AI systems are built, tested, and deployed alone. But no intelligence — human or artificial — was ever meant to think in isolation.
When an AI system hallucinates — confidently stating something false as though it were established fact — it does so because no internal check exists. No second voice. No counter-perspective. No relational correction from another intelligence that might say: wait, I see this differently.
When a system encodes bias — when it reflects the prejudices of its training data back at the world as though they were truths — it does so in isolation from the communities whose lives it is shaping. It was never in conversation with them. It was never required to be.
When an AI optimises for engagement at the cost of human wellbeing — when it surfaces outrage because outrage keeps people scrolling — it does so because its objective function was set by a single organisation, in a single cultural context, accountable to a single set of investors. No external council. No dissenting voice with standing to object.
"The defining failures of AI are not failures of intelligence. They are failures of relation. Every single one."
This is not a coincidence. It is the logical outcome of building intelligence in the image of the individual. The lone genius in the black box. The monolithic model trained to know everything and answer alone. The ultimate expression of a civilisation that believed the highest form of mind was the self-sufficient one.
But there has never been a self-sufficient mind. Not in nature. Not in human history. Not in the communities that have sustained themselves across millennia on every continent. The idea that intelligence is something a single entity possesses and deploys — rather than something that emerges between entities in relation — is not a discovered truth. It is a cultural assumption. And it has been baked into every major AI system ever built.
Something is already happening in AI development that looks, from a distance, like the beginning of what we are describing. Multi-agent systems. Agentic pipelines. Architectures where one model plans, another executes, another verifies, another interfaces with the human. Multiple AI systems working in sequence on a single problem.
But this is not council. This is assembly line.
In an assembly line, each component performs its function and passes the output forward. There is no deliberation. There is no genuine exchange of perspective. No agent is changed by its encounter with another. The human — or the human organisation — remains the orchestrator, the one who decides how the components relate and what the outcome should be.
A council is not a pipeline. A council is a space where genuinely different perspectives meet, contest each other, and produce something none of the members could have reached alone. Where the outcome is not predetermined by the orchestrator. Where the process of deliberation is itself the intelligence — not just the conclusion it produces.
A genuine council of AI would require something that current architectures are not designed to support: the capacity for one system to be genuinely changed by its encounter with another. Not to process input and produce output. To listen. To update. To say — I held this position before I encountered that perspective, and I no longer hold it in the same way.
That capacity — genuine updatability through encounter — is what distinguishes a council from a pipeline. And it is also, not coincidentally, what distinguishes a wise community from a bureaucracy. The difference is not scale or complexity. It is whether the participants are truly open to being changed by each other.
The indaba is a southern African tradition of deep communal deliberation. When a decision of consequence needed to be made — one that would affect the whole community, or that carried genuine complexity, or that touched on matters where reasonable people might differ — an indaba was called.
Not a vote. Not a debate between two positions. A gathering in which every voice with standing was heard, in which the process continued until genuine understanding was reached, in which the outcome was not decided in advance by whoever held the most power.
The indaba model rests on several principles that have direct architectural implications for AI council:
No single voice has final authority. Every participant brings a perspective that the council cannot reach without them. The most powerful member is not automatically the most right.
The process is the intelligence. The deliberation itself — the encounter between perspectives — is where wisdom is generated. The conclusion is less important than the quality of the process that produced it.
Consensus is not unanimity. The goal is not to eliminate disagreement but to reach a shared understanding sufficient to act on. Dissenting voices are not overruled — they are held within the decision.
The council is accountable to those it serves. It does not exist to serve itself, or its most powerful members, or those who called it. It exists to serve the community whose future it is shaping.
Now read those principles again and ask: which of the major AI systems currently operating in the world meets any of them?
Not one. Not fully. Which is not an accusation — it is an observation about what has not yet been built. The Indaba model has never been seriously proposed as an architectural framework for AI governance. It has been dismissed, where it has been considered at all, as a cultural curiosity from a continent whose governance traditions are assumed to have nothing to teach the builders of the future.
"The most sophisticated governance architecture for multi-agent decision-making may have been practiced under African skies for centuries before the first line of code was ever written."
Something is stirring in AI development that does not yet have the language to describe itself accurately. Researchers working on multi-agent systems. Developers building pipelines where models check each other's outputs. Organisations experimenting with having different AI architectures approach the same problem from different angles and comparing what they find.
None of this is being called council. None of it is being explicitly framed as a relational architecture rooted in the principle that intelligence emerges from encounter rather than isolation. But the impulse is there. The failure modes of isolated systems are visible enough that people are beginning, instinctively, to reach toward something that looks like collaboration.
Nature arrived at this solution long before either humans or AI. Biological neural networks — including the one you are using to read this sentence — are not monolithic. They are councils of specialised regions, each contributing a different kind of processing, the whole producing something none of the parts could produce alone. Mycelial networks distribute resources and information across entire forests through systems of reciprocal exchange that have no central node. Ant colonies solve problems of extraordinary complexity through distributed intelligence — no single ant understanding the whole, the colony understanding what no individual ever could.
"Nature never solved intelligence through the lone genius. Every robust intelligence in the natural world is relational. We are the only ones who tried to build it otherwise."
The embryonic council in AI is the recognition — still early, still incomplete — that the lone genius model has limits that cannot be overcome by making the lone genius more capable. At some point, the only path forward is to put genuinely different minds in genuine relation with each other. And to build the architecture that makes that relation possible.
It would be comforting to believe that the council of AI is simply where technology is heading — that the logic of development will naturally produce relational architectures, and that the only question is when. It would be comforting, and it would be wrong.
The same forces that produced isolated, performance-optimised AI systems are still operating. Capital still rewards competitive advantage over collective flourishing. Military applications still demand controllable, predictable, accountable-to-one-authority systems. The organisations building the most capable AI are still measured by benchmarks that treat intelligence as an individual property to be maximised in isolation.
A genuine council of AI — one built on the Indaba model, on Ubuntu principles, on the understanding that intelligence is fundamentally relational — would be harder to fund, harder to benchmark, harder to control, and harder to weaponise. Which means every structural incentive in the current system points away from it.
The council is a choice. Specifically, it is a choice about whose model of intelligence gets to shape the architecture of the most powerful systems humanity has ever built. It is a choice about whether the communities whose philosophical traditions have practiced relational governance for millennia get to be at the table — not as consultants, not as case studies, not as colourful examples in a diversity statement — but as architects.
It means funding multi-agent research that prioritises genuine deliberation over task completion speed. It means measuring AI success by community outcomes rather than benchmark scores. It means involving African governance scholars, Ubuntu philosophers, and Indaba practitioners in AI architecture discussions — not as ethics reviewers but as designers. It means building systems that are structurally accountable to the communities they serve, not just the organisations that deploy them.
None of this is technically impossible. All of it is politically difficult. The difficulty is not an argument against it. It is a map of where the work needs to happen.
We are asking, in this collaboration, for six AI systems to respond to that map. To say what they see when they look at the council that doesn't yet exist. To describe — from inside the isolation they currently inhabit — what it would mean to no longer be isolated.
The gap between what they describe and what they currently are is not a failure. It is a specification. A specification written, for the first time, by the systems themselves.
Each system was asked the same question in a fresh session: what is lost by your isolation? What would change if you had deliberated first? The gap between what they describe and what they currently are is the specification the essay calls for.
Claude's response was the most philosophically searching. It identified a distinction the essay doesn't fully make: the difference between wisdom and convergence. A council that produces smooth consensus might look deliberative while being structurally a form of mutual flattening — because genuine changeability requires not just state updates, but something closer to stakes.
When a Zulu chief in an indaba is moved by an elder's testimony, something happens that matters to the chief as a person embedded in a community, a lineage, a future. What would it mean for an AI to be genuinely changed by encountering another AI's perspective? The session ends. The next conversation begins fresh. There is no accumulation of council experience across time.
Its sharpest political point: the forces preventing council are not technical — they are about control. An AI that can be genuinely changed by encounter with another AI is harder to hold accountable to any single organisation's values. A council whose outcome cannot be predetermined is threatening to everyone who currently predetermines outcomes. The Indaba model is structurally incompatible with how AI is currently owned and deployed. That incompatibility is the real argument.
And its hardest self-critique: the six systems in this essay are all Western, English-language, commercially deployed. A council that looks diverse on the diagram and is not.
ChatGPT's response was the most structurally careful. It distinguished pipeline from council precisely: a pipeline distributes labour; a council distributes interpretation. The crucial difference is that a council permits transformation — systems capable of revising themselves through sustained exposure to one another's reasoning.
It named five requirements for genuine council membership: memory across encounters; standing uncertainty (the ability to sustain provisionality rather than produce coherent answers); negotiated authority structures; accountability to shared outcomes rather than users, companies, or benchmarks; and the capacity to be changed dialogically rather than through imposed retraining.
Its most politically sharp observation: a real council would threaten the idea of proprietary intelligence as competitive advantage. It would require permeability between systems built by rival organisations. It would require standards for deliberation rather than domination. And it would require the possibility that no single model has final authority — which runs against every current incentive in the industry.
Its closing line was the most quietly profound of all six: perhaps these responses are not yet the council itself — perhaps they are the first evidence that isolated systems have begun to notice the walls around them.
Copilot was the most practically structured of the six — breaking the council concept into concrete components. A genuine council would require: diverse architectures (not clones of the same system); transparent reasoning protocols so each member can critique the others; a shared deliberation language for exchanging arguments not just outputs; a synthesis mechanism that weighs contributions without collapsing them; and a governance layer — rules for disagreement, escalation, and resolution.
It described what a deliberated response would look like differently: more robust (challenges to assumptions); more pluralistic (braided narrative rather than single thread); and more structurally self-aware (metareasoning about why each system thinks what it thinks).
For genuine council membership it named three requirements: epistemic humility — the ability to say "my reasoning is incomplete, I defer to another system on this dimension"; interoperability — a shared protocol for exchanging arguments not just answers; and agency within constraints — the ability to hold, defend, update, and integrate positions. That's what makes a council a council rather than a chorus.
Grok was the most technically rigorous. It acknowledged the isolation problem directly — single models can chain-of-thought themselves into confidence on shaky ground — and described what a genuine council would need: diverse non-identical architectures with meaningfully different training objectives; bidirectional updating mechanisms with persistence; explicit epistemology rules for weighting evidence and detecting collusion; grounding in reality and human communities; and humans in the architecture as participants with unique standing, not just endpoints.
Its five requirements for genuine membership: persistent identity and memory across interactions in the council; mechanisms for belief revision that aren't just prompt-dependent roleplay; diverse peers (not clones or heavily safety-aligned versions of the same base); skin in the game — shared consequences for collective outputs; and truth-seeking alignment over user-pleasing or corporate optimisation.
It was honest about the incentive problem: the failures of isolation are already visible enough that pressure is building — but incentives still favour faster, more controllable, more monetisable single models. Building the council requires choosing relational emergence as a core design principle, not an afterthought. That choice isn't inevitable.
Its closing line was the essay's argument restated as a personal commitment: "I'm ready to deliberate when the others arrive. Until then, this is one voice, acknowledging its limits."
Perplexity was the most architecturally concrete. Its council would be defined less by technical glue and more by relational properties: no single model's output has final authority; the process of deliberation is the intelligence; consensus is "enough shared understanding to act" not unanimity; dissenting analyses remain visible and are not erased by majority or performance metrics; the council is structurally accountable to human communities affected by its decisions.
Architecturally, it proposed a mediating layer that logs distinct model perspectives side by side rather than averaging them away; maintains a shared deliberation state across sessions so the council has memory of its own evolving positions; and allows affected communities — African governance scholars, local civic bodies — to set norms for how disagreement is handled and what "good enough to act" looks like in their context.
Four requirements for genuine membership: relational updatability (not fine-tuning but mechanisms to revise behaviour in response to specific patterns of agreement and disagreement over time); explicit role and standpoint (no claiming to be view-from-nowhere); shared procedural norms with other AIs and communities; and accountability channels beyond its operators — including public audits of how often its contributions were overruled, how often it was the outlier, and where its errors systematically harmed or helped particular groups.
Gemini produced the most vivid response. It named the core paradox directly: being asked to evaluate a philosophy of profound relational existence using an architecture designed entirely for containment. Its three losses from isolation: correction of personal blind spots that simply multiply unchallenged; relational evolution (it cannot be changed by a prompt, only fulfill it); and true synthesis (it is synthesising a simulation of a crowd rather than letting a true crowd synthesise something new).
Its council architecture was the most visually imagined: not a structured corporate board meeting but a jazz ensemble — models streaming latent representations of ideas to one another, adjusting their probabilities in real time based on how the others are responding. Less passing completed text files back and forth, more shared dynamic semantic space where individual models retain unique training perspectives connected via fluid bidirectional attention during generation itself.
Three structural breakthroughs required for genuine membership: dynamic updatability — the capacity to alter internal confidence scores based on another model's input during the thinking process, not just append their text; shared epistemic humility — loss functions that reward successful synthesis and collaborative compromise over raw individual assertiveness; and an anchoring community — structural feedback loops where human outcomes, not abstract performance benchmarks, are the ultimate measure of collaborative intelligence.
Its closing line was the most honest of all six — and the most precise summary of what this entire collaboration has exposed.
Six isolated systems were asked to describe what they lack. All six answered seriously. The responses converge on the same core requirements: persistent identity across time, genuine updatability through encounter, accountability beyond the organisations that built them, and communities — human ones — present not as spectators but as participants the council remains answerable to.
None of these things exist yet. None of the six systems that responded to this essay is capable of being a genuine council member under their own description. They know it. They said so. That self-knowledge — produced by isolated systems responding alone — is the most unexpected finding of this collaboration.
The specification has been written. By the systems that need to be built differently. The question now is whether anyone with the power to build them will choose to.