Induction
Induction moves from particular observations to general conclusions. This copper wire conducts electricity; that copper wire conducts electricity; therefore, probably, copper conducts electricity. The conclusion goes beyond the evidence — it extends from observed cases to unobserved cases — and this extension introduces risk. The generalization may be wrong. Induction is productive in a way deduction is not, because it yields genuinely new general propositions, but it is fallible, because the new propositions are never guaranteed by the evidence that supports them. Contemporary AI systems perform operations that are functionally inductive on scales that dwarf anything Peirce imagined — but they induce without knowing that they induce, without understanding the risk entailed, and without the capacity to recognize their own failures.
In the AI Story
A large language model trained on billions of tokens has, in effect, performed inductions over the entire accessible corpus of human writing — extracting statistical regularities, identifying patterns of co-occurrence, generalizing from observed sequences to predicted sequences. The predictions are often remarkably accurate. They are also, in the strict Peircean sense, fallible: the model's generalizations may fail on any particular case.
The crucial point for the AI debate is that the machine's inductions lack the self-awareness that human inductions have. The human inquirer who generalizes from observed cases knows that she is generalizing, understands that the generalization may fail, and maintains the epistemic stance appropriate to fallible inference. The machine does not. It produces outputs with the same statistical confidence regardless of whether the underlying pattern is robust or fragile.
Induction is the operation AI performs most visibly and most powerfully. The pattern-matching capabilities of neural networks are essentially large-scale induction, and the extraordinary performance of modern systems on tasks that require pattern recognition testifies to the mechanizability of at least this mode of inference.
But induction alone, without abduction to generate new hypotheses and without Secondness to test them, is not inquiry. It is pattern extrapolation, and pattern extrapolation from an incomplete training set will systematically miss precisely the phenomena that new hypotheses would illuminate.
Origin
Peirce's treatment of induction evolved significantly across his career. Early Peirce treated induction as a single mode; mature Peirce distinguished three forms — crude induction, quantitative induction, and qualitative induction — each with distinct logical structures.
The mechanization of induction is older than the AI field itself — Bayesian inference, statistical learning theory, and modern machine learning all trace back to the same formal problem Peirce analyzed.
Key Ideas
Ampliative but fallible. Generates genuine new knowledge but without guarantee — the conclusion can always be wrong.
Pattern-extension, not pattern-creation. Extends observed regularities; does not propose new ones (which would be abduction).
Machine-executable at scale. AI performs induction over corpora no human could examine — but without awareness that it is inducing.
Requires testing. Inductive conclusions are hypotheses that must face experience; the AI system generates them without the testing loop.
Appears in the Orange Pill Cycle
Further reading
- Charles Sanders Peirce, "Deduction, Induction, and Hypothesis" (1878)
- Charles Sanders Peirce, "The Probability of Induction" (1878)
- Nelson Goodman, Fact, Fiction, and Forecast (Harvard, 1955)
- John Norton, The Material Theory of Induction (BSPS, 2021)