Collective and Animal Intelligence

Collective intelligence is the capacity of a group of agents, whether people, animals or software, to solve problems or make decisions in ways that depend on the group's organization and not only on the abilities of its individual members. Animal intelligence is the study of problem-solving, learning and social cognition in non-human species. Together these topics supply two contrasts for AI: intelligence that is not human in kind, and intelligence that is not located in a single agent.

Definitions of intelligence built around a single human mind (see FOUND-002 Definitions of Intelligence) leave open two questions that AI research has to answer in practice. First, which cognitive abilities are specific to humans, and which are shared with other animals whose brains have very different structure? Second, can an aggregate of simple or partially informed agents exhibit competence that none of its members has?

Both questions bear on AI design. Comparative cognition gives a set of tasks and a vocabulary (causal reasoning, flexibility, prospection, social cognition) that is reused when AI systems are assessed, and it shows that complex cognition can arise in distantly related lineages with different neural architecture. Collective intelligence supplies mechanisms (stigmergy, voting, diversity of problem-solving methods) that have been turned into algorithms, and it frames current multi-agent LLM systems and the organizational view of AI deployment.

The topic also requires care about claims. Metaphors from biology ("hive mind", "swarm") are widely used in AI discussions, but the empirical literature on group performance is narrower than those metaphors suggest.

Early quantitative work on group judgement is usually traced to Francis Galton's 1907 note "Vox Populi" in Nature, which analysed estimates of an ox's weight made by visitors to a livestock fair and found that the central estimate of the crowd was close to the true value (Galton, 1907). The observation is the classic illustration of error cancellation in aggregated independent judgements.

Comparative psychology developed through the twentieth century around tool use, memory and social learning in primates and other vertebrates. A widely cited review by Emery and Clayton (2004) argued that corvids (crows, jays and relatives) show tool manufacture, mental time travel and social cognition comparable in kind to that studied in apes, and that such abilities evolved more than once in lineages with very different brain structure.

In computing, Reynolds (1987) showed that flocking behaviour in simulated "boids" could be generated from local steering rules, commonly described as separation, alignment and cohesion, without a leader. Dorigo, Maniezzo and Colorni (1996) proposed the Ant System, an optimization method inspired by pheromone-mediated foraging in ant colonies. These methods are now commonly grouped under the label "swarm intelligence" (the origin of the term is not traced in this article). In parallel, Hong and Page (2004) gave a formal model of group problem-solving in which diversity of method can outweigh individual ability, and Woolley et al. (2010) proposed a measurable collective intelligence factor for small human groups, a proposal later examined by Riedl et al. (2021) and questioned by Credé and Howardson (2017). Legg and Hutter (2007) offered a definition of machine intelligence abstract enough to apply to agents of any kind, which is one route to comparing animal, human, collective and artificial intelligence on a single scale.

#4.1 Conditions for crowd accuracy

Suppose $n$ agents each give an estimate $x_i = \theta + \epsilon_i$ of a quantity $\theta$, where the errors $\epsilon_i$ have zero mean and variance $\sigma^2$. If the errors are independent, the variance of the mean estimate is

If errors share a common pairwise correlation $\rho$, the variance becomes

which tends to $\rho\sigma^2$ as $n$ grows. Aggregation therefore removes independent error but not shared bias, which is why the independence and diversity of members matter more than their number. This is a statistical identity, not an empirical claim about any particular crowd.

#4.2 Diversity and problem-solving teams

Hong and Page (2004) model agents as having different representations of a problem and different search heuristics. They show that, under their assumptions, a team drawn at random from a diverse population can outperform a team consisting of the individually best performers, because the best individuals tend to be similar to one another. The result is a mathematical property of the model. Whether real teams meet its assumptions is an empirical matter, and the model should not be read as showing that diversity always beats ability.

#4.3 Stigmergy and local rules

In stigmergy, agents coordinate indirectly by modifying a shared environment. In the Ant System, artificial ants deposit "pheromone" on edges of a graph in proportion to solution quality, and later ants choose edges with probability weighted by pheromone and a heuristic desirability, with evaporation preventing premature convergence (Dorigo et al., 1996). The coordination is carried by the environment's state, not by messages between agents. Flocking models achieve coordinated motion from purely local rules (Reynolds, 1987).

#4.4 Measuring collective intelligence in people

Woolley et al. (2010) applied the logic of psychometric factor analysis to groups. In two studies with 699 people working in groups of two to five, they found converging evidence of a general factor, termed c, that explains group performance across a variety of tasks. They reported that c was not strongly correlated with the average or maximum individual intelligence of members, but was correlated with average social sensitivity, equality of conversational turn-taking, and the proportion of women in the group.

Collective intelligence in biological systems.

• Insect colonies: ant foraging via pheromone trails; honeybee nest-site selection, studied in detail by Seeley (2010), in which scouts advertise candidate sites and the swarm reaches a quorum.
• Vertebrate groups: schools, flocks and herds, where local interaction rules can transmit information across the group.

Collective intelligence in human groups. Aggregation of judgements (polls, prediction markets, "wisdom of crowds"), deliberating teams, and large distributed projects. The field is surveyed in Malone and Bernstein (2015).

Animal intelligence, by domain. Physical cognition (tool use, causal reasoning), social cognition (cooperation, deception, perspective-taking), memory and planning (caching behaviour, "mental time travel" in corvids), and communication. The review by Emery and Clayton (2004) proposes that complex cognition depends on a "tool kit" of causal reasoning, flexibility, imagination and prospection.

Algorithmic derivatives.

• Ant colony optimization (Dorigo et al., 1996).
• Particle swarm optimization and related population methods (not covered by a source in this article's reference list; see TECH-MARL-001 Multi-Agent Systems for the agent-based treatment).
• Multi-agent systems and ensembles in machine learning, in which several models are combined.

• Optimization. Ant-colony methods were proposed for combinatorial problems such as the travelling salesman problem (Dorigo et al., 1996). They are one of several metaheuristic families; the Ant System paper itself reports results on small benchmark instances, and the article does not claim superiority over other heuristics.
• Computer graphics. Reynolds's flocking model became a standard technique for animating groups of creatures (Reynolds, 1987).
• Organizational design. The collective intelligence factor motivates research on team composition and communication norms (Woolley et al., 2010; Malone & Bernstein, 2015).
• Comparative benchmarks. Animal-cognition tasks (object permanence, tool use, delayed gratification) are used as reference points when researchers discuss which abilities current AI systems lack; this use is conceptual, and standardized cross-species and AI benchmarks remain rare.
• Multi-agent AI. Ensembles and multi-agent LLM pipelines are sometimes described as "collective intelligence". Whether such systems display the properties documented for human or insect collectives, as opposed to being engineered pipelines, is a matter of definition and is discussed in TECH-AGENT-001 Autonomous Agents.

Strengths. Distributed systems are robust to the loss of individual members, scale by adding agents, and can combine partial information. Aggregation reduces independent error (Section 4.1).

Limitations.

• Correlated errors defeat aggregation. Herding, shared information sources and social influence reduce the independence the statistical argument requires.
• The collective intelligence factor was proposed from two studies of small groups (Woolley et al., 2010). A later analysis of 22 studies (5,279 individuals in 1,356 groups) by a team including the original authors reports support for a general factor (Riedl et al., 2021), while a published comment disputes the original analysis (Credé & Howardson, 2017; the content of that comment was not reviewed for this audit). It should be treated as a proposed construct whose structure is still debated.
• Diversity results such as Hong and Page (2004) are conditional on model assumptions.
• Comparative cognition faces an inference problem: behaviour consistent with causal reasoning may be explained by simpler mechanisms, and the same task can recruit different mechanisms across species. The conclusion that corvids and apes "share cognitive tools" is an interpretation drawn from reviewed studies and is open to alternative accounts.

Open questions. Whether a general intelligence factor is the right model for groups; how to compare intelligence across species and across biological and artificial systems without anthropocentric task design; and whether collectives of AI agents will show emergent coordination failures or competencies not predictable from single-agent evaluations.

As of 2026-10-15, the following can be stated with differing confidence (classification in Section 12).

• The statistical benefit of aggregating independent estimates is a mathematical result.
• Swarm-inspired algorithms are established as a family of optimization heuristics; their relative performance is problem-dependent.
• Complex cognition in corvids and in apes, and the evolutionary independence of those capacities, is supported by a body of studies reviewed by Emery and Clayton (2004); later work may have revised parts of that picture, and this article has not systematically reviewed it.
• The existence and structure of a collective intelligence factor in human groups is supported by the original studies (Woolley et al., 2010) and by a 22-study analysis from a group that includes the original authors (Riedl et al., 2021); independent assessment is represented here only by a published critical comment (Credé & Howardson, 2017), which re-analysed six previously published samples and concluded that support for the construct is weak. Other independent replications were not reviewed.

This section makes no registered forecast. Observed trends and open directions:

• Near term (0-3 years). Continued use of multi-agent LLM systems; evaluation methods for groups of AI agents remain immature, so claims of emergent collective competence should be checked against single-agent baselines.
• Medium term (3-10 years). Closer integration of comparative-cognition tasks into AI evaluation is plausible but not assured; it depends on whether the field adopts shared protocols.
• Long term (10+ years). Speculative scenarios in which collectives of AI systems coordinate at scale belong to the safety and governance literature (see SAFE-XRISK-001 Existential Risk from AI); they are hypothetical.

• Foundational Concepts:
  • FOUND-002 Definitions of Intelligence: the definitions that group and animal intelligence stretch.
  • FOUND-001 Artificial Intelligence: the discipline within which these analogies are used.
• Related Methods & Technologies:
  • TECH-MARL-001 Multi-Agent Systems: engineered multi-agent coordination.
  • TECH-AGENT-001 Autonomous Agents: single and multiple agent architectures.
  • FOUND-012 Game Theory and Decision Theory: formal models of cooperation and competition among agents.
  • FOUND-013 The Brain and Biological Neurons: the neural substrate of animal cognition.
• Advanced Implications & Future Trajectories:
  • AGI-001 Artificial General Intelligence: questions of generality across species and systems.
  • SAFE-XRISK-001 Existential Risk from AI: multi-agent and collective risk scenarios.