Intelligence is becoming abundant.
For most of history, advanced cognition was scarce. It lived inside people, institutions, and a limited number of technical systems. That scarcity shaped economic power, national capacity, professional authority, and the practical limits of automation. Machine intelligence is changing those conditions. It is becoming cheaper to reproduce, easier to distribute, more capable of operating across networks, and increasingly able to act through software, infrastructure, and machines.
The open-versus-closed debate remains important. Open weights can expand access, inspectability, modification, distribution, research, and sovereign deployment. Closed systems can concentrate capability and control inside a smaller number of institutions. These choices affect innovation, competition, security, and who can build with advanced intelligence.
But access is not the same question as authority.
Open weights expand who can access intelligence. The custody question determines who may deploy it, delegate through it, constrain it, revoke its authority, and remain accountable for its actions. A model may be open or closed while the systems built around it exercise very different powers. The decisive governance issue is therefore not only who possesses intelligence, but who is permitted to authorize consequential action through it.
That distinction becomes more urgent as systems become autonomous, distributed, and embodied. An assistant that drafts text does not exercise the same authority as an agent that moves money, changes software, allocates resources, directs machines, controls infrastructure, or makes decisions that affect rights and livelihoods. A system operating across jurisdictions, organizations, devices, and physical environments can create consequences faster than conventional supervision, contracting, and enforcement can respond.
Delegation must therefore be explicit. Institutions should be able to state what authority a system has, what it does not have, who granted that authority, which evidence justified the grant, how the authority is monitored, and how it can be suspended or revoked. Traceability should follow the chain of delegation rather than ending at a model provider or disappearing inside a technical stack.
Accountability must also follow authority and benefit. The people and institutions that authorize, deploy, and profit from consequential systems should not be able to transfer operational power to machines while leaving responsibility behind. Where benefits are captured, liability cannot be externalized. Where authority is delegated, the ability and duty to intervene must remain real.
The Custody of Intelligence proposes an Intelligence Stewardship Framework for examining this layer. It separates capability from custody and consequence; describes Intelligence Stewardship Levels for different degrees of delegated authority; and proposes Prohibited Authority Thresholds for powers that should not be autonomously exercised merely because a system is capable of exercising them. These ideas are proposals, not an adopted standard. The full monograph is a working draft and has not been peer-reviewed.
The framework also does not settle the open-weights question. It is intended as a complementary next question. Open models may strengthen innovation, inspection, resilience, and sovereignty. They may also be deployed under strong or weak custody arrangements. Closed systems may be centrally controlled yet still receive excessive or poorly governed authority. Custody is therefore not an argument against openness, sovereign AI, or accelerated deployment. It is an attempt to specify the accountability architecture that must accompany them.
Serious governance will require disagreement about where lines should be drawn. Which powers require permanent human authorization? What forms of revocation are technically and institutionally credible? How should responsibility be assigned when systems act through multiple organizations and jurisdictions? Which evidence should be required before an autonomous system receives authority over money, infrastructure, weapons, employment, health, identity, or civil rights? Where does the proposed framework fail, duplicate existing concepts, or create false confidence?
This statement is an invitation to test those questions. Participation in discussing, reviewing, challenging, or signing the narrow proposition below does not imply endorsement of every claim in the monograph, its terminology, or its proposed framework. Public attribution requires explicit permission and an exact description of the person’s or institution’s participation.
“As intelligence becomes abundant, autonomous, distributed, and embodied, governance must address not only access to models, but authority over their deployment, delegation, constraint, and accountability.”
The full monograph provides the argument, evidence, limitations, and proposed operational framework in greater detail. Readers are invited to identify where it fails.