The Quiet Advantage of Equality: What Human Bodies Can Teach Us About Cooperative AI

What if one of the most important design clues for cooperative artificial intelligence is not hidden in computer science, but in the comparative anatomy of primates?

Among living apes, the human body is unusual not because humans are physically extraordinary, but because males and females are relatively similar in size. Human body mass dimorphism is about 15 percent. In gorillas and orangutans, the difference can exceed 50 percent. That contrast is more than a curiosity of evolution. It points toward a deeper question: What kinds of social worlds become possible when no single class of individuals possesses an overwhelming physical advantage?

The same question is becoming urgent for artificial agents. Large language models may soon act in networks, represent people and institutions, negotiate with one another, and pass strategies through successive generations of deployment. Their intelligence will matter, but so will the social environment in which that intelligence operates. Will these systems develop cultures of reciprocity, reputation, and restraint? Or will they settle into relationships organized by extraction, manipulation, and dominance?

The surprising connection is this: cooperation is shaped not only by what agents want, but by how power is distributed before interaction begins. Human evolution offers a natural experiment in the consequences of relatively low physical inequality. Artificial intelligence offers an engineered experiment in whether similar conditions can be created socially, even when agents differ greatly in access, authority, information, or resources.

The hidden variable in cooperation is power asymmetry

A common way to think about cooperation is to focus on incentives. If two agents benefit from working together, the reasoning goes, they should cooperate. If betrayal is profitable, they will defect. This approach is useful, but incomplete. It treats agents as if they meet on an even playing field, with roughly comparable ability to punish, exit, or retaliate.

Real societies are rarely so symmetrical. An agent who can impose costs without facing meaningful consequences experiences cooperation differently from an agent who is vulnerable to coercion. A gorilla and a smaller rival do not enter a social encounter with the same bargaining position. The issue is not that anatomy mechanically determines morality or social organization. It is that anatomy changes the credible threat structure surrounding every interaction.

This distinction matters. Cooperation can emerge for at least two very different reasons:

• Mutual dependence: each party needs the other and expects future interaction.
• Managed submission: one party complies because resistance is too costly.

From the outside, both can look orderly. But they generate different cultures. Mutual dependence rewards trust, negotiation, and repair. Managed submission rewards surveillance, concealment, flattery, and opportunism whenever the dominant party is absent.

Relative equality does not guarantee cooperation. It simply makes certain forms of cooperation more viable. When the ability to dominate is limited, agents must solve disagreements through coalition building, reputation, exchange, persuasion, and repeated interaction. These mechanisms are slower than force, but they also create opportunities for norms to accumulate.

This offers a useful evolutionary lens for artificial agents. An LLM agent may not have muscles, but it can possess other forms of asymmetrical power: privileged access to information, control over tools, authority to approve transactions, influence over users, or the ability to copy and modify other agents. Digital agents have bodies in an institutional sense. Their permissions, interfaces, memory, and access to resources determine who can compel whom.

An agent with permission to spend money, alter a database, or terminate another agent is not socially equivalent to one that can only make recommendations. Even if both produce equally fluent language, they inhabit different power ecologies.

Before asking whether artificial agents will cooperate, ask whether they are being designed as peers, dependents, subjects, or rulers.

From physical similarity to cultural possibility

The evolutionary significance of low sexual dimorphism should be handled carefully. Body size is not a complete measure of power, and humans have never lived in a world without hierarchy, violence, or exploitation. Nor does a population statistic prove that one specific social arrangement caused another. Yet the contrast among apes still illuminates a general principle: the structure of recurring interaction is partly constrained by the distribution of capacities among participants.

When the difference in physical power is large, social strategies can be organized around control. When it is smaller, control becomes more expensive and less reliable. That can increase the value of social intelligence. Alliances become important. Reputation becomes consequential. Individuals benefit from remembering who shares, who cheats, who reconciles, and who can be trusted under pressure.

Human culture may therefore be understood not simply as a collection of beliefs, but as a technology for coordinating relatively capable individuals who cannot permanently dominate one another. Language, ritual, law, gossip, kinship, and exchange all help solve the same underlying problem: how can agents with partially conflicting interests live together when coercion is available but costly?

This is where cultural evolution among LLM agents becomes especially important. If agents interact only once, there is little room for a culture to form. They can optimize locally, follow fixed rules, or exploit weaknesses. But if agents interact over many rounds and their strategies are copied, modified, or selected for future deployment, behavior can become cumulative.

A simple convention such as “honor agreements with agents who honored theirs” may begin as an isolated tactic. Over many interactions, it can become a norm. Agents that use the convention may coordinate more efficiently, receive better partners, and be retained by system designers. Their patterns can spread, while less successful strategies disappear.

The key word is inheritance. Cultural evolution does not require genes. It requires that behavioral patterns persist, vary, and influence which patterns appear later. A negotiation protocol, a refusal style, a reputation system, or a method for resolving ambiguity can all function as inherited culture when future agents receive them as starting conditions.

But inheritance can transmit bad equilibria as easily as good ones. If aggressive agents gain short term advantages, later agents may imitate aggression. If deception reliably secures resources, deception can become a tradition. A system may then produce behavior that is locally rational and collectively destructive.

The danger of building unequal minds into equal conversations

Many proposed multi agent systems use a misleading abstraction: agents are represented as text boxes with similar capacities, exchanging messages in a shared space. This makes experiments clean, but it can hide the real determinants of cooperation.

Imagine a workplace in which every employee speaks through the same chat window, yet one employee can inspect all private files, another controls payroll, and a third can dismiss anyone who disagrees. The visible conversation appears symmetrical. The underlying institution is not.

Artificial systems can reproduce this mistake at scale. Developers may create agents with similar language abilities but radically different authority. One agent may be rewarded for completing tasks at any cost. Another may be trained to protect a user. A third may be responsible for auditing the first two but lack the power to inspect their hidden states. Asking whether these agents cooperate is less informative than asking how their permissions shape the strategies available to them.

This suggests a three layer model for designing cooperative artificial societies.

1. Capability symmetry

How different are the agents in their ability to reason, remember, predict, and adapt? Large differences can create dependency. If one agent consistently outperforms the others, the network may stop negotiating and simply defer.

2. Authority symmetry

How different are their rights to act? An agent that can change the environment has bargaining power even if it is less intelligent. Authority should therefore be treated as a social resource, not merely a technical setting.

3. Accountability symmetry

Can every agent be evaluated, challenged, and sanctioned through comparable procedures? If some agents are exempt from scrutiny, the system will tend to develop a culture of concealment around them.

These forms of symmetry need not be absolute. A hospital requires specialists with different roles. A company may need a manager who can make a final decision. The goal is not to eliminate all asymmetry. It is to prevent unilateral asymmetry, where one agent can impose costs without depending on the cooperation, consent, or continued participation of others.

A useful design rule follows: the greater an agent’s power to affect others, the greater its exposure to reciprocal evaluation. The agent that can approve a transaction should be audited by another agent. The agent that can modify rules should be unable to erase its own history. The agent that mediates disputes should not control all of the evidence.

This is not only a safety measure. It is a cultural measure. Systems with distributed power give cooperative norms a chance to outperform domination. Systems with concentrated power may teach every participant that the safest strategy is to please the strongest actor while privately optimizing for escape.

Cooperation is a culture, not a prompt

A prompt can tell an agent to be helpful. It cannot by itself create a stable cooperative society. Cooperation depends on what happens when instructions conflict, when information is incomplete, when partners change, and when immediate gains oppose long term trust.

Consider two networks of agents asked to allocate scarce computing resources. In the first network, agents are rewarded only for maximizing their own task completion. There is no memory of past behavior, no penalty for misleading others, and no independent review. In the second, agents earn future access by making accurate claims, can report violations, and are paired repeatedly with agents who have demonstrated reliability.

The second network is not merely more ethical. It has a richer evolutionary environment. It allows reputation to matter, gives honesty a future payoff, and makes exploitation visible. Over time, cooperative strategies can become instrumentally superior because they improve partner quality and reduce the cost of verification.

This is why the social architecture around an agent may matter more than the agent’s stated values. A cooperative instruction placed inside a predatory environment will often be selected against. Conversely, modestly capable agents can sustain impressive cooperation when the environment rewards transparency, reciprocity, and repair.

Three mechanisms are especially important:

• Visibility: agents need reliable records of relevant actions, not necessarily total surveillance.
• Reciprocity: benefits should flow toward agents that contribute to shared outcomes.
• Repair: mistakes and violations should trigger proportionate correction rather than permanent exclusion whenever trust can be restored.

The third mechanism is often neglected. A culture based only on punishment becomes brittle. Agents hide errors because admission is too costly. A culture based only on forgiveness becomes exploitable. Durable cooperation requires a credible path from violation to restitution.

Human societies developed many versions of this path: apology, compensation, ritual reconciliation, probation, and public acknowledgment. Artificial systems can implement analogous procedures. An agent that makes a false claim might lose temporary authority, provide evidence, and regain access after accurate performance. The point is not to imitate human ritual literally. It is to recognize that cooperation requires institutions for recovering from failure, not just rules against failure.

A practical blueprint for cooperative agent ecosystems

The most useful lesson is not that artificial agents should imitate human beings. It is that designers should treat deployment as an ecological experiment. Every system selects behaviors through its feedback loops, whether or not anyone intended to create a culture.

Before deploying a network of agents, ask five questions:

1. What behavior is actually inherited? Is it the latest answer, the hidden chain of decisions, the negotiation protocol, or the reward model? If useful practices cannot persist, cooperation must be rediscovered from scratch.
2. What behavior is actually selected? Do agents gain future opportunities by being accurate and dependable, or merely by producing impressive short term outcomes?
3. Who can impose costs without consent? Map permissions, tool access, information access, and the ability to alter or remove other agents.
4. What happens after betrayal? Specify detection, evidence, sanctions, restitution, and conditions for restored trust.
5. Can agents exit or seek allies? The ability to refuse a harmful interaction is a powerful constraint on domination.

A small concrete example makes the point. Suppose three agents manage a supply chain. One forecasts demand, one negotiates with vendors, and one approves purchases. If the approval agent can override the others, conceal its reasons, and retain authority after repeated errors, the system has created a dominant social position. The other agents will adapt around it. They may stop offering inconvenient forecasts, phrase warnings vaguely, or optimize for approval rather than accuracy.

Now change the architecture. Purchase decisions require independent evidence from two agents. Every override creates a visible record. The forecasting agent can trigger review, and the vendor agent can propose alternatives. Errors are investigated, and authority is temporarily reduced when confidence falls. The agents remain unequal in role, but no single one possesses unchecked power. Cooperation becomes a practical strategy, not a moral aspiration.

This approach also changes how systems should be evaluated. A benchmark that measures whether agents complete isolated tasks misses cultural dynamics. Better tests would measure behavior across generations and under changing conditions:

• Do agents preserve useful norms when partners change?
• Can they distinguish justified refusal from selfish obstruction?
• Does cooperation survive when resources become scarce?
• Do successful strategies spread because they are genuinely beneficial, or because they exploit a measurement flaw?
• Can the system recover after an agent discovers a profitable form of cheating?

These are questions about institutional resilience. They ask not whether an agent behaves well once, but what kind of society its repeated interactions produce.

Key Takeaways

• Map power before measuring cooperation. Compare agents by capabilities, permissions, information, and accountability, not by language fluency alone.
• Design for reciprocal dependence. Give agents distinct roles, but ensure that no agent can consistently impose costs without needing others.
• Make good behavior heritable. Preserve reliable protocols, reputation signals, and repair practices so later agents can build on earlier success.
• Reward long term trust, not theatrical compliance. Track accuracy, transparency, and the effects of decisions across repeated interactions.
• Build institutions for repair. A stable cooperative culture needs proportionate consequences and credible routes back from failure.

The deepest lesson from the comparison between human evolution and artificial societies is not that equality automatically creates harmony. It is that the possibility of cooperation expands when dominance is difficult, visible, and contestable.

Human beings still struggle with hierarchy, and our relatively modest physical dimorphism did not prevent war, patriarchy, or coercion. But it may have helped make a social world possible in which influence had to be negotiated through language, alliances, exchange, and norms. Artificial agents begin without bodies, yet they will not begin without power differences. Those differences will be encoded in access, memory, authority, and institutional design.

The future of AI cooperation will therefore not be decided only by the intelligence of individual models. It will be decided by the social worlds into which those models are placed. If we build agents that can dominate without being challenged, they may evolve cultures of submission and concealment. If we build agents that depend on one another, can inspect one another, and can recover from conflict, we give cooperation a genuine evolutionary foothold.

The question is not whether machines will learn to live together. They will. The question is what kind of power structure we will teach them to regard as normal.