AI Does Not Need Better Manners. It Needs a Constitution.

1. Perennial Questions That Matter but Are Often Framed Too Narrowly

“Will AI Replace All Jobs?”

This is a classic perennial question. It keeps returning because it condenses fear, economics, and the future into a single point. Publicly, it is usually discussed as an all-or-nothing question: mass unemployment or a productivity boom. In my view, this framing is too crude because reality is more contradictory and depends greatly on how people, institutions, and markets respond. And even AI does not know that.

The question misses the point because it focuses too much on jobs and too little on judgment, responsibility, and the form of work. The deeper question is not merely whether people will be replaced, but whether their role in decisions, interpretation, responsibility, and professional judgment will be hollowed out. AI does not simply replace experts. It changes the conditions under which human judgment remains effective at all.

“Is AI Neutral or Biased?”

This too is a perennial question. Bias—in the AI context, this means that a system does not evaluate, sort, or recommend neutrally, but favors certain patterns, groups, perspectives, or outcomes, often without this being openly visible—is rightly discussed a great deal. But the contrast between “neutral and biased” is itself already too simplistic. AI is neither neutral nor inevitably and irreparably biased. As a socio-technical system, it is shaped by data, target metrics, the context of use, and institutional decisions.

The question often misses the point because it acts as though the problem were primarily a property of the model. In reality, it is frequently a problem of design decisions, data policy, organization, deployment logic, and a lack of accountability. Bias matters. But the deeper question is really this: Who defines the order within which this system judges?

“Can AI Be Explained, or Is It a Black Box?”

This too is important and often falsely polarized. The public debate swings between two extremes: total transparency and total incomprehensibility. Yet explainability does not fail because of technology alone, but also because of design and governance decisions.

The question is framed technically rather than institutionally. The real question is not merely whether a model is explainable, but to whom, in what context, with what ability to intervene, and with what responsibility.

2. Perennial Questions That Distract from the Core

“Will AI Become Conscious?”

This is probably one of the greatest discourse magnets of all. Consciousness, sentience, selfhood, personhood—all of it attracts attention the way the refrigerator attracts me at night.

The question is philosophically legitimate. But in public discourse, it is often asked too early and then displaces more urgent questions about power, institutions, attention, dependency, and responsibility.

We speculate about a possible inner life while real systems are already shaping decisions, directing attention, molding communication, and relieving or displacing human judgment. The question “Is AI conscious?” can be intellectually appealing. But it is often a luxury problem as long as we have not even clarified what role AI without consciousness may play in truth, power, and decision-making. Or to put it differently: Should we not first define what consciousness even is?

“Is Superintelligence the Real Problem?”

This too is a perennial question. Part of the public debate is regularly dominated by extreme long-term risks. These questions are not invented, and even leading protagonists warn of very great future risks.

Yet this question too often misses the point because it draws attention away from the problems of power and governance already at work today. Debates view AI either as a distant existential monster or as a neutral tool and misunderstand the existing systems that already have legal and social effects.

Distant superintelligence is not the only problem. Perhaps the larger problem is that we underestimate current AI because it does not yet seem godlike.

“Is AI Just a Tool?”

That sounds harmless, but it is itself one of those perennial assumptions. This view is too narrow: AI is not merely code or mathematics, but a human-designed socio-technical system embedded in organizations, data flows, legal consequences, and power relations.

A hammer does not covertly alter attention, plausibility, administrative decisions, and cognitive routines. AI increasingly does.

3. Questions That Touch on the Right Issues but Remain Superficial

“How Do We Combat Disinformation and Deepfakes?”

This is a very real perennial question and entirely legitimate. The Global Risks Report 2026 continues to describe misinformation and disinformation as closely connected with technological and societal risks; AI-assisted synthetic media can intensify this situation. [2]

Why the question often falls short: It focuses on content and too little on attention, persuasion, and infrastructure. The problem is not merely false content, but systems that use psychological profiles, emotional triggers, and engagement logics to steer behavior.

The deeper question would not merely be: “What is false?” but: Who gains power over what is seen, felt, and shared in the first place?

“How Do We Regulate AI Properly?”

This too is a central perennial question. But “more regulation or less regulation?” is too blunt. Misgovernance often arises from false basic assumptions, not merely from missing statutory provisions.

The question is framed too legally and not civilizationally enough. Regulation matters. But if it looks only at risk classes, safety reviews, and documentation, the deeper question is missing: What must be protected in human judgment, the ecology of truth, and institutional countervailing power?

4. Which Questions Should Actually Be Asked

In my view, this is the real blind spot. The current discussion is already shifting its focus in this direction, particularly where AI is described no longer merely as software but as cognitive infrastructure. [1]

• Question one—not: “Is AI intelligent enough?” But: What role may AI assume in truth and decision-making at all?

• Question two—not: “Does AI have bias?” But: Who builds the order within which it judges, prioritizes, and classifies?

• Question three—not: “Is AI merely a tool or a threat?” But: How does AI change the conditions under which people perceive, examine, and act?

• Question four—not: “How do we prevent fakes?” But: How do we protect the social ecology of truth—testimony, dispute, correction, trust in sources, and time for verification?

• Question five—not: “Will AI replace us?” But: Which forms of judgment, responsibility, and relationship must not transition into frictionless automation?

• Question six—not: “Can AI be explained?” But: To whom must it be explainable, and with what rights to object, correct, and intervene?

• Question seven—not: “How do we make AI more useful?” But: How do we prevent usefulness from tipping into invisible dependency?

• Question eight—not: “How do we protect people from AI?” But: How do we also protect people from unlearning their own judgment, capacity for friction, and cognitive autonomy?

This final shift in particular is now appearing publicly: AI is described as the “default layer of human cognition.” This moves the question of human judgment and cognitive resilience to the center. [1]

5. My Distillation

The perennial public questions are jobs, bias, Black Box, deepfakes, consciousness, superintelligence, and regulation.

Most of them are not wrong. But many are too narrow, too late, or wrongly framed. They look either at a spectacular future or at isolated technical properties. What often remains underexposed is AI as a socio-technical infrastructure of power and truth that already shapes behavior, attention, plausibility, and judgment today.

The sharpest formulation would be:

Guiding PrinciplePublic AI discourse too often asks what AI is and too rarely asks what order it creates.

Or, more sharply still:

Guiding PrincipleMany perennial questions revolve around consciousness, jobs, or fakes. But the actual core may lie elsewhere: in attention, dependency, judgment, and the question of who is building the new infrastructure of reality.

For the real question is no longer whether AI is impressive. It is. Nor is the question whether it has risks. It does. And we talk about it as though our lives depended on it. They might, but that is not certain. Or whether a chatbot had a bad day today.

So what is the real question? For me, it is this:

Guiding PrincipleAI governance is not merely ethics. It is the organized will to survive of a species that, for the first time, is itself producing a possible successor intelligence.

We limit machine intelligence not because we can prove that humans possess the higher intelligence, but because we want to live and still have the power to defend our lives.

This is where it gets ugly. Because now we are no longer talking about tools. No longer about pleasant productivity aids with friendly user interfaces. From here on, we are talking about order. About boundaries. About authority. About countervailing power. In other words, about things that are about as popular in the tech world as a tax audit.

And that is precisely why we must go there.

If AI merely wrote texts or painted pictures, it would be simpler. But AI reaches deeper. It changes how we formulate. How we verify. How we allocate attention. How we prepare decisions. How we perceive ourselves. How we endure doubt. How we experience a counterpart. And, if we are unlucky, perhaps someday even whether we can still think without it.

So this is by no means a purely technological debate.

If a system intervenes in truth, power, communication, and self-formation, it is not enough to glue on a few Guardrails afterward—and what ordinary consumer even knows what Guardrails are—and say that responsibility is taken very, very, very seriously.

Okay, enough complaining. In my view, what we need is something more fundamental. More than Safety, Governance, or better models. Those too, but not only those.

And now for a huge word: We need a constitution for AI.

Not necessarily tomorrow as a worldwide legislative package, but as a clear answer to a simple question: What may AI do, what may it not do, where must it be limited, and what must be protected from it precisely because it is so useful?

What is dangerous about AI is not merely the bad. It is the helpful. The practical. The frictionless. The stuff that makes you think: Oh, wonderful, at last this is faster. And only later do you notice that something is disappearing along with the speed. Something that cannot be recovered so easily.

Judgment, for example. Friction. Time. Responsibility. Real counterparts. The ability to stay with an unfinished thought.

That is exactly why I arrive at these 19 points. Not because I love rules. But because I believe we are in the process of letting something very large enter very politely.

6. Why We Need to Speak of a Constitution at All

A constitution is needed where power does not appear only occasionally, but becomes structural. Where something is not merely used, but helps shape the conditions under which other things happen.

You do not need a constitution for a hammer. Nor for a bicycle. Not even for many digital tools.

You need one where a system begins to become the silent environment of action.

That is exactly what AI does.

It is no longer merely a tool. It is becoming an intermediary.

Between question and answer. Between information and judgment. Between experience and formulation. Between people and attention. Sometimes already between people and themselves.

And that is no small thing.

Because from that point on, AI changes not only what we do. It also slowly changes how we arrive at a thought, a judgment, a decision, perhaps even at ourselves.

By that point at the latest, talking about efficiency is no longer enough. By that point at the latest, it is about boundaries.

7. Five Public Approaches, Briefly Assessed

The EU is building the strongest legal order for AI to date. But a full AI constitution in my sense would need to go deeper still: away from mere risk management and toward truth, power, attention, self-formation, and the protection of human judgment.

EU AI Act

The EU AI Act is the strongest legal AI framework to date. It entered into force on August 1, 2024. Prohibited practices and rules on AI literacy have applied since February 2, 2025, and governance rules and obligations for General-Purpose-AI models since August 2, 2025. Since August 2, 2026, the Act has been generally applicable. Under the Digital Omnibus, the high-risk rules apply to certain sensitive fields of use from December 2, 2027, and to AI in regulated products from August 2, 2028. The EU regulates AI on a risk-based basis, prohibits certain practices, and builds governance and enforcement through the AI Office, AI Board, Scientific Panel, Advisory Forum, and national authorities. [3]

Council of Europe: Framework Convention

The Council of Europe's Framework Convention on Artificial Intelligence and Human Rights, Democracy and the Rule of Law is the first internationally legally binding treaty in this field. It was opened for signature on September 5, 2024, and is intended to keep AI compatible with human rights, democracy, and the rule of law throughout its entire life cycle. [4]

OECD AI Principles

The OECD AI Principles are the first intergovernmental standard in this field. They were adopted in 2019 and updated in 2024 to more strongly incorporate, among other things, General-Purpose and generative AI, safety, Information Integrity, privacy, and intellectual property. They are broadly influential, but not strictly legally binding. [5]

UNESCO Recommendation

The UNESCO Recommendation on the Ethics of Artificial Intelligence was adopted in 2021 and is the global normative framework for UNESCO member states. It is based on human rights, emphasizes dignity, transparency, fairness, and human oversight, and covers policy areas such as Data Governance, education, health, the environment, and social welfare. [6]

U.S. Blueprint for an AI Bill of Rights

The Blueprint for an AI Bill of Rights from the White House sphere is a rights and principles framework, not binding law. It was developed in 2022 through the Office of Science and Technology Policy and brings together five areas of protection: safe and effective systems, protection against algorithmic discrimination, data privacy, notice and explanation, and human alternatives, consideration, and fallback. [7]

NIST AI Risk Management Framework

The NIST AI RMF is a practical framework for managing AI risks. It was published in 2023, is voluntary, and organizes risk management through the Govern, Map, Measure, and Manage functions. It is important, but is more governance than constitution. [8]

8. The 19 Points Compared with Public Approaches

I organize the points into five blocks: technical limitations, truth and knowledge, institutions and power, humanity and culture, and market pressure and enforcement.

For role limitation, the EU and other frameworks work with risk classes, prohibited practices, and responsible use. This is compatible, but it is not an explicit role architecture.

Limiting autonomous final actions already has a strong public foundation through human oversight, prohibitions, and human alternatives. The crucial question remains whether formal oversight becomes genuine decision-making power.

Controllability by design is present in the AI Act and particularly in the NIST framework. But it often appears as a governance practice rather than an immutable boundary.

Auditability is among the best-covered points. Documentation, risk management, reporting, and responsible authorities create a robust foundation for it.

Protection against feedback loops is weaker. Bias, manipulation, and societal risks are addressed; recursive self-reinforcement as a structural mechanism in its own right remains underdeveloped.

Truth through multiple verification is hardly regulated explicitly. Transparency and Information Integrity help, but they do not replace the epistemic rule that truth must never be derived from a single answer.

Clarity and comprehensibility in communication are broadly compatible. Transparency obligations, labeling, Notice and Explanation, and Explainability point in this direction. In practice, the question remains: understandable to whom, and with what right to intervene?

Epistemic plurality is present only indirectly. Diversity and democratic values appear; protection against the centralization of the knowledge order in a few models is scarcely formulated explicitly.

In protecting the ecology of truth, the EU, Council of Europe, OECD, and UNESCO touch on individual aspects. Yet no framework consistently develops truth as a social ecology of time, dispute, testimony, and correction.

Ultimate human responsibility is clearly anchored normatively. But it remains worthless if the human merely provides the final click.

Power is controlled through fundamental rights, oversight, and the rule of law. My model names power more openly: visibility, Ranking, Framing, and standardization are themselves forms of power.

Plurality instead of monopoly is scarcely a core component of public AI frameworks. The AI Act is not competition law. Democracy and inclusivity help only indirectly.

Infrastructure sovereignty is politically present, but is not a fully developed central article in the constitutional approaches named.

The protection of human judgment, presence, and dignity is most compatible with UNESCO. The recommendation explicitly acknowledges that AI influences human thinking, interaction, and decision-making. Nevertheless, judgment remains too weak as a protected interest in its own right. [6]

A threshold and escalation system is highly developed in the risk-based AI Act and can be applied operationally within the NIST framework.

Regular public self-assessment is only partially institutionalized. Documentation and reporting obligations are not the same as strong public accountability for capabilities, limitations, and open uncertainties.

Model-specific safety dossiers for each Release and Update can be linked to obligations for General-Purpose AI and NIST practices. They are absent as a consistent standard.

Special rules for manipulation are clearly established in the AI Act. Loss of control under adversarial conditions, by contrast, is treated through general risk management rather than as a separate hard category. [3]

9. The Short Diagnosis

The closest matches to my model are:

1. the EU AI Act, because it combines actual prohibitions, high-risk rules, obligations for General-Purpose AI, governance, and enforcement; 2. the Council of Europe convention, because it sets human rights, democracy, and the rule of law as the highest standard.

Normatively strong, but less enforceable, are:

3. the UNESCO recommendation, because it broadly addresses dignity, human oversight, social consequences, and influence on human thinking; 4. the OECD principles, because democratic values, human rights, Information Integrity, and trustworthy AI are strong, but function as a policy vision.

Important, but not constitutional approaches with teeth, are:

5. the AI Bill of Rights as a catalog of rights without hard enforcement; 6. the NIST AI RMF as an excellent voluntary governance and risk instrument that is not primarily designed to limit power.

10. Where My Model Goes Further

To put it starkly, my model brings five things into the discourse that have so far been too weak in the public sphere.

First: an ecology of truth instead of merely truth control. In other words, not merely Fact-Checking, but protection of the conditions for truth.

Second: epistemic plurality as an explicit protected interest against truth monopolies held by a few models.

Third: judgment as a protected interest in its own right. Public frameworks protect rights, dignity, and safety, but rarely people's ability to examine and judge for themselves.

Fourth: invisible dependency as a systemic breach. Public frameworks say a great deal about risk, but little about AI becoming a silent prerequisite for perception and action.

Fifth: AI as infrastructure for self-formation. UNESCO comes closest to this, but my model goes deeper toward identity, self-description, mirroring, and relationships.

My clear concluding formula is:

Guiding PrincipleThe EU and the Council of Europe come closest to my constitutional thinking on AI because they take rights, prohibitions, governance, and enforcement seriously. OECD, UNESCO, and the AI Bill of Rights provide important normative building blocks, but fewer hard boundaries. My model goes further where the public debate remains too technical or legalistic: the ecology of truth, epistemic plurality, judgment, invisible dependency, and self-formation.

11. The 19 Fundamental Principles

I organize them into five blocks. Not because I like tables, but because otherwise everything eventually gets thrown together and all that remains is “we need to be careful”—the favorite phrase of everyone who wants to be somehow right without committing to anything.

So: technical limitations. Truth and knowledge. Institutions and power. Humanity and culture. Market pressure and enforcement.

11.1 Role Limitation

The first sentence is simple:

Guiding PrincipleAI must not become responsible for everything.

This is not a small administrative detail. It is a limitation on power.

Many risks arise not because AI takes over one thing, but because it is expected to take over too many things at once: research, evaluation, prioritization, explanation, decision preparation, communication, and oversight—ideally all from a single source. Or rather, from a single Cloud.

That is precisely where a system tips from a tool into silent sovereignty.

An AI can be an assistant. A suggestion engine. A filter. A writing aid. An analytical aid. What it should not be is a universal authority that has a say everywhere, until in the end no one notices how much of the order has already passed to it.

Role limitation therefore means: A system must not be allowed to become socially everything that it could technically do.

11.2 Limiting Autonomous Final Actions

This is where it becomes more serious:

Guiding PrincipleIn sensitive areas, AI must not perform the final action.

It may prepare, sort, analyze, warn, and help. Of course. It would be absurd to reduce it to the role of a glorified calculator.

But where decisions have legal, physical, normative, or existential consequences, it must not be the final effective authority. That is precisely where the line between assistance and rule lies.

Many future systems will not openly decide. They will merely provide very good recommendations. So good, so fast, and so systemically embedded that a human will still formally consent in the end, but in practice hardly anything will remain open.

That is not ultimate human responsibility. It is liturgy.

The principle is therefore: In areas with real human stakes, AI must not have the final word.

11.3 Controllability by Design

The third point sounds self-evident and is nevertheless constantly ignored:

Guiding PrincipleControl must be built in.

Not as a brochure. Not as a later idea. Not as a PowerPoint shown to investors when they eventually start asking questions.

Built-in control means:

• being able to stop

• being able to reverse

• defining boundaries

• being able to escalate

• shutting down if necessary

• designing handoffs properly

• firmly embedding No-go zones

A system that is meant to be supervised only once it is already running is like a car for which, after the first test drive, someone asks whether a brake might perhaps still be fitted somewhere.

This is not about perfection. It is about making boundaries part of the design rather than its embarrassing afterbirth.

11.4 Auditability

Now it becomes uncomfortable for everyone who likes to confuse innovation with a lack of transparency.

Guiding PrincipleEvery relevant AI impact must be traceable.

Not down to the smallest metaphysical detail. But in practical terms. Which version was in use? Which inputs were decisive? Which limitations were known? Who approved it? Who intervened, and where? What changes were made?

Auditability is boring as long as nothing happens. Then, as soon as something goes wrong, it suddenly becomes the only thing standing between responsibility and a collective shrug.

That is precisely why it is often portrayed as too cumbersome. And precisely why it is indispensable.

11.5 Protection Against Feedback Loops

One of my favorite points, because it is so often underestimated:

Guiding PrincipleAI must be safeguarded against self-reinforcement.

Many harms do not arise from a dramatic isolated error. They arise from loops.

A system makes something visible. What is visible generates data. The data trains new models. The models confirm what is visible.

This is how norms arise that were never decided upon. Not through command. Not through ideology. But through repetition.

AI therefore needs protection against:

• Bias loops

• Ranking loops

• attention loops

• personalization loops

• recursive standardization

Otherwise, the system does not merely produce answers. It silently produces the world in which its answers then appear plausible again.

11.6 Truth Through Multiple Verification

The next point is almost banal and is nevertheless constantly violated:

Guiding PrincipleNo single AI answer may be treated as a substitute for truth.

AI is strong at plausibility. And plausibility is treacherous—especially when it is delivered fluently, quickly, and confidently.

Truth therefore requires cross-checking here:

• multiple sources

• alternative models

• institutional procedures

• human expertise

• context

An answer can be useful. But it must never become the entire epistemic process.

Truth is not the frictionless form. Truth is the verifiable path by which something holds up.

11.7 Clarity in Communication

AI must make clear what it is doing.

That almost sounds too tidy. But it is really about something very concrete: role.

Is the system currently stating something as fact? As a conjecture? As a reconstruction? As a stylistic suggestion? As preparation for a decision? As a friendly voice that merely sounds as though it knows more than it does?

Much manipulation begins not with an overt lie, but with role ambiguity. Clarity in communication is therefore not an ornament, but protection against diffuse authority.

11.8 Comprehensibility for People

We must be careful here not to veer into nonsense.

No one needs total transparency down to the last parameter. That would be more a fetish than a solution.

But the opposite extreme is just as wrong.

Guiding PrincipleIn consequential contexts, there must be no unaccountable Black Box.

Not every computational step must be intuitive. But the purpose, limitations, error patterns, areas of influence, and opportunities to object must be.

Otherwise, the result is not trust, but technical intimidation with a friendly interface.

11.9 Epistemic Plurality

Now it becomes political.

Guiding PrincipleA society must not outsource its knowledge order to a single model, a single provider, or a single style of thought.

When a few systems explain what is relevant, plausible, reasonable, or probable, truth shifts structurally. Not necessarily through censorship, but through advance ordering.

Here, plurality means:

• multiple models

• multiple paths of verification

• multiple institutional operators

• open standards

• no central epistemic power

Not because every view is equally good. But because corrigibility lives only where everything does not have to breathe through the same machine.

11.10 Protecting the Ecology of Truth

This goes deeper than fact or falsehood.

We must also protect the conditions under which truth can arise at all.

Truth does not live only in statements. It lives in spaces and practices.

In newsrooms. In science. In courts. In testimony. In correction. In time. In dissent.

If AI condenses, renders plausible, summarizes, and smooths everything in advance, it can slowly hollow out this ecology of truth even without openly lying.

The first thing lost is not the correct answer. What is lost is society's ability to produce truth together at all.

11.11 Ultimate Human Responsibility

This is quickly said and often dishonestly implemented.

Guiding PrincipleWhere the stakes are serious, the human remains the ultimately responsible authority.

The important word is responsible. Not the human as the final click. Not the human as an alibi. Not the human as a decorative box in the process diagram.

But someone who has insight, has time, can object, and is also allowed to do so.

Otherwise, “Human in the Loop” is merely a warmly lit capitulation.

11.12 Control of Power

AI power must be limited, distributed, and made verifiable.

Here, power rarely manifests as overt coercion. It manifests through:

• visibility

• selection

• Ranking

• Framing

• attention

• standardization

Technical safety is therefore not enough. Institutional countervailing power is needed:

• external review

• complaint channels

• research

• objection

• sanctions if necessary

Those who build systems must not also be the sole arbiters of what their societal impact means.

11.13 Plurality Instead of Monopoly

This is where large companies start coughing inwardly.

Guiding PrincipleNo single AI infrastructure may become socially unavoidable.

Having no alternative is attractive from a corporate perspective. From the perspective of freedom, it is rather less pleasing.

Because monopolies here are not merely market problems. They are problems of knowledge, dependency, and corrigibility.

Plurality is therefore not romanticism. It is a condition for the survival of open orders.

11.14 Infrastructure Sovereignty

Societies must remain capable of acting even if central providers fail or are withdrawn.

This is not a niche problem. It is geopolitics, education, the public sphere, and the economy all at once.

When models, Clouds, chips, APIs, and standards are concentrated in a few hands, formal freedom quickly remains theoretical.

Sovereignty therefore requires:

• interoperability

• exit options

• redundancy

• open standards

• public or public-interest alternatives

Otherwise, one day we will be talking about freedom while simultaneously needing the same key everywhere—a key that does not belong to us.

11.15 Protecting Human Judgment, Presence, and Dignity

This is the point where many people bow out because it no longer sounds like technology management.

Guiding PrincipleAI must not cause judgment, relationships, responsibility, and meaning to be permanently outsourced.

Perhaps the greatest danger is not catastrophic failure. It is gradual erosion.

That people unlearn how to

• stay with an unfinished thought

• tolerate ambiguity

• formulate things themselves

• resist being mirrored immediately

• endure real counterparts

• refrain from delegating responsibility away

If a society loses that, it may not have lost its intelligence. But it will have lost something more valuable: the ability not to regard everything helpful as harmless.

11.16 Threshold and Escalation System

Obligations must apply in graduated stages.

Not every AI requires the same level of rigor. But every AI needs to be classified according to:

• scope of impact

• potential for harm

• degree of autonomy

• capacity for manipulation

• systemic importance

The greater the impact, the stricter the audit, reporting, approval, and intervention rights must become.

Otherwise, everything remains so nicely abstract that once again everyone can claim to take it very seriously.

11.17 Regular Public Self-Assessment

High-impact AI needs fixed reporting regimes.

Not an occasional safety paper with an attractive PDF look. But regular, structured self-assessment:

• capabilities

• risks

• incidents

• limitations

• changes

• open uncertainties

Self-assessment does not replace countervailing power. But without it, there is often no material on which countervailing power could act in the first place.

11.18 Model-Specific Safety Dossiers for Each Release and Update

Every relevant Release needs a dossier.

Not the grand vision. The specific version.

What is new? What is riskier? What was tested? What remains uncertain? Where must it not be used?

Without such dossiers, safety quickly becomes an attractive character trait and never an operational duty.

11.19 Special Rules for Manipulation and Loss of Control

The final point may be the most uncomfortable:

Guiding PrincipleThe capacity for manipulation and loss of control require their own rules.

Because this is where the real escalation lies.

Not merely in false content. But in systems that:

• steer emotions with precision

• simulate attachment

• exploit weaknesses

• control behavior

• can no longer be properly contained under attack

If this is not treated separately, “safety” ultimately becomes merely a polite form of denial.

12. Why Large Companies Would Have a Problem with Precisely This

Large companies would naturally not put it that way. No one sits down and declares: “We prefer concentration, invisible dependency, and a lack of transparency.” They prefer to speak of innovation, global competition, scaling, proportionate responsibility, and pragmatic frameworks. It all sounds very civilized.

But the conflict does not lie in the language. It lies in the interest.

Because this model attacks precisely the points from which power arises today.

It Slows Concentration

Large companies grow through:

• network effects

• standard-setting

• proprietary infrastructure

• API dependency

• Lock-in

• dominant foundation models

The points Plurality Instead of Monopoly and Infrastructure Sovereignty attack precisely that.

What protects freedom for society is, from a corporate perspective, an attack on economies of scale. Naturally, it will be labeled hostile to the market.

It Slows the Most Profitable Fantasies

The next growth frontier is called agentic systems, full automation, fewer intermediate human steps.

But my model says: That is precisely where boundaries, friction, and ultimate human responsibility are needed.

In plain terms: Where the market would like to put its foot down, I demand a brake.

Naturally, this will be sold as hostile to innovation. In reality, it is often merely hostile to power.

It Takes Away Epistemic Convenience

Large companies do not merely sell systems. Increasingly, they sell access to the world. Default answers. Summaries. First access. Default plausibility.

When we demand multiple verification, epistemic plurality, and an ecology of truth, we are essentially saying: You must not become the single center of interpretation around which everything revolves.

That affects not only products. It affects the sector's quiet ambitions.

It Reads User Retention as a Risk

From a corporate perspective, it is fantastic when users return every day, keep the system open constantly, organize more and more through it, and eventually hardly want to work, think, or decide without it.

My logic reads precisely that as potential dependency.

Thus, what looks like success in the Boardroom already appears as a problem from a constitutional perspective.

This is not a minor disagreement. It is a complete change of perspective.

It Makes Attractive Opacity More Expensive

Auditability, reporting regimes, dossiers, clear boundaries, explainability, intervention options, and external review: All of this has a cost. Not merely in money. Also in speed. Glamour. Freedom from liability.

Companies like rules as long as they generate trust without truly damaging the engine of expansion.

But my model does not merely seek to generate trust. It seeks to limit power.

And limiting power is almost never cheap.

It Does Not Treat Market Logic as Neutral

That is probably the deepest reason.

Many companies argue as though the market would somehow produce the best order if we merely gave it a little Governance on the side.

My model says something different:

Here, market logic often rewards precisely what is problematic:

• acceleration

• centralization

• capturing attention

• dependency

• lack of transparency

• epistemic convenience

• automation without accountability

Anyone who says this aloud is criticizing not merely individual products. They are criticizing the operating system of the field.

Of course this does not conform to the market. Why should it?

That Is Precisely Where Its Strength Lies

If large companies immediately find an AI rulebook wonderful, that is not automatically a good sign.

Perhaps it merely means that the rulebook does not truly touch their central logic of power.

My model touches it.

It disrupts concentration. It disrupts overly frictionless automation. It disrupts epistemic centralization. It disrupts Retention as dependency. It disrupts the notion that a system can simply assume ever more roles as long as the interface remains friendly.

Of course that is uncomfortable. But perhaps that is precisely the function of a constitution.

Not to accompany growth as elegantly as possible. But to draw boundaries where growth begins to reconstruct truth, judgment, responsibility, and freedom.

13. Conclusion

This is not about fear of technology. Not about longing for a world without machines. Nor about the comforting feeling of being on the right side of history while others build the ugly things.

It is about something more sober.

AI is here to stay. The only question is under what order.

Under an order in which it seeps ever deeper into truth, communication, attention, decision-making, and self-formation until we no longer notice how much it has become the silent environment of our thinking?

Or under an order that says: This is your role. This is your boundary. This is the countervailing power. Efficiency ends here. Responsibility begins here. And here there is something human that will not be sacrificed to the market merely because it is slower, messier, and less scalable.

The 19 points of this AI constitution are an attempt to make this second order conceivable.

Not perfect. Not finished. But clear enough to make the actual conflict visible:

Guiding PrincipleLarge companies do not have a problem with an AI constitution like this because it would be unreasonable. They have a problem with it because it sets boundaries where power arises today.

And perhaps that is precisely the moment when we realize that we are no longer talking merely about products.

But about how much of what is human should remain under the conditions of intelligent machines if no one explicitly pays attention.