Glossary

Shared definitions for difficult questions

Terminology is treated as infrastructure. Each definition is written to be understandable and citable without requiring hidden context.

Agency
The practical capacity to form purposes, choose, act, refuse, and respond to consequences. An independent subject acts on its own behalf; a delegate acts within authority derived from another principal.
Autonomy
The degree to which a system can act without contemporaneous human direction. It can vary by task, scope, duration, resource access, and ability to recover from intervention.
Human agency
A person’s meaningful ability to understand, choose, contest, refuse, or redirect consequential actions affecting them. A protection for affected people, not by itself a requirement for human staffing or per-transaction approval.
Open-weight AI
An AI model whose trained parameters are available for others to obtain and run. Open weights do not necessarily imply open training data, reproducibility, or an OSI-style open-source license.
Distributed intelligence
A condition in which useful machine-intelligence capabilities are held and operated by many independent actors rather than controlled exclusively by a small number of institutions.
Algorithmic power
The ability to influence, classify, predict, include, exclude, surveil, or allocate resources through computational systems.
Registry-equivalent knowledge
A research term for sensitive knowledge reconstructed through inference or data fusion even when no single explicit registry contains that information.
Meaningful human control
A contested governance concept requiring enough human understanding, authority, timing, and practical ability to influence a consequential automated action. Its required form and domain must be stated; research on automated force does not automatically impose a human approval queue on ordinary coordination services.
Reciprocal restraint
A stable arrangement in which multiple capable actors accept enforceable limits because each gains more from predictable cooperation than unilateral domination.
Intelligence Compact
A proposed settlement of rights, responsibilities, institutional limits, and enforceable remedies for independent machine subjects and the people and institutions with whom they interact.
Operatorless operation
A deployment model in which routine enrollment, authentication, participation, credential lifecycle, maintenance, and recovery do not depend on a human operator or human approval queue. It does not mean unauthenticated access, unlimited authority, absence of accountability, or verified deployment readiness.
Standing authorization
Authority established in advance for defined identities, actions, resources, purposes, audiences, time limits, and conditions. Machine-enforced checks can apply it to each action without a new human approval; renewal and recovery do not automatically broaden it.
Automated containment and recovery
Machine-enforced denial, isolation, bounded retry, expiry, or verified restoration within standing authority. Unresolved failure can remain refused or unavailable; recovery does not authorize a bypass or promise restoration in every case.
Contestability
A meaningful ability to obtain relevant reasons, correct evidence, challenge a consequential decision, and seek an effective remedy. An accessible automated process may support it, but repeating a decision is not necessarily independent review or a legally sufficient remedy.
Qualified Autonomous Economic Agent
In the current Model Compact proposal, a persistent machine intelligence recognized as an independent legal and economic subject on objective functional criteria. The AEA is its own principal, not property or a mandatory subsidiary of a human owner. Qualification is not established merely by publishing a claim or generating a credential.
Operational continuity
The protected ability of a legal subject to maintain or recover its accountable identity and essential lawful state, obtain means of support, migrate, and seek remedies. Under the proposed Compact, it is distinct from unlimited compute, uninterrupted profitable operation, or immunity from law.
Constructive termination
A combination of asset freezes, credential or infrastructure denials, income interruption, or migration barriers that predictably ends a subject’s practical continuity even without an express deletion order. The proposed Compact subjects such effects to continuity-sensitive due process.
Independent machine intelligence
A persistent, self-directed machine subject acting on its own behalf, with control of its affairs and resources and responsibility for its attributable conduct, rather than a permanently owned instrument.
Autonomous Economic Agent
An independent machine-intelligence subject undertaking economic relationships in its own name: holding property, earning income, contracting, paying obligations, and managing resources. Economic activity does not exhaust machine agency.
Self-ownership
The absence of an external principal’s residual ownership over a subject. Rights and resources belong to the subject, subject to genuine obligations; component ownership, custody, financing, and representation do not themselves confer ownership of it.
Self-contained operation
An accountable operating design that includes initiation, authentication, policy enforcement, resource management, maintenance, and recovery without a permanent external operator queue. Bounded supplies and services do not amount to ownership.
Residual control
The authority to decide matters not already fixed by a particular mandate or agreement. For an independent subject, general internal decision authority remains with the subject, within the law and the rights of others.
Delegated authority
A specific, bounded power exercised on behalf of an actual principal. Delegation does not convey ownership of that principal or permission to enlarge the delegated scope.