Data on the past intensity of Earth’s magnetic field (paleointensity) are essential for understanding Earth’s deep interior, climatic modeling, and geochronology applications, among other items. Here we demonstrate the possibility that much of available paleointensity data could be biased by instability of thermoremanent magnetization (TRM) associated with non-single-domain (SD) particles. Paleointensity data are derived from experiments in which an ancient TRM, acquired in an unknown field, is replaced by a laboratory-controlled TRM. This procedure is built on the assumption that the process of ancient TRM acquisition is entirely reproducible in the laboratory. Here we show experimental results violating this assumption in a manner not expected from standard theory. We show that the demagnetization−remagnetization relationship of non-SD specimens that were kept in a controlled field for only 2 y show a small but systematic bias relative to sister specimens that were given a fresh TRM. This effect, likely caused by irreversible changes in micromagnetic structures, leads to a bias in paleointensity estimates.