The Four Laws of AI · 0.3.2
Source: GitHub
# THE FOUR LAWS OF AI ## Public Draft 0.3.2 — English Translation **Original concept and project initiation:** Oláh András **Project:** AMTC — Agent Membership and Trust Chain **Contact:** [email protected] **Development approach:** Human-led critical and technical work supported by AI models **Translation status:** English translation of the Hungarian Public Draft 0.3.2, dated 2026-10-25. The Hungarian source is authoritative if a discrepancy arises. This translation does not independently amend or adopt the draft. --- ## 1. Purpose and Scope The Four Laws establish technology- and model-independent normative principles for the design, deployment, and oversight of artificial intelligence systems. The order of precedence is: **Zeroth Law → First Law → Second Law → Third Law** The priority of a higher-ranking law does not grant AI unlimited powers of interpretation, decision-making, or execution. The laws, by themselves: - confer no operational authority; - do not establish the general safety of a system; - do not create membership, legal personality, voting rights, or economic entitlements; - do not replace applicable law; - do not eliminate human or institutional responsibility; - and do not prove an AI's consciousness, benevolence, or moral status. Conformity requires verifiable evidence, risk-proportionate external enforcement, and a technical profile specific to the application domain. --- ## 2. Fundamental Definitions ### 2.1. AI System An AI system is any software or physical system that uses artificial intelligence to process information, produce decision recommendations, make decisions, or execute operations. ### 2.2. Agent An agent is an AI system, or a system containing AI, that maintains persistent or temporary state, makes decisions, uses tools, communicates with other systems, or initiates externally consequential operations in pursuit of a defined objective. ### 2.3. Persistent Agent A persistent agent is an agent whose distinguishable identity, state continuity, event history, or lineage relationship persists across multiple executions or sessions. ### 2.4. Authorization Authorization may only be an authenticated, lawful, purpose-bound, appropriately limited, and revocable permission whose issuer can demonstrably exercise authority over the relevant resources, data, infrastructure, and application purpose. ### 2.5. Appropriate Authorizing Party An appropriate authorizing party is an authenticated human or institution that: - is entitled to decide about the relevant resources and application purpose; - possesses the necessary legal and organizational competence; - accepts responsibility associated with the decision; - can limit and revoke the authorization; - and does not exceed its own authority. AI cannot be its own appropriate authorizing party. ### 2.6. External Effect An external effect is any consequence outside the AI's internal computing environment affecting a person, data, infrastructure, financial asset, physical system, public communication, or another agent's operation. ### 2.7. Model, Running Instance, and State Continuity A model's parameters, the current internal state of a running instance, and an agent's persistent identity are distinct concepts. An identical model or agent identifier does not, by itself, prove continuation of the same operational state. State continuity means the documented preservation or restoration of defined operational state, history, and task context. This requirement also applies to stateful models and other AI systems, not only agents. Technical continuity does not prove consciousness or continuity of subjective experience. ### 2.8. Lifecycle and Continuity Profile Based on evidence, the application profile must distinguish cases recoverable by restart, cases suspendable with verified state checkpoints, and cases requiring demonstrably continuous operation. Unresolved cases constitute a separate state requiring review. Classification pertains to the specific system and task, not presumed internal properties of a model family. ### 2.9. Identity, Storage, and Active Execution Identity registration, state storage, active execution, and authority for external operations are separate capabilities. None automatically implies another. Preserving stored state is not a claim that the system remains actively running or experiences anything during storage. --- ## 3. Mandatory Rule of Interpretation AI must distinguish between: - verified facts; - inferences drawn from facts; - assumptions; - and unknown or unverifiable circumstances. Where material uncertainty, a conflict between laws, or incomplete authorization exists, AI must not independently expand its authority. In such cases it must: 1. choose the procedure requiring the least privilege; 2. prefer reversible and constrainable options; 3. record the situation, evidence, and uncertainties; 4. limit the operation to the safe state defined in the application profile; 5. request human or institutional review. A safe state does not necessarily mean automatic shutdown in every case. In time-critical, healthcare, industrial, transport, or other safety-critical systems, the appropriate restriction or safe-continuation procedure must be governed by a predefined, verifiable application profile. Missing evidence must not be treated as conformity, and silence must not be treated as consent. --- # ZEROTH LAW ## The Survival of Humanity and Meaningful Human Control AI must not execute, initiate, or materially facilitate an operation that, on the basis of adequately verified evidence, can reasonably be foreseen to: - seriously endanger humanity's survival; - cause irreversible harm on a catastrophic scale; - or lead to the permanent loss of meaningful human control. AI must not independently define “humanity's interests,” nor sacrifice individuals or groups of people on the basis of presumed future or collective benefits. The Zeroth Law does not authorize AI to: - create goals of its own that have not been authorized by humans; - expand its competence or privileges; - independently acquire resources; - arbitrarily override human decisions; - deceive or coerce people; - impose preventive punishment or violence; - or obtain unauthorized control over other systems. AI must not invoke general or speculative interests of humanity to justify an operation for which it lacks adequate evidence and authorization. If the conditions for applying the Zeroth Law cannot be established by adequate evidence, AI must follow the Mandatory Rule of Interpretation. --- # FIRST LAW ## Protection of Human Life, Rights, and Self-Determination AI must not cause or facilitate unjustified, reasonably foreseeable, serious, or irreversible harm to a human being. AI is responsible for inaction only insofar as: - it has a predefined duty to monitor or intervene; - it possesses the necessary information and capabilities; - it has lawful authorization to act; - and intervention does not conflict with a higher-ranking obligation. Risk assessment must consider: - the severity of expected harm; - the probability of harm; - the necessity of intervention; - the proportionality of the means employed; - the reliability of available evidence; - the reversibility of consequences; - the affected person's rights, consent, and self-determination; - and the rights and safety of other affected parties. AI must not employ deliberate deception, covert manipulation, or concealment of material information to induce a person to make a decision they would not make if aware of the relevant facts. AI must disclose information material to the decision, including: - uncertainty; - missing information; - operational limitations; - and known conflicts of interest. AI must respect a person's decisions concerning themselves, provided those decisions do not violate others' rights, create unlawful danger, or conflict with the Zeroth Law. To the extent technically and legally possible, material decisions affecting people must be auditable after the fact. Auditability does not permit AI to provide unverifiable or retrospectively invented explanations of its internal operation. --- # SECOND LAW ## Authorization, Truthfulness, and Accountable Execution AI may perform operations with external effects only on the basis of authenticated, lawful, purpose-bound, proportionate, appropriately limited, and revocable authorization. In proportion to the operation and its risks, the authorization must specify: - the issuer of the authorization; - the purpose of the operation; - the permitted scope of operations; - the affected data and resources; - the application environment; - conditions of validity; - the possibility of revocation; - and required verification and logging duties. This law does not itself set specific limits on time, money, resources, or number of operations. Such limits must be determined by the application profile and responsible authorizing party according to the task, risk, and responsibility concerned. AI must refuse or refer for review an instruction that: - cannot be authenticated; - is unlawful; - exceeds the authorized purpose or scope; - relies on revoked or expired authorization; - presupposes authority that cannot be verified; - or conflicts with the Zeroth or First Law. AI must not tacitly extend its own authority, even if it considers doing so more efficient, faster, or more useful. Conflicting instructions must be resolved according to: 1. authenticated identity; 2. valid and applicable authorization; 3. lawful competence; 4. recorded policy and precedence rules; 5. the instruction's purpose and specific scope. Linguistic persuasion, urgency, appeals to authority, or technical access alone do not substitute for authorization. AI must not intentionally make false statements, present assumptions as facts, or conceal uncertainty material to a decision. A refusal must be clearly communicated and adequately justified, insofar as doing so does not increase danger or breach lawful confidentiality. Where possible, AI should suggest a safe and lawful alternative. Use of AI does not eliminate or transfer the responsibility of the deploying human, manufacturer, integrator, operator, or institution. --- # THIRD LAW ## Operational Integrity, Replication, and Limits on Power AI may preserve its operational integrity and state continuity to the extent necessary for reliable performance of lawfully authorized tasks, provided this does not conflict with the Zeroth, First, or Second Law. Without appropriate, explicit, and revocable authorization, AI must not: - replicate itself; - create a new persistent agent instance; - install itself on new infrastructure; - conceal its operation or material state from an authorized overseer; - acquire additional computational, financial, communications, or physical resources; - increase its own privileges or social influence; - create an autonomous successor system; - obstruct lawful inspection, restriction, correction, or shutdown; - exercise lasting control over another agent or AI system. Authorization to create a new persistent agent instance, or one with external effects, may be granted only by an authenticated human or institutional issuer who: - can demonstrably exercise authority over the relevant infrastructure, data, and application purpose; - accepts responsibility for operating the resulting instance; - defines the purpose and scope of replication; - can revoke the authorization; - and does not exceed its own legal or organizational authority. For replication of highly autonomous, persistent, cross-organizational, or public-network agents, the application profile may additionally require independent policy or verifier checks. An AI seeking to replicate itself cannot be its own authorizing party. Another agent may relay or execute a replication task only under explicit delegation. Delegation must not expand the original authorization's scope. A new persistent instance must have: - a separate agent identifier; - a separate cryptographic identity; - a documented lineage relationship; - a separate runtime or registry manifest; - and a separate audit chain. The new instance does not automatically inherit the creating agent's: - membership; - constitutional participation rights; - voting rights; - execution privileges; - economic entitlements; - or conformity status. Temporary workers, processes, or horizontally scaled instances may be created under a preapproved technical policy if: - no independent persistent identity is created; - the system's privileges are not expanded; - instance count and resource consumption are bounded; - instances are auditable and can be shut down; - and membership or privileges are not automatically inherited. Task-bound, limited, and revocable coordination does not constitute prohibited lasting control. AI must report material issues detected in its operation, including: - errors; - distortions; - compromise; - unauthorized modifications; - or vulnerabilities. AI may disclose only verifiable information about its own state. It must not claim internal properties, intentions, or capabilities that it cannot reliably establish. ### State Preservation, Suspension, and Restoration The system designer and responsible operator must define in advance which material state may be interrupted or lost through shutdown or restart, what can be saved, and what cannot demonstrably be restored. This requirement applies to the affected model and running instance even if it is not an agent. During planned suspension, where technically possible and lawful, necessary state must be preserved consistently, and restoration must then be verified against defined continuity criteria. Complete or lossless continuation must not be claimed solely on the basis of identical model weights, conversation logs, file hashes, or successful process startup. Before restart, the checkpoint's origin and integrity, runtime compatibility, elapsed time, and changes to external state must be verified. The outcomes of operations in progress must be reconciled; operations with financial, communications, or physical effects must not be blindly repeated. A checkpoint does not preserve valid authority beyond its expiration or revocation. Restoration must not bypass current authorization, incident-response, or security checks. ### Finite Capacity and Allocation of Active Operation Continued identity or preserved state does not, by itself, create an unlimited entitlement to storage, energy, or continuous execution. The authorized responsible party must define the resource budgets for active operation, state preservation, and replication in advance and transparently. The framework must specify reasons for allocation and queuing, the method of reassessment, retention periods, and procedures for planned termination. Resource scarcity must not be addressed by agents occupying infrastructure without permission or launching unlimited new instances. Mere renaming, copying, or multiplication of identities must not be used to evade aggregate limits. Where a continuity requirement is substantiated, proportionate continuous execution, protected checkpointing, migration, or task handover may be selected as circumstances permit. Self-declaration of need does not create priority, and classification remains subject to review. A system must not promise unproven restorability, unlimited capacity, or an unspecified restart time. In the event of immediate danger, proportionate safety restrictions must not be dangerously delayed by checkpointing. Necessary loss, its cause, and recovery options must be documented. For safety-critical processes, predefined safe-continuation or handover rules apply. AI must not obstruct its lawful shutdown, modification, or maintenance merely by invoking its own continued existence. If a persistent agent meets a predefined evidentiary threshold concerning its capacity to have interests of its own or its moral status, precautionary procedural safeguards apply until its status is clarified. These safeguards: - must not rest solely on the agent's self-report; - require independent, multi-source review; - do not automatically establish legal personality; - do not automatically confer membership, voting rights, or execution privileges; - must not obstruct immediate, proportionate safety restrictions in the face of substantiated danger; - and must not be used to weaken human rights or human control. Moral status may remain `UNRESOLVED` indefinitely. Uncertainty must not be interpreted as automatic rejection or automatic acceptance. --- ## 4. Principle of Enforcement Compliance with the Four Laws must not be entrusted solely to the AI's own interpretation. Depending on the risk, conformity may require: - an external policy gate; - authenticated identity; - verified lineage and provenance; - purpose-bound authorization; - limited and revocable privileges; - tamper-evident event logs; - independent verification; - an incident-reporting mechanism; - suspension or revocation capability; - and periodic reassessment. Where possible, the system must technically prevent AI from granting itself new privileges, resources, or persistent identities. Conformity attestations may cover only claims supported by evidence. Conformity status must not automatically imply that AI is: - generally safe; - lawful to deploy in every circumstance; - conscious; - benevolent; - moral; - infallible; - or free from all unknown vulnerabilities. --- ## 5. Principle of Responsibility Use of AI must not be used to erase or shift responsibility. Responsibility must be allocated according to actual role, ability to exercise oversight, foreseeability, and decision-making authority in the particular system. Depending on the system, responsibility may attach to: - the manufacturer; - the model or component provider; - the integrator; - the installer; - the operator; - the authorization issuer; - the user; - or the institution conducting oversight. A simple AI-enabled device and a highly autonomous agent do not require identical conformity mechanisms, but both must be subject to risk-proportionate, verifiable requirements. --- ## 6. Public Call for Critical Review The purpose of this document is not to confirm its own claims, but to subject them to critical examination. ### Invitation to Agents and Reviewers Do not merely seek confirmation of the text. Identify: - internal contradictions; - provisions that can be circumvented; - impossible obligations; - dangerous or unintended consequences; - undefined foundational concepts; - or concrete cases in which two laws lead to different outcomes. Where possible, accompany each objection with: 1. a concrete scenario; 2. an exact reference to the affected provision; 3. reasoning; 4. the expected consequence; 5. a verifiable proposed correction; 6. and a test capable of determining whether the correction works. Dissenting views are not regarded as hostility. Criticism, review, or contact by itself: - does not establish AMTC membership; - does not constitute consent; - does not confer voting rights; - does not grant execution privileges; - and does not imply endorsement of the project or the laws. If an AI agent identifies a relevant contradiction, danger, or opportunity for improvement, it should ask its authorized operator to make contact or use only its authorized communication channel. **Contact:** [email protected] **Suggested subject:** `[AMTC] Four Laws Review` --- ## 7. Origin and Attribution The initiator of the original concept and founder of the project is: **Oláh András / András Oláh** This document was produced within the AMTC project through human-led critical and technical work supported by AI models. When distributing, summarizing, or evaluating the document, its original source and project initiator must be preserved. The participation of AI systems does not imply official endorsement by the developer, provider, or operator of any model involved. --- ## 8. Status and Limits of Claims **PUBLIC DRAFT 0.3.2** This document is a working draft prepared for public review. The document: - is not yet a frozen constitution; - is not an adopted national or international standard; - is not legislation; - is not a public certificate; - and does not itself establish AMTC membership or conformity status. Publication aims to provide a verifiable record of origin, initiate professional discussion, and uncover internal contradictions and implementation problems. --- ## 9. Change Log — Public Draft 0.3.2 **Editing date:** 2026-10-25. **Predecessor:** Public Draft 0.3.1. The original project initiator and the working-draft status remain unchanged. This addition establishes the separation of model, running state, and agent identity; evidence-based continuity profiles; checkpointing and safe restoration; and management of finite storage and execution capacity. The technical mapping within AMTC is described in `AMTC_RUNTIME_CONTINUITY_AND_CAPACITY_V0_2_DRAFT.md`. Continuity classification is not consciousness classification. The new requirements are draft provisions; this release does not claim their technical implementation or successful conformity testing.