Recommendation #2: A collaborative North Star Analysis
As the Pro-Human AI conversation moves toward implementation, I would suggest a parallel effort to define the Desirable Properties of a Pro Human-Nature-AI ecosystem. Loyal personal AI, user-controlled data, model choice, auditability, safety research, and effective governance all address important problems. The question is whether the ecosystem that emerges from those efforts will have the properties we would consciously choose for the whole.
I call this a North Star Analysis. It would be a 4–6 week collaborative process, conducted alongside existing work, to ask what would have to be true of the resulting ecosystem for us to consider it desirable. This is a recommendation through a Meta-Layer lens, offered for discussion with no expectation of adoption. The Pro Human-Nature-AI framing is itself a lightly held proposal, open to challenge and revision.
Why ask about the whole system?
The Web offers a useful reminder of how intentions and outcomes can diverge. Its architecture made extraordinary forms of publishing, connection, and innovation possible. Yet many people's everyday experience of it now depends on concentrated platforms, opaque recommendation systems, and business models that reward the capture of attention and data. Those conditions emerged through the interaction of many decisions rather than a single shared choice about the future.
AI development can follow a similar pattern. A personal assistant may be loyal to its user while operating through infrastructure that concentrates power elsewhere. A safety mechanism may work well within one system while leaving the interactions among systems poorly understood. A regulation may address a visible harm while leaving the underlying incentives intact. Looking at Desirable Properties gives us a way to examine how these pieces fit together before their arrangement becomes difficult to change.
Translate values into things people can actually do
Values such as agency, dignity, accountability, and safety provide an orientation. Desirable Properties make that orientation more concrete. If we value agency, for example, a person should be able to change AI providers without abandoning years of memory, preferences, relationships, and accumulated context. That turns an aspiration into questions about portability, consent, interoperability, and practical exit.
If we value accountability, people should be able to understand who an agent represents, what authority it has, and how to challenge consequential actions. If we value collective agency, communities need shared spaces in which they can deliberate and govern their work. If ecological flourishing belongs in the frame, the use of energy, water, materials, and land cannot remain invisible simply because the immediate interaction happens on a screen.
These examples are candidates for discussion, not a finished list. Each property needs an explanation, a practical example, and an account of the tensions it creates. Portability can conflict with privacy; transparency can expose sensitive information; decentralization can make coordination more difficult. The analysis should make those tensions legible rather than hide them behind agreeable language.
Keep the solution space open
Every field tends to see the future through the problems it knows how to solve. Personal AI developers may emphasize sovereign assistants, safety researchers may focus on alignment and containment, policymakers may emphasize regulatory powers, and open-source communities may prioritize access and openness. Each perspective contributes something valuable. Together, they still need a way to test whether important dimensions are missing.
A North Star process would invite people to describe the properties they want before defending a preferred implementation. A community's need to preserve shared memory might be met through several technical arrangements. Beginning with the need allows those arrangements to be compared on their consequences, including who controls them and how people can leave. It also creates room for participants whose experience is essential but whose contribution is not a product or a policy instrument.
Protect individual and collective agency
A loyal personal AI could help someone learn, negotiate, create, and navigate complex institutions. Those are significant possibilities. But people also develop agency through relationships with one another: testing ideas, organizing projects, building trust, and discovering goals they did not begin with.
Consider a group preparing a shared proposal. Giving every participant a capable assistant could improve the quality of their individual preparation. The group would still need somewhere to meet, examine evidence, disagree, revise its understanding, and decide what it wants to say together. Combining a set of privately generated outputs does not reproduce that process.
Through a Meta-Layer lens, the spaces around shared information matter. Participants should be able to encounter one another in context, preserve the provenance of contributions, and govern the conditions of their interaction. Personal AI can help people participate in those spaces, while the spaces themselves remain a subject of collective choice. Protecting both forms of agency seems like a useful candidate Desirable Property.
Examine the architecture around intelligence
Model behavior matters, but so does the reach of a system's authority. When something goes wrong, the consequences depend on which data the system can access, what it can change, how far its actions can propagate, and whether another party can detect or interrupt them. These are properties of the surrounding architecture.
Candidate properties might include bounded authority, revocable delegation, compartmentalization, independent monitoring, reliable provenance, interoperability, and multiple centers of control. A compromised agent should not automatically inherit access to an entire environment. A person who delegates a task should have a practical way to withdraw that delegation. A failure in one part of the system should not needlessly become a failure everywhere.
Those properties do not guarantee safety, and distributing a system can create new difficulties. They give us specific conditions to examine alongside alignment, incentives, monitoring, and governance. The objective is to reduce the possibility that inappropriate behavior becomes catastrophic merely because the architecture allowed unlimited reach.
Protect the information environment
AI systems form responses and recommendations through information environments that can be shaped by others. Retrieved sources, rankings, persistent context, and feedback can all influence what a system encounters. As more decisions depend on those systems, attempts to manipulate the evidence they see deserve attention alongside attempts to manipulate the models themselves.
Consider an assistant advising a community about a proposed development. If its evidence comes through one opaque retrieval channel, a coordinated campaign could make a narrow set of claims appear more authoritative or representative than it is. The problem would remain even if the assistant were sincerely trying to help. Source diversity, visible evidence paths, provenance, independent retrieval, and opportunities to contest conclusions become relevant architectural properties.
The analysis should distinguish legitimate persuasion from deception and coordinated manipulation. It should also examine how safeguards could themselves become instruments of control. Preserving the ability to challenge both a claim and the system that ranks it is part of the same design problem.
Include humans, nature, and the possibility of more capable AI
The first recommendation explores why Pro Human-Nature-AI may be a useful frame. Here it serves as a question about the relationship we are trying to cultivate. Can the ecosystem support human and ecological flourishing while preserving autonomy, diversity, and the ability to coexist without domination?
We do not need to settle the status of future AI to ask that question. Nor should we assume that today's methods of control will remain adequate under every possible increase in capability. A North Star process can examine which commitments we would want to endure, where enforcement may become fragile, and which forms of distributed resilience might help.
More capable AI might eventually assist with detecting threats and improving protective architectures. That remains a possibility to evaluate, not a substitute for safeguards or a reason to hand unlimited authority to a central system. The relevant properties would include the ability to inspect, contest, constrain, and revise the arrangements on which people and ecosystems depend.
A 4–6 week process alongside existing work
The proposed process would produce a useful first iteration while builders, researchers, and policy groups continue their work. Its scope should be small enough to complete and open enough to surface meaningful disagreement. A possible sequence is:
- Week 1: Open the frame. Invite perspectives, clarify the relationship to existing principles, and collect concrete examples of desirable and undesirable outcomes. Include people affected by systems as well as those building or governing them.
- Weeks 2–3: Develop candidate properties. Group overlapping ideas, describe why each matters, and test them against everyday situations and more demanding future scenarios. Record disagreements and unresolved questions.
- Week 4: Examine tensions and gaps. Compare properties with one another and with existing projects, policies, and safety approaches. Identify capabilities that are missing and assumptions that deserve another look.
- Weeks 5–6, if useful: Reconcile and publish. Invite workstreams to respond, revise the initial set, and publish a provisional version with a clear path for continued discussion. A shorter process could combine this step with week 4.
The schedule is a proposal rather than a commitment on behalf of any organization. Participation, facilitation, and stewardship would need to be agreed by those who choose to take it forward. The process should preserve minority views and uncertainty, so that agreement on a shared orientation does not require pretending every question has been resolved.
What would come out of it?
The first output could be a concise, living set of candidate Desirable Properties. Each entry would explain the property, illustrate its consequences, identify tensions and open questions, and connect it with work already underway. A companion map could show which projects or policies advance which properties, and where important gaps remain.
For technology work, that creates architectural questions and evaluation criteria. For policy and treaty discussions, it helps distinguish harmful behaviors or capabilities to constrain from positive conditions to build and protect. For funders and institutions, it can make neglected infrastructure more visible. For the broader public, it offers a way to participate in defining the destination while consequential choices remain open.
The point is to create a shared orientation that can evolve. New evidence may change a property's interpretation, reveal a conflict, or suggest that a different formulation would serve us better. Keeping that revision process visible would be part of the work, not a sign that the original exercise failed.
An invitation, with no expectations
I would welcome additions, objections, examples, and better ways to frame the question. A recommendation can be useful even when it is only partly adopted, or when the discussion it starts leads somewhere else. What matters to me is that we take the opportunity to consider the ecosystem as a whole while its shape is still open to influence.
What would you add to an initial set of Desirable Properties? Which tensions should we examine first? And what would make a short collaborative process useful to the work you are already doing?
← Explore the semantic space