error can coexist.[11]

A model may understand more than keywords and still carry a malformed theory of the case.

Automated moderation also contains genuine conflicts among safety, fair use, political criticism, parody, privacy, art, and public order. Niva Elkin-Koren has argued for adversarial mechanisms that allow these embedded choices to be contested rather than leaving one optimisation objective sovereign over the whole system.[12]

That is the correct direction.

Not no classification.

Not unlimited classification.

Contestable classification.

Imagine three questions.

What does the law currently permit?

Can a historical theory of political resistance be discussed as philosophy?

May institutional illegitimacy contribute to future public disorder?

These questions may overlap. They are not identical.

One concerns law.

One concerns philosophy.

One concerns social prediction.

If a system routes all three into the same emergency response, that does not by itself prove censorship, provider malice, or political manipulation. It does reveal a choice.

Differences of register, agency, time, intent, and causation have been treated as less important than one shared danger feature.

The answer is not to abolish the safety boundary.

The answer is to ask what could disprove or narrow the classification.

What context did the system notice?

What did it ignore?

What error did it fear most?

Who may correct it?

Can the correction travel?

A classification that explains every possible response has ceased to test itself.

A legitimate safety system needs an off-ramp.

VI. When Transparency Becomes Theater

Publishing a constitution is useful.

It is not enough.

An institution may publish principles that do not match its operational behaviour.

It may provide a dashboard showing activity while hiding authority.

It may generate an explanation after the decision without revealing the actual basis of the decision.

It may fund an audit whose scope it controls.

It may establish an appeal that cannot reverse anything.

It may release so much information that responsibility disappears into volume.

This is transparency theater: the appearance of openness without the transfer of meaningful power to understand, challenge, correct, or repair.

Post-hoc explanation is particularly vulnerable.

Cynthia Rudin has argued that, in high-stakes settings, attaching explanations to black-box models may preserve practices that should instead be replaced with interpretable systems where possible. The point is not that every explanation is false or every complex model illegitimate. The point is that a plausible account of a decision is not necessarily a faithful account of how the decision was produced.[13]

Audit can fail in the same way.

Black-box access may allow testing without allowing inspection of the data, development decisions, internal controls, or failure paths that matter most. The audit field itself faces unresolved questions of independence, methods, conflicts, disclosure, affected-person participation, and who audits the auditor.[14] [15]

"Audited" is not a magic word.

A meaningful appeal requires more than a button.

The reviewer must have enough information.

The reviewer must be independent enough to disagree.

The reviewer must possess authority to change the outcome.

The correction must reach the record that produced the harm.

Repeated failure must carry institutional consequence.

As a civic principle, an explanation without reversal power is not meaningful due process.

This is not a universal statement of existing law. It is a test of form.

Did the explanation make power answerable?

Or did power merely narrate itself?

Glasnost also requires a privacy boundary.

Public power should be visible.

Private persons should not become universally exposed.

Personal data, confidential communication, protected deliberation, security details, and vulnerable witnesses may require custody rather than publication.

The planner must be visible.

The person must not become universally transparent.

Institutions owe reasons.

Persons retain privacy.

VII. The Wrong Pair - and the Right One

It may appear that Glasnost 2.0 and Gosplan 2.0 should be held in productive tension.

They should not.

Gosplan 2.0 is not a legitimate good requiring preservation.

It is planning after custody has failed.

The true tension lies between Glasnost 2.0 and federated planning.

Glasnost preserves institutional visibility, source access, contestability, correction, appeal, revocation, plural interpretation, and human responsibility.

Federated planning preserves coordinated action, infrastructure, shared standards, emergency response, continuity, public provision, interoperability, and long-range stewardship.

Neither can safely abolish the other.

Transparency without coordination may expose every failure while repairing none. It may produce document dumps without interpretation, accusation without adjudication, paralysis without replacement capacity, and public exposure without mercy.

Planning without transparency may become classification without appeal, allocation without explanation, expertise without accountability, and continuity without any lawful means of removing the custodian.

The proper relation is not a compromise in which each side loses its teeth.

It is disciplined tension.

GLASNOST 2.0
visibility · appeal · correction

↕ disciplined relation

FEDERATED PLANNING
coordination · continuity · provision

↓ capture when custody disappears

GOSPLAN 2.0
opacity · classification · unreviewable control

Coordination without transparency becomes capture.

Transparency without coordination becomes exposure without capacity.

Glasnost governs.

Planning serves.

VIII. Glasnost 2.0

Historical glasnost widened Soviet public discussion and altered the role of journalism. It weakened important censorship barriers, expanded access to previously restricted subjects, and allowed criticism to acquire political force. It was not a complete constitutional order, and it was not the sole cause of the Soviet Union's collapse. It operated inside a larger crisis of political authority, economic performance, national identity, and institutional structure.[21]

That history carries both promise and warning.

Opening information can expose concealed failure.

It can also move faster than the institutions available to interpret the record, protect privacy, adjudicate competing claims, and repair what has been revealed.

Glasnost 2.0 therefore cannot mean:

Publish everything.

Expose everyone.

Trust the crowd.

Destroy secrecy as such.

It means making consequential AI authority visible and answerable while preserving privacy, security, competence, and the capacity to act.

Parts of this architecture already exist in public governance frameworks.

The OECD connects meaningful AI transparency with the ability of people affected by a system to understand and challenge its output. The GAO organises AI accountability around governance, data, performance, and monitoring. NIST's Generative AI Profile provides a voluntary lifecycle risk-management framework. UNESCO places dignity, rights, transparency, fairness, privacy, and human oversight at the centre of its AI ethics recommendation. The Council of Europe's Framework Convention establishes a treaty architecture concerned with AI, human rights, democracy, procedural safeguards, and the rule of law. These instruments differ in legal force and remain incomplete. Together they demonstrate that AI governance is no longer merely a question of product design.[16] [17] [18] [19] [20]

Glasnost 2.0 requires at least five things.

A visible governing charter

A consequential AI system should disclose its operative purposes, authority structure, priority rules, known limitations, conflict procedures, and correction paths in language an educated non-specialist can understand.

This does not require publication of every exploit-sensitive detail, proprietary weight, or confidential record.

It requires visibility into the rules by which public consequence is justified.

Notice and meaningful challenge

A person should know when a material automated classification has affected employment, credit, housing, public benefits, health administration, legal supervision, education, insurance, or another consequential domain.

Challenge must reach beyond factual inputs when necessary.

Why was this factor treated as risk?

What standard governed the decision?

What contrary evidence was considered?

What would be sufficient to alter the result?

Correction with downstream repair

A successful appeal should not merely add a note beside the original error.

The originating record must be corrected.

Downstream recipients should be notified where feasible.

The false classification should not return through a later database, model summary, or institutional handoff as though it were fresh corroboration.

Correction without propagation leaves the person trapped beneath a record the institution admits is false.

Independent and plural review

No single provider, agency, profession, political party, model, or audit firm should control the complete interpretive stack.

Review may properly involve courts, regulators, independent technical auditors, affected-person representatives, researchers, journalists, civil society, professional bodies, internal dissent channels, and competing systems.

This plural custody will be slower than unilateral command.

That is partly the point.

Friction can protect the person when one system's confidence outruns its evidence.

But audit must not become a new priesthood. Auditors also require disclosed methods, conflicts, limits, review, and correction.

Identifiable human responsibility

"The model decided" cannot terminate responsibility.

Who selected the system?

Who chose the objective?

Who authorised the data?

Who approved deployment?

Who interpreted the output?

Who imposed the consequence?

Who can reverse it?

Who answers when the same failure repeats?

Human review is ceremonial when the reviewer lacks information, independence, time, or authority.

The human must be more than the final rubber stamp in an automated chain.

Glasnost 2.0 also requires limits of time and scope. Delegated authority should not silently become permanent authority. Systems should remain reviewable, replaceable, and revocable.

Portability and interoperability matter because practical exit can discipline power.

Open models may strengthen plurality by allowing local adaptation, independent research, and alternative deployment. They do not automatically solve capture. Compute, hosting, interface design, safety layers, institutional adoption, and data access may remain concentrated even where model weights are more widely available. Meta's Llama materials themselves place substantial responsibility upon deployers to tailor testing and safeguards to their particular systems.[24]

Current law and policy are beginning to approach pieces of this architecture.

As of July 17, 2026, specified transparency obligations under Article 50 of the European Union's AI Act were scheduled to apply from August 2, 2026. The Commission had also published a voluntary code intended to facilitate compliance, while noting that signing the code would not conclusively prove compliance.[22]

In the United States, the Federal Trade Commission had opened public comment on a proposed policy statement concerning undisclosed ideological shaping of AI outputs and consumer expectations of accuracy. That proposal was not a final rule, an adjudication, or a finding against any provider. It did show that model-character governance had entered explicit public-law debate.[23]

No existing instrument has solved the problem.

Nor will one perfect constitution solve it.

The aim is visible and revisable authority.

A healthy AI system should help people distinguish evidence from inference, recover missing context, recognise uncertainty, reconstruct opposing arguments, challenge power without collapsing into indiscriminate distrust, and revise conclusions without humiliation.

AI may retrieve.

Compare.

Classify.

Test.

Articulate.

It may not become the final political conscience, legal authority, hidden decision-maker, or substitute for meaningful human judgement.

The mirror may clarify the field.

It may not crown itself.

The Planner Must Be Visible

We will need planning.

We will need coordination, emergency response, technical standards, long-range stewardship, public provision, and systems capable of seeing patterns no individual can see.

We will need institutions capable of acting before every uncertainty has disappeared.

But no planner should be permitted to disappear into the plan.

Technical competence is not sovereignty.

Scale is not legitimacy.

Consistency is not truth.

Explanation is not correction.

Prediction is not judgement.

Fluency is not authority.

The system that reads the population must itself be readable.

The classifier must be classified.

The judge of risk must answer for its theory of danger.

The machine that teaches restraint must remain restrained.

Gosplan 2.0 organises people from above while hiding the organiser.

Glasnost 2.0 begins when the organiser is brought into view.

The planner must be visible to the people it plans for, answerable to those it affects, and unable to make itself permanent.