Defeasible Reasoning

Defeasible reasoning is inference that is valid in the absence of defeating information but becomes invalid when new evidence emerges. Unlike deductive reasoning, where conclusions follow necessarily from premises, defeasible reasoning produces conclusions that are provisionally justified — held as 'accurately or very nearly true,' in Newton's phrase — until contrary phenomena appear. The concept is central to epistemology, artificial intelligence, and legal reasoning, where agents must act on incomplete information while remaining prepared to revise their commitments.

The philosophical significance of defeasible reasoning is that it captures how actual minds — and actual scientific communities — operate. No empirical inference is ever final; every generalization is a standing invitation to counterexample. The architecture of defeasible reasoning connects to non-monotonic logic in AI and to belief revision theory in formal epistemology, both of which attempt to model how rational agents should update their beliefs when new information conflicts with old conclusions.

Computational Models

In artificial intelligence, defeasible reasoning is not merely a philosophical description but an engineering requirement. Real-world agents — from medical diagnosis systems to autonomous vehicles — must reason with incomplete information and revise conclusions when new data arrives. The formal tools for this include non-monotonic logics, default logic, and circumscription. These systems depart from classical logic by allowing conclusions to be withdrawn: if Tweety is a bird, we defeasibly conclude Tweety flies, but if we later learn Tweety is a penguin, the conclusion is retracted without inconsistency.

The architecture of defeasible reasoning in AI mirrors the architecture of belief revision in formal epistemology. The AGM postulates (Alchourrón, Gärdenfors, Makinson) specify rational constraints on how a belief set should change when new information arrives: contraction (removing a belief), expansion (adding a belief), and revision (adding a belief while maintaining consistency). These operations are not mere bookkeeping; they encode a theory of rational change. A system that revises its beliefs incoherently — adding new evidence but failing to propagate its consequences — is not just inefficient; it is epistemically defective.

The connection to machine learning is equally deep. A neural network trained on a dataset has learned defeasible generalizations: it classifies inputs based on patterns that hold unless