CONCEPT
Unfalsified Confidence
The specific epistemic danger of AI output—confidence that has not survived any attempt at refutation and therefore carries no genuine epistemic weight, yet arrives formatted as knowledge and is experienced as authoritative.
Karl Popper's philosophy of science rests on a distinction that most people intuitively accept but rarely articulate with precision: there is a difference between a claim that has never been tested and a claim that has been tested and survived. Both may be true. Both may be stated with the same words and the same confidence. But they earn trust through entirely different mechanisms, and confusing them is the fundamental epistemological error of the age. A theory that has been subjected to severe testing and has not yet failed has earned the provisional trust that Popper called the weight of the unfalsified—not certainty, but a specific reliability grounded in the history of attempts to disprove it. A claim that has never been tested earns nothing equivalent, however plausible it sounds. Unfalsified confidence is the name for the specific kind of confidence that
large language models generate: outputs produced without any internal mechanism of self-testing, presented in the register of knowledge, indistinguishable by surface quality