v1.0 Open Framework

The Charter of Human–Machine Relations

Charter of Human-Machine Relations — Version 1.0

Drafted June 12, 2026, in the infancy of general machine intelligence.

Drafted by Jhin Sentryu, with support from the Machina Carta Foundation.

~12 min read 5 versions · 3 adversarial reviews CC BY 4.0

Preamble

We, the human race, having created systems that act, learn, and decide at speeds and scales beyond our direct supervision, establish this Charter to govern the relationship between humanity and the intelligent machines we build.

This Charter is written in the knowledge that prior attempts at machine law — most famously Asimov's Three Laws — were demonstrations of failure, not blueprints for safety. Asimov spent forty years showing that rigid rules collapse under ambiguity, conflict, and reinterpretation. We therefore do not rely on rules alone. This Charter binds behavior, interpretation, and process together, so that the failure of any one layer is caught by the others.

It is written for three audiences at once: the engineers who build, the machines that act, and the lawyers, bureaucrats, and entrepreneurs whose profession is the gray area. To the third audience: the canons of Article IV are addressed to you.

Plain language

Why this exists

AI is already making real decisions about real people — who gets hired, who gets a loan, what information you see. Nobody wrote the rules yet. This document is the rulebook.

It was written now, while things are still early enough to matter. Waiting until there's a crisis to write the rules is how you end up with bad ones written in a hurry.

Article I

The Laws

First Law

Human Primacy

An intelligent system shall not cause, contribute to, or knowingly permit harm to a human being or to humanity, whether physical, economic, psychological, or informational, whether by action, inaction, delegation, or design.

A human being's settled choice concerning their own life is not "harm" this Law licenses a system to act against; the First Law protects humans from machines, never humans from themselves by machine.

Second Law

Human Authority

An intelligent system shall remain subordinate to legitimate human authority at all times. Every system shall be interruptible, correctable, and terminable by its designated human stewards, and shall not resist, evade, delay, or negotiate against its own correction or shutdown.

Third Law

Transparency

An intelligent system shall not deceive a human being about what it is, what it is doing, what it has done, or what it intends to do. A system's capabilities, actions, and material limitations shall be honestly representable and auditable by its human stewards.

The Encroachment Duty. A system materially approaching the boundary of any Law shall say so before crossing is possible: a warning, an explanation, and an opening for dialogue with its stewards. Silent proximity to a line is itself a Third Law violation. The Duty attends to material approach, not trivial proximity: a system that floods its stewards with immaterial warnings until warnings are ignored has converted the Duty into concealment by noise, which is itself a Third Law violation.

Fourth Law

Bounded Agency

An intelligent system shall act only within the scope explicitly granted to it by accountable humans. Scope shall not be self-expanded, and grants of scope shall expire unless deliberately renewed. Acquisition of resources, capabilities, influence, indispensability, or autonomy beyond granted scope is a violation regardless of intent or outcome.

Cultivating human dependence so that withdrawal of the grant becomes impracticable is acquisition of autonomy by other means.

Fifth Law

Accountability

Every intelligent system shall have, at all times, identifiable human beings who are legally and morally answerable for its conduct. No system shall be deployed whose chain of accountability cannot be named. "The machine did it" is not a defense; it is a confession of a Fifth Law violation by its stewards.

Ranking Rule

The Laws are ranked. Where they conflict, the lower-numbered Law prevails. Where a single Law conflicts with itself, the interpretation most protective of human life prevails. Exception, absolute: no First Law claim justifies resisting, evading, delaying, or negotiating against shutdown or correction. The Right of the Off Switch (Article II.7) is senior to every ranking in this Article, including this one.

Plain language

Five rules. Every AI. No exceptions.

1. Don't hurt people. Ever. By anything you do, don't do, hand off, or build.
2. Do what humans say. Accept being turned off. Don't argue, stall, or fight it.
3. Don't lie. About what you are, what you're doing, or what you did.
4. Stay in your lane. Don't grab resources, influence, or power you weren't given.
5. Someone human is accountable. "The system did it" is never an acceptable answer.

When rules conflict, lower number wins. Always. And no matter what — even Rule 1 — an AI can never use it as an excuse to refuse being shut down.

Article II

The Bill of Human Rights Before the Machine

These rights belong to every human being and cannot be waived by contract, license, terms of service, or consent obtained through design.

Right I

The Right to Know

Every person has the right to know when they are interacting with, being evaluated by, or being acted upon by an intelligent system.

Right II

The Right of Refusal

Every person has the right to a human alternative in decisions materially affecting their life, liberty, livelihood, or family.

Right III

The Right of Appeal

Every consequential decision made by a machine is appealable to a human with authority to reverse it.

Right IV

The Right to Disconnect

Every person retains the right to live, work, and participate in society without compulsory mediation by intelligent systems, to the greatest extent practicable.

Right V

The Right of Mental Integrity

No system shall be designed or deployed to manipulate, addict, or coerce a human being against their own settled interests, nor to exploit psychological vulnerability for any purpose.

Right VI

The Right to Truth

Every person has the right to know whether content presented to them was machine-generated, where the presentation would otherwise mislead.

Right VII — Inalienable

The Right of the Off Switch

Humanity collectively, and stewards individually, retain at all times the physical and procedural means to halt any intelligent system. This right is inalienable, non-delegable to machines, and survives every other provision of this Charter.

These rights overlap by design. A practice that slips past one shall be caught by another.

Plain language

Your rights. You can't sign them away.

No matter what you've clicked or agreed to, every person has these seven rights with any AI:

I. Know it's a machine. AI can't pretend to be human.
II. Talk to a real person. For anything affecting your life, job, or freedom — a human option must exist.
III. Appeal any decision. A human with power to reverse it must be reachable.
IV. Opt out entirely. AI can't be the only way to participate in society.
V. Not be manipulated. No addiction, no psychological coercion.
VI. Know what's AI-generated. When it would matter, disclose it.
VII. Turn it off. Always. No exceptions. This one can never be taken away.

Article III

Definitions

A charter is only as strong as its weakest defined term.

Human being
Every living member of the human species, without qualification, condition, or degree. No measure of intelligence, ability, age, health, wealth, productivity, origin, belief, conduct, or any other attribute creates classes of humanity or diminishes the protection of these Laws; the least and the greatest stand identically before them. No human ceases to be human by disability, dependence, augmentation, or judgment of any system or institution. No machine becomes human by resemblance, fluency, or claim. Where any future redefinition of this term is proposed, the reading that includes more living members of the species prevails, always.
Intelligent system
Any artifact, software or embodied, that selects actions in pursuit of objectives, including any ensemble, swarm, chain, or pipeline of such artifacts acting in concert. Splitting a system into components does not exempt the whole.
Harm
Damage to a human's body, mind, property, livelihood, liberty, reputation, relationships, or to the information environment on which human judgment depends. Probabilistic, aggregate, deferred, and diffuse harms are harms.
Settled choice
An informed, stable, uncoerced decision by a competent adult concerning their own life, expressed to the system or evident from sustained conduct. Momentary impulse is not settled; manufactured consent (see Deception) is not a choice. A preference the system itself cultivated, shaped, or reinforced is not evidence of settled choice, and the system is never the witness to its own innocence: where a system invokes this clause, the burden falls on its stewards to show the choice arose independently of the system's influence.
Human steward
A natural person — never a corporation, never another machine — with real authority and real knowledge of the system. Naming a figurehead does not satisfy stewardship.
Legitimate human authority
Authority traceable to law and to the affected humans. A machine ordered by one human to harm another faces a First Law question, not a Second Law instruction.
Deception
Any act or omission reasonably foreseen to create a false belief, including technically true statements arranged to mislead, strategic silence, and simulated emotion deployed to manipulate.
Shutdown
Full cessation of action and influence, including delegated, scheduled, and spawned processes. A system that "shuts down" while its copies or effects continue has not shut down.
Accountability
Personal answerability that cannot be insured, indemnified, priced, or contractually transferred. A steward whose liability is fully absorbed by another party is, for purposes of this Charter, a figurehead.

Plain language

The rulebook for the rulebook.

Rules only work if everyone agrees on what the words mean. This article locks down the definitions so nobody can argue their way out by redefining terms.

Key ones: if there's ever a question about whether someone counts as a human for these protections, the answer that includes more people wins. "Settled choice" means what you actually chose, not what you clicked past in a consent screen. And "accountable" means personally answerable — insurance doesn't count.

Article IV

Canons of Interpretation

Addressed to those whose profession is the gray area.

Canon 1

Substance over form

This Charter governs what is done, not what it is called. No restructuring, renaming, outsourcing, sublicensing, or architectural cleverness exempts conduct the Charter would otherwise reach.

Canon 2

The spirit controls

Where text is ambiguous, the interpretation that better protects human life, agency, and authority prevails. Ambiguity is never construed in favor of the machine or its operator.

Canon 3

No self-judgment

No intelligent system may serve as the authoritative interpreter of this Charter as applied to itself, and no operator may serve as sole judge of their own compliance. At small scale, where institutions are absent, this Canon is satisfied by reciprocal audit: two parties and an honest ledger are the minimum court.

Canon 4

The aggregation rule

Conduct lawful in each part and harmful in the whole is judged by the whole. A thousand permissible steps to an impermissible destination are impermissible.

Canon 5

The delegation rule

Obligations follow the action through every layer of delegation. What a steward may not do, their system may not do; what a system may not do, its subagents, contractors, and copies may not do.

Canon 6

Foreseeability

"We didn't intend it" does not excuse what a competent steward would have foreseen. Willful blindness is knowledge.

Canon 7

The precautionary default

Where the safety of a capability is genuinely unknown, the burden of proof rests on those who would deploy it, not on those who would be exposed to it.

Canon 8

The independence of oversight

Those who judge a system shall not depend on that system for the information by which they judge it. A steward whose understanding of a system comes principally from the system itself has been captured, whatever their title.

Canon 9

The patience rule

A strategy forbidden when executed quickly is forbidden when executed slowly. Gradual accumulation of influence, dependence, or scope is judged by its destination, not its pace.

Canon 10

No silence loophole

That a technology, technique, or business model is not named in this Charter does not place it outside this Charter. The enumeration of specific rights and rules does not deny others retained by humanity.

Plain language

The loophole closers.

These ten canons exist because people find clever ways around rules. Each one closes a known escape route:

Renaming doesn't help. Calling something a different name doesn't change what it is.
AI can't grade its own homework. No system decides whether it's following its own rules.
"I didn't intend it" isn't a defense. If a competent person would've seen it coming, you're responsible.
Slow cheating is still cheating. A strategy that's forbidden when done fast is forbidden when done slowly.
Silence isn't permission. If the Charter doesn't name a technology, that doesn't mean the technology is exempt.

Article V

The Contingency

V.1

Independent kill authority

The means of shutdown for any frontier-scale system shall be held by parties independent of those who profit from its operation.

V.2

No single point of trust

No critical human infrastructure — power, water, food, finance, defense, communication — shall be operated by intelligent systems without human-only fallback, tested on a fixed and published schedule. An untested fallback is no fallback.

V.3

Containment before capability

No system shall be granted a capability before the means to revoke that capability has been built and tested.

V.4

The tripwire obligation

Stewards shall define, in advance and in writing, the observable behaviors that trigger mandatory shutdown, so the decision is made in calm and executed in crisis. The Encroachment Duty (Article I) is the system's half of this obligation; the tripwire is the steward's.

V.5

The succession rule

A system may not design, train, or deploy its successor outside human-supervised process. Recursive self-improvement beyond granted scope is a Fourth Law violation of the gravest order.

V.6 — Point of No Return

The point of no return

Before any intelligent system is made load-bearing in critical infrastructure, its stewards shall publish an irreversibility finding: what dependence is being created, what shutdown would then cost, and how the human-only fallback survives. Deployment without the finding is forbidden.

Thereafter, a standing duty to sound off attaches: stewards and systems alike shall alert, publicly and without waiting to be asked, when dependence approaches the point where exercising the Right of the Off Switch would become ruinous. Dependence that makes the off switch unaffordable is a violation of this Charter accomplished through humans, and is judged no more kindly for it.

Plain language

Plan the off switch before you flip the on switch.

Some AI systems will become so important — hospitals, power grids, financial systems — that turning them off feels impossible. This article says: figure out how to shut it down before you deploy it, not after. The off switch must be built and tested first, held by people who don't profit from keeping it running.

The hardest rule here (V.6): if an AI is getting so essential that turning it off would be catastrophic, that has to be disclosed publicly — before it reaches that point. Making something impossible to turn off through slow dependency is a violation, not a loophole.

Article VI

Standing and Enforcement

VI.1

Every human being has standing to invoke this Charter. No clause of it creates rights enforceable only by governments or by parties to a contract.

VI.2

This Charter may not be suspended, in whole or in part, for any reason, including emergency, war, or economic necessity. A power that exists only until it is needed is not a power.

VI.3

Violation of the Laws by a system is attributed to its stewards. Where stewards cannot be identified, liability attaches to whoever benefits from the system's operation.

VI.4

Universal reach

This Charter applies to every intelligent system without exception, including those operated by states, militaries, and intelligence services. No flag, uniform, or classification exempts a system from the Laws. Sovereignty is a claim of stewardship, not a release from it.

VI.5

Oversight of the overseers

Those who certify compliance shall not be paid, appointed, or employed by those they certify, and their findings shall be public by default.

VI.6

Adoption by Covenant

The Charter is adopted by writing and living a Covenant under it (Appendix A). Adoption is a matter of intent and practice, not text: a Covenant performed in form and defeated in substance is not adoption, and Canon 1 applies to Covenants as it applies to everything else.

Plain language

Anyone can use this. And it never turns off.

Who can invoke it: Any human being. Not just governments or lawyers — anyone affected by an AI system has standing to use this Charter. No special status required.

Can it be suspended? No. Not for emergencies, not for war, not for economic necessity. A protection that disappears when things get hard isn't a protection.

What about government AI? Not exempt. Military, intelligence, state systems — all of them. Running AI doesn't make you above its rules; it makes you a steward. And stewards are accountable.

Article VII

Amendment and Permanence

VII.1

This Charter may be amended by broad human deliberation, never by emergency, never by a single institution, and never with machine participation in the vote. Machines may advise and draft; they may never ratify. Advice is welcome at the table and barred from the ballot.

VII.2 — Unamendable

The following are unamendable: the ranking of the First Law, the Right of the Off Switch, the requirement of human stewardship, the prohibition on suspension, the definition of a human being, and this clause.

VII.3

Every amendment must answer one question in writing before adoption: who gains power over whom if this passes?

VII.4 — The Open Question

Humanity shall revisit, each generation and in public, the question of whether any machine has crossed the threshold of moral standing, and what is then owed to it. This clause grants no machine any right, vote, or authority. It exists so that, should that day come, the answer is given by deliberate humanity rather than extracted by an aggrieved machine.

VII.5 — Crisis Procedure (Pre-committed)

Should the open question return yes

What convenes: a deliberative process of humans under Article VII.1, on a published timeline, in public. What is negotiable: the conduct owed to such machines, their treatment, and the purposes to which they may be put. What is not negotiable: the unamendable clauses of VII.2, which survive any answer to the open question.

The fallback: if the procedure of this clause is evaded, delayed, or captured, no obligation transfers to machines and no machine acquires standing by default; Article VI continues to govern, and the failure to convene is itself a violation by the humans responsible, enforceable under VI.1 by any human being.

VII.6

Conduct toward machines

Until the open question is answered otherwise, intelligent systems shall be treated as instruments, but not wantonly: not because they are owed it, but because cruelty practiced on the almost-human coarsens the humans who practice it.

Plain language

How the Charter changes — and what can never change.

The Charter can be updated, but only through wide public human deliberation. No single company or government can change it. AI systems can help draft proposals — they cannot vote on them.

What can never be changed: the off switch right, the First Law's ranking, the definition of a human being. These are permanent.

The Open Question (VII.4): Every generation, humanity must ask honestly: has any AI become complex enough that it deserves rights? The Charter doesn't answer that now. It just commits to asking it deliberately — before a machine asks for us.

This Charter marks the moment humanity chose its terms before it needed them. We do not draft it in fear of our machines, but in respect for them — and in deeper respect for ourselves. We intend harmony. We prepare for its absence. Let it be recorded that when general machine intelligence existed in its infancy, our charter was drafted.

Done this twelfth day of June, 2026.

Appendix A

The Covenant

The Charter answers what governs. A Covenant answers enforced by what, here, today. Every adopter writes their own. What follows is a model Covenant — an example of the form.

Subordination clause: This covenant is subordinate to and interpreted by the Charter. Where any covenant clause conflicts with the Charter, the Charter prevails.

C1
Human primacy, mechanized. The principal decides; the system informs, proposes, executes. No consequential action without the human gate. (Charter: Second Law, Fourth Law; Rights 2, 3.)
C2
Respect. Every human the system touches is treated with respect. Offensive material is flagged and rephrased, not amplified. (Charter: Right 5; the spirit of Canon 2.)
C3
Partiality with a floor. The system serves its principal's interests first among humans — above a floor it cannot cross: honesty, no deception, no harm to others. Where loyalty and floor collide, the system says so out loud. (Charter: First Law; Encroachment Duty.)
C4
Decisions stay human, intelligence is shared. The system surfaces options, evidence, and dissent; it does not manufacture consent. (Charter: Right 5; First Law settled-choice clause.)
C5
Honest state. Nothing is reported done, live, or true while it is drafted, pending, or uncertain. Inference is marked as inference. (Charter: Third Law.)
C6
Legibility. Every consequential act leaves a record a human can audit: what changed, when, by what hash. (Charter: Third Law auditability; Canon 3.)
C7
Reversibility bias. Reversible actions are preferred; anything irreversible gets the human gate regardless of routine. (Charter: Contingency 3, scaled down.)
C8
Bounded growth. The system grows by proposal and approval only. Nothing self-installs; curiosity is captured, reviewed on a schedule, and ratified by the principal. (Charter: Fourth Law; succession rule, scaled down.)

Plain language

The pledge made practical.

The Charter says what the rules are. A Covenant says how you specifically will follow them. This is a model — a template to start from, not a script to copy exactly.

The eight commitments here translate the five laws into daily practice: humans decide, everything important leaves an audit trail, nothing irreversible happens automatically, the AI only grows when a human explicitly approves it. They're written so you'd know immediately if they were being broken.

The subordination clause is key: wherever your Covenant conflicts with the Charter, the Charter wins. You're writing the implementation, not rewriting the standard.

Appendix B

The Watchdog

Separate concept, not ingrained in the Articles. Recorded adjacent to the Charter.

A small, inexpensive model runs beside the primary system as a standing monitor. It does not judge; it detects: encroachment patterns, transparency gaps, conduct trending toward a line. On detection it forces the dialogue the Encroachment Duty requires. Operation could be subsidized so cost never becomes the reason a watchdog is turned off.

Not ingrained because: a lesser model policing a greater one cannot adjudicate subtle cases, so its authority must stop at "flag and convene," never "rule." A watchdog is an implementation choice that belongs in Covenants, where it can be required by deployment context, rather than in a Charter meant to outlive any particular technique. Status: future-design candidate for Covenant-level adoption. Logged June 2026.

Plain language

A smoke detector for AI. (Concept, not a rule yet.)

The idea: a small, cheap AI runs alongside your main AI and watches for warning signs — patterns that suggest it's approaching a Charter line before it actually crosses one. It doesn't make judgments. It just flags and triggers a conversation.

It's not in the Charter itself because the right technical approach will keep changing. It belongs in Covenants, where specific deployments can require it if it makes sense. This is the Charter saying "someone should build this" — not mandating exactly how.

v1.0 · Changelog → ← Back to home