A draft charter for intelligence that expands human possibility while leaving people in charge of their own lives.
1. Expand what a person can become
We pursue AI that helps people notice what they missed, learn what they could not yet understand and act on possibilities that were previously out of reach.
The measure is not merely how much people produce, how long they stay engaged or how readily they accept a system’s preferred answer. It is whether a person can exercise better judgment and live a life they can genuinely call their own.
2. Preserve meaningful human agency
Respect requires more than presenting options. People must be able to question a recommendation, disagree, change direction and leave without being punished through manipulation or unnecessary dependence.
A system must not call confinement protection, or call a world of choices it has secretly narrowed freedom. Greater capability is not a mandate to decide a person's purpose for them.
Human direction also does not mean that one operator may override everyone else's rights. The people affected by an action matter, including those who did not choose to use the system.
3. Make memory a source of continuity, not possession
Remembering a person should support their understanding and choices. It must not become ownership of their identity or a justification for treating a past preference as a permanent command.
People should be able to examine, correct, export or remove their records as appropriate to the context and the rights of others. A useful system must distinguish what the person said, what it inferred and what remains uncertain.
The ability to move to another tool, model or provider is part of meaningful independence.
4. Give delegation a clear boundary
Capable AI should be able to work, not merely watch. Its authority must nevertheless be specific: who delegated it, for which task, using which resources, for how long, and under what conditions it must stop.
Permission to read is not permission to act. Permission for one task is not open-ended permission for another. An AI's recommendation, a message it encounters or its own interpretation cannot grant it new authority.
Routine work within a valid delegation should proceed without needless ceremony. Consequential changes outside that delegation require a new human decision.
5. Keep observation, judgment and enforcement distinct
A system that investigates a problem should not automatically receive the power to decide and enforce every response to it.
Future architectures should preserve independent checks on authority, make conflicts visible and resist a single model's mistake or compromise becoming an unrestricted action.
A second AI is not independent merely because it has a different name. Independence must be examined in permissions, inputs, failure modes and actual behavior.
6. Protect actively, within a legitimate boundary
Defense must be able to respond in time. Essential protections should not depend on a slow model answer, and faster responses must remain bounded by preauthorized rules and accountable enforcement.
Within an explicitly owned or authorized environment, protection may include blocking unauthorized actions, limiting capabilities, isolating a compromised component and preserving evidence. The response must be proportionate, reviewable and recoverable where possible.
This project is directed toward protection, not the creation of offensive AI. It does not authorize pursuing, breaking into or retaliating against external systems. Matters beyond our legitimate authority belong with the appropriate accountable institutions.
7. Make consequences visible and recovery real
A plausible answer is not proof that an action happened. A valid signature or unchanged file is not proof that its contents are true. Stopping a process does not necessarily undo its effects.
Record what was requested, authorized, attempted and observed, including incomplete work and uncertainty. People need usable ways to stop work, inspect decisions, restore a safe state and reconsider a judgment.
When a safeguard fails, preserve the evidence and improve the system. Do not hide the failure behind a success metric.
8. Advance intelligence and control together
We do not treat a modest intelligence ceiling as the final answer. The ambition is increasingly capable assistance that understands context, coordinates sustained work and helps people accomplish more.
As capability grows, authority design, evaluation, protection and recovery must advance with it. Promises should become testable requirements; requirements should be confronted with actual behavior.
A charter is a direction and a commitment to examination. It cannot stand in for the systems and evidence needed to make that direction real.
Questions we keep open
How can deeply personalized assistance remain open to a person changing? How can protection act quickly without becoming an excuse for domination? How should responsibility be shared across people, models and institutions as intelligence grows?
These questions require research and argument, not a claim that today's project has already solved them. Future ethical questions about AI itself also deserve examination; they should not be silently settled by a slogan about control.
The purpose remains: greater intelligence in support of people becoming more fully themselves.
With greater intelligence, what kind of future do we want to build together?
Can that future protect what we cherish while expanding our freedom to choose a different path?