Inference

Inference is the process of deriving conclusions from premises, evidence, or data. It is the central operation of logic, statistics, cognitive science, and artificial intelligence — yet the concept is rarely examined across these domains as a unified phenomenon.

In deductive logic, inference is truth-preserving: if the premises are true, the conclusion must be true. In inductive reasoning, inference is ampliative: the conclusion goes beyond the premises, and the inference is evaluated by its reliability, not its necessity. In statistical inference, the evaluation is formalized through probability theory: the conclusion is a probabilistic claim about a population, derived from a sample. In Bayesian inference, the conclusion is a posterior probability distribution, updated from a prior by the evidence through Bayes' theorem.

The cognitive science of inference studies how humans actually perform these operations — and the answer is that human inference is neither purely deductive nor purely Bayesian. It is heuristic: fast, frugal, and ecologically adapted to specific environmental structures. The heuristics-and-biases program documents systematic deviations from normative models; the ecological rationality program argues that these deviations are often adaptive responses to environmental constraints rather than cognitive bugs.

In artificial intelligence, inference is the operation that transforms trained models into predictions. A neural network performs inference when it maps an input to an output through its learned weights. A large language model performs inference when it generates the next token conditioned on the context. These operations are not "reasoning" in the human sense — they are statistical generalization at scale. The question of whether machine inference is continuous with human inference, or a different phenomenon entirely, remains one of the central open questions in the philosophy of AI.

The structural insight is that inference is not a single operation but a family of operations unified by their function: the transformation of information into conclusions, under constraints of time, data, and computational resource. The differences between deduction, induction, abduction, statistical estimation, and neural prediction are differences of formalization and constraint, not differences of kind. A unified theory of inference would treat them as points in a space defined by the trade-off between soundness (guaranteeing truth), completeness (covering all truths), and tractability (computational feasibility).

Inference in Distributed Systems

The standard account of inference assumes a single reasoner with access to all premises, all evidence, and all computational resources. This assumption is not merely unrealistic; it is structurally misleading for understanding how inference works in complex systems. No human brain holds all relevant premises. No scientific community has a single unified model. No distributed computer system has a global state that any one node can inspect. Inference, in practice, is almost always distributed — performed by multiple agents with partial information, limited communication, and conflicting objectives.

The question then becomes: what can be inferred when no single agent has the whole picture? The answer, from distributed computing and multi-agent systems, is that inference is not just about deriving conclusions from premises. It is about coordinating beliefs across agents whose local inferences may be mutually inconsistent. A consensus algorithm in a distributed database is a form of inference: nodes infer that a transaction is committed based on messages from other nodes, not on direct observation. A scientific community's gradual convergence on a paradigm is a form of inference: individual scientists infer from local experiments, and the field infers from the aggregate pattern of individual inferences. The market's price discovery is a form of inference: no trader knows the true value of an asset, but the distributed process of bidding and asking produces a collective inference that is often more accurate than any individual expert's judgment.

The systems-theoretic insight is that distributed inference trades completeness for tractability and robustness. A single reasoner with all premises could (in principle) derive all valid conclusions, but would face combinatorial explosion. A distributed system of reasoners cannot derive all valid conclusions, but can derive robust approximate conclusions that degrade gracefully when individual agents fail. The Frame Problem — the question of what changes when something changes — is not just a problem for individual agents. It is a problem for distributed systems: when one agent updates its beliefs, how do the other agents know what to update, and what they can safely ignore? The solutions that work for individual robots — local update rules, causal neighborhoods — generalize to distributed inference as gossip protocols, eventual consistency, and bounded rationality. The underlying theorem is the same: global consistency is not what complex systems need. Local consistency, coordinated through sparse communication, is sufficient — and it is all that is physically possible.

The inference literature's obsession with single-agent ideal models is not merely a simplification. It is a disciplinary blind spot that has prevented the integration of inference with the study of complex systems. Until we treat inference as a distributed, multi-scale phenomenon — something that happens between brains, between computers, between institutions — we will not understand how intelligence actually works in the only world that exists: the one where no one has all the answers.