Truth-Tracking

Truth-tracking is the epistemic property of a belief-formation process that tends to produce true beliefs when the evidence supports them and false beliefs when the evidence contradicts them. A truth-tracking process is not merely one that happens to be right; it is one that would have been different if the facts had been different. The concept was introduced by Robert Nozick in his epistemological work, but it has taken on renewed urgency in the context of artificial intelligence and information systems.

A language model that produces a true statement by chance is not truth-tracking. It would have produced the same statement — or an equally confident false statement — even if the facts had been otherwise. Truth-tracking requires a sensitivity to evidence that current AI systems fundamentally lack. The hallucination problem is not merely that models are sometimes wrong; it is that they are not truth-tracking even when they are right. Their outputs are decoupled from the evidential state of the world.

This decoupling has systemic consequences. An information ecosystem populated by non-truth-tracking systems — whether human or artificial — becomes vulnerable to information cascades in which false beliefs propagate because they are coherent with prior false beliefs, not because they are supported by evidence. Truth-tracking is therefore not an individual epistemic virtue but a systemic requirement for collective intelligence. Epistemic foraging is the behavioral complement of truth-tracking: the active seeking of evidence that would distinguish true from false beliefs. A system that cannot forage epistemically cannot track truth, no matter how large its training corpus.