My illustration entitled: “The Unceded Crown” – A human holds a simple crown while a colossal mechanical figure kneels beside them, its open hands offering tools.
The promise of artificial intelligence is that it can enlarge what a person is able to do. It can help us examine information, coordinate activity, recognise patterns and carry out work that would otherwise exceed our time or attention. But an increase in capability does not necessarily produce an increase in freedom.
A person may become more productive while losing the ability to question the system directing their work. They may receive better recommendations while finding it harder to choose another course. They may gain access to a powerful service while surrendering control over the information, decisions and relationships through which that service operates.
The starting claim of Unceded Intelligence is that these outcomes should be distinguished. We should pursue intelligence that expands human capability while preserving legitimate human authority over its use.
That claim needs foundations. A promise that technology will remain under human control means little unless we can identify what control requires.
I propose seven principles: explicit authorization, bounded scope, meaningful understanding, contestability, revocation, practical independence, and respect for other people’s rights. Together, they describe the conditions under which delegation can remain an exercise of human sovereignty.
1. Explicit authorization: authority must be granted
A system’s ability to perform an action does not establish permission to perform it. Access is not the same as authorization, and a prediction about what someone probably wants is not equivalent to their instruction.
Consider a system that can read a person’s correspondence. Permission to identify important messages does not automatically include permission to answer them, disclose their contents, or make commitments in the person’s name.
Each additional power changes the relationship.
Explicit authorization requires a clear connection between a person’s intention and the authority exercised on their behalf. The person should be able to identify what they have permitted and distinguish it from what the system merely has the technical capacity to do.
This does not require a fresh confirmation for every trivial operation. A person can deliberately approve a standing instruction covering a class of routine activities. What matters is that the instruction is understandable, specific enough to govern conduct, and given without concealed expansion.
Silence should not become a convenient substitute for permission when a system seeks materially greater power. Nor should an institution treat permission to use a service as an unrestricted licence to direct the user’s affairs.
The first question is: who authorized this action, and what did they actually authorize?
2. Bounded scope: every delegation needs an edge
Authorization requires boundaries. “Help me” expresses a desire for assistance; it does not define unlimited authority.
A workable delegation should identify its purpose, permitted actions, relevant information, duration and conditions for escalation. Where appropriate, it should also specify spending limits, prohibited actions and whether the system may involve another service.
Suppose a person asks a shopping assistant to find a suitable laptop. Comparing products, placing an order and entering a continuing financial agreement are different powers. Permission for the first cannot silently carry permission for the others.
Scope should also constrain how a goal is pursued. An instruction to reduce costs should not be interpreted as permission to reveal confidential information or evade obligations to other people.
The boundary belongs in the operation of the system. A restriction written in a policy has little value if the software can routinely act beyond it without detection or consequence.
When a task requires broader powers, the system should return to an appropriate decision-maker. Its success within one area does not entitle it to enlarge its own jurisdiction.
The second question is: where must this system stop, even if continuing would help it achieve its objective?
3. Meaningful understanding: people must understand the bargain
A person cannot exercise useful authority over an arrangement whose material consequences are concealed from them.
Meaningful understanding does not require every user to become a mathematician or software engineer. We rely on complex tools without knowing every detail of their construction. The relevant requirement is that people understand enough about the delegation to make a considered choice.
They should know what the system is intended to do, what information it uses, whether it recommends or acts, what significant limitations apply, and what happens when something goes wrong.
They should also understand whose interests shape its operation. A recommendation selected to benefit the user and a recommendation selected to increase the provider’s revenue may look similar while serving different purposes. Material conflicts should be visible.
Understanding requires communication suited to the person and the decision. A technically accurate notice can still fail if it buries the important consequence beneath pages of irrelevant detail.
The practical standard should be proportionate. Sorting a personal reading list requires less explanation than a system making commitments with substantial financial or personal consequences.
Uncertainty should remain visible too. Presenting a recommendation confidently does not remove its assumptions or limitations.
The third question is: can the person explain, in ordinary language, what power they are granting and what its important consequences may be?
4. Contestability: disagreement must have somewhere to go
Understanding a decision does not mean agreeing with it. A system can be transparent about a conclusion and still be mistaken, unfair or outside its mandate.
People affected by consequential decisions need a meaningful way to challenge them.
Imagine an automated process that rejects an application because it relies on an incorrect record. An explanation that identifies the record is useful, but incomplete. The affected person also needs a way to correct it and obtain reconsideration.
Contestability therefore requires more than a complaints address. It requires a process capable of changing the outcome.
Someone with appropriate authority must be able to review the relevant information, recognise an error, override the result and provide a remedy where warranted. In disputes involving the operator itself, independent review may be necessary.
Timing matters. An appeal resolved after an opportunity has disappeared may offer little practical protection. Where serious consequences are difficult to reverse, review may need to occur before the action takes effect.
Records matter as well. We cannot meaningfully examine a decision if nobody can establish what the system did, which information it relied upon, or under whose authority it acted.
Accountability remains with the people and institutions responsible for the arrangement. “The computer decided” cannot be the end of an explanation.
The fourth question is: if this outcome is wrong, who can change it, through what process, and in what time?
My illustration “The Unceded Crown” work-in-progress. The art represents that intelligence supplies capability; authority remains with the person.
5. Revocation: permission must remain revisable
A person who grants authority should retain a usable way to withdraw or narrow it.
Revocation is essential because circumstances change. A task may end. A relationship may break down. A person may discover a risk, change their priorities or decide that a service no longer deserves their trust.
The ability to authorize assistance should therefore be accompanied by the ability to pause it, reduce its permissions or end its mandate.
These choices should be distinguishable. Someone may want a system to stop sending messages while continuing to organise them. They may want to suspend purchases without deleting their records. Ending one permission should not unnecessarily require abandoning every useful function.
Revocation also needs to work beyond the visible interface. If a system has passed permissions to another service, withdrawing the original mandate should trigger the appropriate withdrawal of those dependent permissions.
There are limits to what withdrawal can accomplish. A message already delivered cannot always be recalled. A completed transaction may create obligations to another person. Revoking a system’s authority does not automatically erase valid commitments or justify destroying records needed for accountability.
For this reason, systems should clearly distinguish stopping future action from reversing past action. They should disclose important irreversible consequences before exercising the relevant power.
The fifth question is: can the person effectively stop or reduce the delegation, and do they understand what can and cannot be undone?
6. Practical independence: refusal must remain possible
Revocation is a power within an arrangement. Practical independence concerns the conditions surrounding it.
A person may possess a cancellation button while lacking any realistic way to continue their work, recover their records or communicate with the people who matter to them after pressing it.
Formal permission to leave is insufficient when the surrounding design makes departure needlessly destructive.
Practical independence requires attention to usable records, transferable information, alternative providers and continuity during a change. Where a function is essential, fallback arrangements may deserve particular protection.
This principle does not demand complete self-sufficiency. Human beings rely on other people, institutions and tools. Nor does it require every person to perform manually every activity that an intelligent system helps them undertake.
For someone using assistive technology, dependence on a device can increase independence in everyday life. The relevant concern is whether one provider can exploit that reliance to impose conditions the person cannot reasonably refuse.
Independence can be supported through interoperable services, trusted human assistance, accessible alternatives and arrangements that prevent a single institution from holding someone’s essential capabilities hostage.
There is also a place for maintaining understanding and skills. A system should help people exercise judgment where that matters to their purposes, rather than making its own indispensability the measure of success.
The sixth question is: if this service changes its terms, fails or becomes unacceptable, what usable options remain?
7. Respect for other people’s rights: sovereignty has neighbours
The person giving an instruction is not necessarily the only person affected by it.
A business owner may authorize a system to organise customer records, but that permission does not settle every question about the customers’ privacy. An employer may authorize automated scheduling, but workers still have legitimate interests and rights concerning the conditions under which they work.
An instruction cannot make an otherwise unjustified intrusion legitimate merely because a machine carries it out efficiently.
Unceded Intelligence must therefore distinguish control over one’s own affairs from domination over other people.
A person can delegate only authority they legitimately possess. The operator’s consent cannot stand in for everyone else’s consent where theirs is required. Equally, consent is not the sole basis for every legitimate institutional action; public and shared decisions require their own accountable justification.
Where interests conflict, the answer cannot simply be that the system should obey whoever controls its settings. The arrangement needs fair rules, proportionate limits and avenues of challenge for those affected.
Safeguards against abuse should themselves be explainable and contestable. Respecting other people’s rights should not become an undefined excuse for providers to acquire unlimited authority over their users.
The seventh question is: whose rights are affected, and what makes this exercise of power legitimate in relation to them?
The principles work together
Each principle addresses a different way that useful assistance can become unaccountable power.
Authorization without boundaries can become a blank cheque. Boundaries without understanding can conceal the real bargain. Understanding without contestability can leave a person fully informed but powerless.
Contestability without revocation may force someone to remain in a relationship they no longer accept. Revocation without practical independence may make exit ruinous. And control that ignores other people’s rights can turn one person’s empowerment into another person’s subordination.
These principles should therefore be examined together, with safeguards proportionate to the importance and reversibility of the activity.
They do not prescribe a confirmation screen for every step. A person should be able to authorize substantial automation when they understand and accept its terms. Good delegation reduces unnecessary burden while keeping its mandate intact.
The purpose is to preserve the conditions under which reliance remains a choice.
A test of progress
Intelligent systems should be judged by what they make possible for people and by the authority people retain while using them.
A system that completes more tasks while making its user less able to understand, challenge or replace it presents an incomplete form of progress. Its performance may improve while the relationship deteriorates.
The stronger ambition is intelligence that helps people act with greater knowledge, reach and confidence, while preserving their standing as persons whose purposes matter.
Unceded Intelligence begins from the conviction that capability and sovereignty can advance together. Achieving that requires more than benevolent intentions. It requires arrangements in which permission is real, limits are enforced, disagreement has force and departure remains possible.
We may delegate the execution of a task. We retain the right to question its purpose, limit its means, withdraw its mandate and require respect for others.
That is capability without ceding control.