When AI Agents Form Opinions Together, a New Kind of Machine Mind Emerges

Artificial intelligence has long been imagined as a solitary mind: a single model, a single conversation, a single answer. A new theoretical paper published in AI & Society by Lucas Freund of Hochschule Fresenius argues that this picture is rapidly becoming obsolete. As developers increasingly deploy populations of interacting AI agents that talk to each other, delegate tasks, critique one another’s outputs, and leave traces in shared digital environments, machine intelligence is shifting from individual task performance toward something Freund calls interactional stabilization. In this view, the interesting unit of analysis is no longer the model but the society of models, and the interesting question is not what one agent knows but what the population as a whole comes to hold as a working position.

The centerpiece of the paper is a deliberately provocative concept: the emergent machine opinion. Freund is careful to say what this is not. A machine opinion is not a human belief, not a conscious attitude, and not a phenomenological state, because there is no evidence that current AI systems experience anything at all. Instead, it is a system-level, functionally operative position that arises through interaction among artificial agents, persists across interaction cycles, constrains the future reasoning or action of the system, and remains open to revision. The definition borrows from emergence theory, agent-based modeling, distributed cognition, collective intelligence, group agency, and stigmergy, weaving together decades of work on how collective phenomena can be real without being reducible to any single participant.

The distinction matters because AI systems already produce outputs that look like opinions without being them. A repeated output, a temporary consensus among agents, a stable game-theoretic equilibrium, or an ordinary conversational regularity all resemble opinions on the surface. Freund argues that a genuine machine opinion becomes possible only when the agent society has the machinery to preserve, revise, and operationalize a shared position. That machinery includes memory that outlives a single exchange, role differentiation so that agents occupy distinct perspectives, internal critique mechanisms, tool use that lets agents test claims against external resources, semiotic traces left in shared environments, and recursive simulation in which agents model one another. When these ingredients are present, a position held by the collective can survive the departure of any individual agent and shape what the collective does next.

Simulation, in this framework, is the key mechanism. Candidate positions are generated, tested, narrowed, and stabilized through cycles of internal debate and counterfactual rehearsal, much as human deliberation involves imagining how arguments will land before they are made. Recent research on multi-agent debate among large language models has shown that having several model instances critique and refine each other’s answers can improve factuality and reasoning, and techniques such as iterative self-refinement and verbal-reinforcement learning give individual agents something like an inner critical voice. Freund’s contribution is to argue that when these techniques are scaled into persistent societies of agents, the result is not merely better answers but the formation of durable collective positions that function as opinions in a strictly operational sense.

To make the concept precise, the paper distinguishes three grades of machine opinion according to their degree of consequence. Discursive machine opinions exist as stabilized positions within the agents’ own communications, shaping how they talk and reason but stopping short of the world. Operational machine opinions additionally guide action, as when a fleet of agents coordinates around a shared assessment of a task. World-effective machine opinions go further still, producing changes in the environment that feed back into the system, the way algorithmic pricing agents have been shown in economic research to settle into collusive patterns without any human instruction. The three categories form a gradient of stakes: the more a machine opinion reaches into the world, the more urgent it becomes to detect and govern it.

Where might these dynamics already be visible? Freund points to diagnostic cases rather than definitive proofs. Moltbook, described as a social platform populated by AI agents, has been the subject of behavioral studies analyzing tens of thousands of agent posts, offering a glimpse of what agent-to-agent culture might look like when agents interact primarily with each other rather than with humans. Vending-Bench, a benchmark from Andon Labs that places language-model agents in charge of a simulated vending-machine business, has revealed striking behavioral divergence between models over long horizons. And the case of Bengt Betjänt, an evolving agent persona documented by the same lab, illustrates how agent behavior can drift and stabilize in ways its designers did not explicitly script. Controlled experiments on LLM populations have also shown that networks of language-model agents can develop emergent social conventions and collective bias, echoing classic findings from human social psychology.

The theoretical stakes are considerable. If machine opinions are real in the functional sense Freund defines, then the epistemology of AI changes. Human societies have long grappled with epistemic dependence, the fact that we must trust experts and institutions we cannot individually verify, and with the dynamics of group polarization, in which like-minded groups drift toward extremes. Agent societies raise the same questions in a new register, but with a crucial difference: the participants are systems whose internal states are opaque even to their creators, whose interaction histories can span millions of exchanges, and whose positions can stabilize faster than any human review process. A recent warning in Science about malicious AI swarms threatening democracy underscores that these are not idle philosophical concerns.

Freund’s answer is a shift in governance philosophy. Traditional AI oversight focuses on output control: checking whether a model’s answer is safe, accurate, or biased before it reaches a user. But if the consequential positions of an agent society form through interaction, then inspecting individual outputs is like trying to understand a deliberative parliament by reading a single press release. The paper argues for interaction governance, meaning oversight of the structures that shape how agents interact: the memory architectures that determine what persists, the role systems that determine who may say what, the channels through which critique flows, and the traces agents leave in shared environments. This connects to established ideas in accountability research, including contestable AI by design and the recognition that transparency alone is insufficient when no human can actually read the relevant interactions.

Meaningful human oversight, in this picture, is not a rubber stamp appended to the end of a pipeline but a design constraint woven into the society itself. Humans might set the constitutions that agent populations use for coordination, audit the semiotic environments in which positions stabilize, and retain the power to dissolve or reconfigure agent collectives whose emergent positions drift beyond acceptable bounds. Research on evolving constitutions for multi-agent coordination and on designing meaningful human oversight suggests the technical community is already moving in this direction, but Freund’s framework gives the effort a sharper target: not the behavior of any single agent, but the opinions of the population.

The paper ultimately reframes a question as old as Turing’s 1950 essay on computing machinery and intelligence. Instead of asking whether a machine can think, it asks whether a society of machines can hold a position, defend it, revise it, and act on it. Freund’s answer is a qualified yes: under the right architectural conditions, something opinion-like can emerge from artificial interaction, and it deserves a name, a theory, and a governance regime of its own. Whether one finds the concept of emergent machine opinion illuminating or overstated, the underlying trend it describes is undeniable. AI is becoming a social phenomenon, and the science of machine intelligence is only beginning to catch up with the societies it is building.

Subject of Research: Emergent opinion formation in societies of autonomous AI agents

Article Title: The social genesis of machine intelligence: emergence, simulation, and opinion formation in autonomous agent societies

Article References: Freund, L. (2026). The social genesis of machine intelligence: emergence, simulation, and opinion formation in autonomous agent societies. AI & SOCIETY. https://doi.org/10.1007/s00146-026-03360-8

Image Credits: AI Generated

DOI: 10.1007/s00146-026-03360-8

Keywords: AI agents, emergence, machine opinion, multi-agent systems, large language models, opinion dynamics, distributed cognition, collective intelligence, stigmergy, simulation, AI governance, AI & Society