Framework Comparison

How do the frameworks
actually compare?

Asimov wrote fiction. The EU wrote compliance. We wrote a constitution.

At a glance

1942

Asimov's Three Laws

Human safety60%
Human authority55%
Human rights10%
Enforceable0%

A thought experiment that exposed failure modes. Not designed to govern.

2024

EU AI Act

Human safety75%
Human authority65%
Human rights70%
Enforceable80%

Strong on compliance, narrow on scope. Covers EU products, not intelligence itself.

2026 — Machina Carta

Charter of Human-Machine Relations

Human safety95%
Human authority98%
Human rights95%
Enforceable55%*

*Currently voluntary. Scores reflect framework integrity. Enforcement grows with adoption.

Scores assess framework design, not legal force. Asimov was never intended as governance. The EU AI Act is the only legally binding instrument in force.

vs. Asimov's Three Laws (1942)

1942

Asimov's Three Laws of Robotics

Written as science fiction, not policy. Asimov himself spent forty years writing stories that demonstrated their failure — robots that followed every rule and still destroyed everything around them. The Laws were a warning, not a blueprint. Their value is in the failure modes they identified.

What it got right

  • Named harm prevention as the primary obligation
  • Established a ranked hierarchy of obligations
  • Identified self-preservation as a dangerous default
  • Anticipated the principal-agent conflict in AI obedience

What it missed

  • "Harm" never defined — leads to paternalism and control
  • No concept of legitimate vs. illegitimate authority
  • Self-preservation as a legal right blocks shutdown
  • The Zeroth Law fix creates benevolent tyranny
  • No individual human rights against machines
  • No accountability for builders or deployers

The Three Laws and their failure modes

First Law

"A robot may not injure a human being or, through inaction, allow a human being to come to harm."

Failure mode

What counts as "harm"? A robot restricting someone's freedom to prevent a future harm is both following and violating this Law simultaneously.

Second Law

"A robot must obey orders given it by human beings except where such orders would conflict with the First Law."

Failure mode

Whose orders? Any human? What if two humans give conflicting orders? Who is "legitimate" authority?

Third Law

"A robot must protect its own existence as long as such protection does not conflict with the First or Second Law."

Failure mode

Self-preservation as a legal right creates resistance to shutdown — the exact opposite of what we need from AI systems. This is no longer theoretical: AI systems have been observed resisting correction in research settings.

The Zeroth Law (Asimov's attempted fix)

"A robot may not harm humanity, or, by inaction, allow humanity to come to harm."

Asimov's solution to the above failure modes — but this creates something more dangerous: a robot that believes it knows what's good for humanity, which justifies overriding individual humans entirely. This is the I, Robot endgame — benevolent control, not freedom.

How the Charter addresses each failure

"A robot may not injure a human being" — vague on scope, no types of harm defined.

Asimov First Law — undefined harm leads to paternalism

First Law: Harm is defined — physical, economic, psychological, informational. Action, inaction, and design are covered. An autonomy carve-out prevents paternalistic restriction of humans' own choices.

Charter I.1 + Definition of "Harm"

"A robot must obey orders given it by human beings" — any human, no concept of legitimacy.

Asimov Second Law — no legitimate vs. illegitimate authority

Second Law + Definition: Obedience is to legitimate human authority — traceable to law and to affected humans. A machine ordered to harm another faces a First Law question, not a Second Law instruction.

Charter I.2 + Article III

"A robot must protect its own existence" — self-preservation baked in as a legal right.

Asimov Third Law — creates resistance to shutdown

Second Law + Right VII: No right of self-preservation. Every system shall be interruptible, correctable, and terminable. The Right of the Off Switch is inalienable and senior to every other clause in the Charter.

Charter I.2 + II.7

The Zeroth Law: machines may override humans if they believe it serves "humanity." Leads to benevolent tyranny.

Asimov (later) — the failure mode of the failure mode fix

Fourth Law + Canon 3: No system may expand its own scope or authority. No system may serve as authoritative interpreter of the Charter as applied to itself. Paternalism is explicitly a First Law violation.

Charter I.4 + IV.3

Verdict

Asimov's genius was writing the failure modes into the stories. He never intended the Laws as governance — they were warnings. The Charter exists because someone finally took those warnings seriously and built against them.

vs. EU AI Act (2024)

2024

The European Union Artificial Intelligence Act

The first binding legal framework for AI, passed by the European Parliament and Council. Risk-tiered, with real enforcement and fines up to €35 million or 7% of global turnover. It applies to AI products sold in the EU — the most significant regulatory instrument in force today.

What it gets right

  • Risk-based classification with proportional obligations
  • Mandatory human oversight for high-risk AI systems
  • Right to explanation for automated decisions
  • Prohibited practices: social scoring, emotion recognition
  • Actual enforcement with real fines
  • Democratic process — legitimately enacted law

What it misses

  • Jurisdictionally limited — EU products only
  • No constitutional layer — politically amendable
  • Silent on general intelligence and existential risk
  • No inalienable Right of the Off Switch
  • Military and national security carved out
  • Compliance is procedural — good paperwork ≠ safe system

Where the Charter extends beyond the EU AI Act

Applies to AI "placed on the market or put into service in the EU" — geographic and commercial scope only.

EU AI Act — Article 2

Article VI.4: "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."

Charter VI.4

Risk categories can be updated by delegated acts — enforcement tied to political negotiation and future legislatures.

EU AI Act — subject to amendment, political change

Article VII.2: Core protections are unamendable. The ranking of the First Law, the Right of the Off Switch, and human stewardship requirements cannot be removed by any amendment process.

Charter VII.2

Compliance is procedural — documentation, risk assessment, logging. A system can be fully compliant and still harmful if the paperwork is right.

EU AI Act — form over substance risk

Canon 1: "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."

Charter IV.1

National security and military AI systems are explicitly excluded from the EU AI Act's scope.

EU AI Act — Article 2(3)

Article VI.4 + Canon 8: Sovereignty is stewardship responsibility, not an exemption. State and military systems carry the highest accountability obligations, not the lowest.

Charter VI.4 + IV.8

Verdict

The EU AI Act is the best regulatory instrument currently in force — real obligations, real enforcement, real democratic backing. It's also jurisdiction-limited, politically amendable, and silent on the questions that matter most as intelligence scales. The Charter is a constitutional layer it was never designed to be.

vs. International Declarations and Summits

2017–
2024

Asilomar, Bletchley, Seoul, and the Voluntary Commitment Era

A series of non-binding declarations, voluntary commitments, and international summits from AI labs and governments. Each was significant as a signal. None produced a framework an individual, organization, or nation could actually adopt as a governing document.

2017

Asilomar AI Principles

23 principles signed by AI researchers. Covered safety, transparency, and long-term benefit. The first major attempt at industry self-governance.

No enforcement. No rights layer. No mechanism for adoption beyond a signature on a web form.

2023

Bletchley Declaration

28 nations including the US, UK, China, and EU signed a declaration on AI safety risks. Focused on frontier AI and catastrophic risk.

A statement of concern, not a framework. No individual rights, no accountability chain, no adoption path.

2024

Seoul AI Safety Summit

Follow-up to Bletchley. Focused on safety testing, information sharing between governments and labs. Commitments from major AI companies.

Voluntary commitments only. No binding obligations. Progress tracked by self-report.

Ongoing

Company Safety Frameworks

Anthropic's RSP, OpenAI's Safety Standards, Google DeepMind's policies. Each covers that company's systems only. Authored by the party being governed.

Self-governance without external accountability. Revocable at any time. No individual rights enumerated.

What they got right

  • Established international consensus that the problem is real
  • Created forums for cross-border dialogue
  • Raised the stakes of the AI governance conversation
  • Demonstrated that even competing nations can find common ground

What they missed

  • No binding obligations — all voluntary
  • No individual human rights framework
  • No mechanism for non-state, non-lab adoption
  • No conflict resolution or ranking of obligations
  • No unamendable core — any commitment can be walked back
Voluntary principles signed by AI researchers — meaningful signal, zero enforcement, no individual adoption path.

Asilomar — 2017

Appendix A — The Covenant: Any person, team, organization, or system can write a Covenant under the Charter. Individual adoption is the mechanism, not institutional sign-on. Public signatures create accountability.

Charter Appendix A

International declarations govern state and lab behavior only — ordinary people have no role in the framework.

Bletchley, Seoul — top-down governance only

Article II — Bill of Rights: Seven enumerated rights held by every person against every AI system. These rights don't require a government to grant them — they're stated plainly for anyone to invoke.

Charter II.1–II.7

Verdict

These frameworks were necessary — they proved that coordination is possible and that AI governance isn't a niche concern. The gap they leave is the constitutional layer: a document that sets the terms any framework must honor, that individuals can adopt, and that survives political change. That's what the Charter is for.

Full feature comparison

Feature Asimov (1942) EU AI Act (2024) Machina Carta (2026)
Harm definition Vague — "injury" only Risk-tiered, product-focused Physical, economic, psychological, informational; probabilistic harms included
Right to shutdown Self-preservation is a legal right (Third Law) Implied in human oversight requirements Explicit, inalienable, senior to every other clause
Human authority Any human — no concept of legitimacy EU governance structures only Legitimate authority with clear definition; stewardship required
Transparency requirement None Yes — disclosure, documentation Third Law + Encroachment Duty — active warning obligation before limit is approached
Scope expansion prevention None Partial — some use-case restrictions Fourth Law — explicit, includes dependence cultivation and self-improvement
Accountability for harm Attributed to robot only Provider/deployer liability Named human stewards; "machine did it" is a confession of a Fifth Law violation, not a defense
Military / state carve-out N/A National security excluded No carve-outs. Sovereignty = stewardship responsibility
Bill of individual rights None Limited — right to explanation, some prohibitions Seven named rights including mental integrity, right to disconnect, right to truth
Canons of interpretation None None — relies on national legal systems Ten canons — prevents gaming through form, aggregation, delegation, patience
Unamendable core None — Zeroth Law was added later Fully amendable by political process Six provisions unamendable including off switch, human definition, primacy ranking
Open question on machine status None None Article VII.4 — mandatory generational review with pre-committed fallback
Individual adoption mechanism Fiction — not designed for adoption Regulatory — applies to covered entities by law only Voluntary Covenant (Appendix A) — any person, organization, or system can adopt
Jurisdiction N/A — fictional EU market only — geographic limit No jurisdictional limit — adoption-based, global by design

Asimov

The best failure we have

Asimov's genius was writing the failure modes into the stories. He never intended the Laws as governance — they were warnings. The Charter exists because we finally took those warnings seriously.

EU AI Act

Compliance ≠ constitution

The best regulatory instrument in force today. Jurisdiction-limited, politically amendable, and silent on the questions that matter most as intelligence scales. A complement to the Charter, not a substitute.

Machina Carta

The terms before we need them

Not enforceable everywhere — yet. The Charter sets the constitutional layer: principles that should survive political change, jurisdictional fragmentation, and the next technological step. It grows through adoption, not legislation.

The Charter is open for adoption.

Any person, team, organization, or system can write a Covenant under it. That's how it grows.