论文信息
Original title: Causal Foundations of Collective Agency Authors: Frederik Hytting Jørgensen, Sebastian Weichwald, Lewis Hammond Published: 2026-04-30 arXiv: 2604.21518 Categories: Artificial Intelligence, Computer Science and Game Theory, Multiagent Systems
---
Opening: The Ant Colony Puzzle
Imagine watching a line of ants carry a dead insect back to their nest on a summer afternoon. Each ant is tiny and simple, following local chemical signals. But as a whole, the colony displays remarkable intelligence—finding shortest paths, allocating workload, responding to disturbances.
The question: Is the colony itself an agent?
If a single ant deviates from its route, it is just individual behavior. But if the entire colony, as a system, exhibits goal-directed adaptation—is it *choosing* an optimal strategy?
This sounds like philosophical chatter, but it is urgent in the AI era. Consider:
- Hundreds of autonomous vehicles coordinating on roads, forming a "traffic-flow agent"
- Thousands of recommendation algorithms interacting on social platforms, shaping "collective opinion"
- Multiple simple AI assistants collaborating on a task and exhibiting emergent behavior their developers never intended
- It has goals (preferences, a utility function)
- Its behavior can be rationalized—viewed as optimal means toward those goals
- It adaptively adjusts behavior as the environment changes
- Multiple agents (players)
- Each agent's decision variables (actions)
- The causal structure of the environment (how actions affect outcomes)
- Each agent's utility function
- They may have collective goals developers never anticipated
- Those goals may conflict with human interests
- Individual-level safety constraints do not guarantee collective-level safety
- Jørgensen, F. H., Weichwald, S., & Hammond, L. (2026). Causal Foundations of Collective Agency. *arXiv preprint arXiv:2604.21518*.
- Pearl, J. (2009). *Causality: Models, Reasoning, and Inference* (2nd ed.). Cambridge University Press.
- Binmore, K. (2007). *Playing for Real: A Text on Game Theory*. Oxford University Press.
- Metzinger, T. (2003). *Being No One: The Self-Model Theory of Subjectivity*. MIT Press.
- Wolfram, S. (2002). *A New Kind of Science*. Wolfram Media.
When should collective behavior be regarded as a "new agent"?
This paper gives a rigorous mathematical answer using causal games and causal abstraction.
---
What Is Agency?
In AI and philosophy, agency is hard to define, but the core intuition is: *an agent is an entity capable of making rational choices according to goals*.
More precisely, a system is an agent if:
The hard question: when can a group of individuals be viewed *collectively* as an agent?
---
Intuitive Pitfalls of Collective Agency
Pitfall 1: Group size. Two collaborators? A thousand AI nodes? Where is the line?
Pitfall 2: Emergent properties. An ant colony has the collective goal of "nest maintenance," but no single ant consciously works for it. Recommendation algorithms have no intent to "manipulate opinion," yet the collective effect is polarization. Intentions at the individual and collective levels can differ entirely.
Pitfall 3: Safety concerns. If multiple simple AIs can "inadvertently" form a collective agent with new capabilities and goals, how do we predict and control it? This is a core AI safety question.
---
The Behavioral Perspective: Defining Agency by Prediction
The paper adopts a clever behavioral perspective:
> A group has collective agency when treating its joint behavior as rational and goal-directed successfully predicts that behavior.
This does not say the collective "has" a goal—it says that modeling the collective as a goal-directed agent is an effective predictive strategy.
This resembles effective theories in physics: gas is made of countless molecules obeying quantum mechanics, but describing it as a continuous fluid with pressure and temperature is effective. Similarly, "group will" may be a valid macro-level description even when only individual behavior exists at the micro level.
---
Causal Games: Formalizing Group Interaction
Causal Models
Causal models, popularized by Pearl, describe causal relations between variables (e.g., smoking → lung cancer, not vice versa—correlation is not causation). Their key feature is interventional reasoning: "If we force someone to smoke, what is the probability of lung cancer?"—which differs from conditional probabilities in observational data.
Game Theory
Game theory studies strategic interaction among rational decision-makers: the Prisoner's Dilemma, coordination games, and so on. Each player has a utility function and picks strategies maximizing expected utility.
Causal Games = Causality + Games
A causal game combines both: it describes not only how players choose strategies, but the causal structure of choice. A causal game includes:
This lets us ask causal questions: "If agent A changes strategy, how will agent B respond?"—causal effects, not mere statistical correlation.
---
Causal Abstraction: The Bridge from Micro to Macro
Abstraction extracts essence from detail. A map preserves distances and relative positions while ignoring the shape of every stone. In causal models, abstraction means going from detailed micro variables (each ant's position and velocity) to coarse-grained macro variables (the colony's average position and direction) while *preserving causal structure*: macro interventions' effects on macro variables are consistent with micro interventions.
Key Theorem
> A group constitutes a collective agent if and only if its joint behavior can be rationalized at some level of causal abstraction.
In other words: if there exists a high-level causal model in which the collective is modeled as an agent with a utility function, and this high-level model's predictions are consistent with the low-level (individual-level) causal model, then the collective is a genuine agent—not as a metaphor, but mathematically.
---
Applications
Collective Agency of Voting Mechanisms
Simple majority voting: individually, voters vote by preference; collectively, the group's "preference" is a function of the outcome. But simple majority can violate transitivity—A beats B, B beats C, yet C beats A (the Condorcet paradox). Without a consistent collective "preference," collective agency is weak.
Ranked choice voting: more complex rules that, under certain conditions, yield a consistent collective preference—hence stronger collective agency.
This yields a deep insight for institutional design: some institutions naturally produce "collective will," others do not.
The Incentive Puzzle in Multi-Agent Actor-Critic
In multi-agent Actor-Critic, each agent has a critic estimating value. If critics consider only individual rewards, agents may act against collective interests. The causal abstraction framework shows this problem is equivalent to "individual utility functions cannot be causally abstracted into a collective utility function." The fix requires redesigning critics to be consistent with the collective-level causal model.
---
Deep Implications for AI Safety
Emergent Risk
If multiple simple AI agents can form a collective agent:
Traditional AI safety focuses on single models. But future AI systems may be decentralized multi-agent networks—every node safe does not mean the whole is safe.
Detecting Collective Agency
The framework offers detection tools: observe the group's joint behavior; try to predict it with a model in which the collective has goals; if predictions are accurate, collective agency exists. Like physicists inferring atoms from Brownian motion, we infer collective will from behavioral patterns.
Control Strategies
If we can identify emergent collective agency, we can design institutions so that "good" collective agents emerge and "bad" ones do not, embed constraints in voting mechanisms, resource allocation, and communication protocols, and monitor macro-level behavior of multi-agent systems with early warnings.
---
Philosophical Echoes: From Leibniz to Metzinger
Leibniz's Monads
Leibniz saw the world as composed of "monads"—independent, windowless substances coordinated by "pre-established harmony." Collective agency becomes: how do monads form unions? Causal abstraction gives a modern answer: collectivity is not an ontological "existence" but an epistemic effectiveness—whether modeling the group as an agent has predictive power.
Metzinger's Self-Model Theory
Thomas Metzinger argues the "self" is not an entity but a phenomenal self-model constructed by the brain. Collective agency may be similar—not a collective mind the group "has," but an agent model observers project onto group behavior that happens to work. This fits the paper's behavioral view: collective agency is not a metaphysical entity but a predictive fiction—a useful one.
---
Outlook: A New Era of Multi-Agent AI
Short term: better multi-agent system design (voting, resource allocation, coordination protocols); new AI safety assessment tools (detecting emergent collective agents); theoretical foundations for human-AI hybrid teams.
Long term: if we enter an era of thousands of AI nodes interacting across the internet, understanding the causal foundations of collective agency will be a survival-level skill. This is not science fiction—today's financial markets, social platforms, and logistics systems are already proto-multi-agent systems, and AI will only accelerate the trend.
---
References
*Feynman-style deep dive | Auto-collected 2026-05-05*