Abstract We consider the learning of formulas in the model of Bārzdiņš et al. [Proc. EuroCOLT'97, 1997, p. 272] and Osherson et al. [J. Symbolic Logic 56 (1991) 661]. We show that, in this model, a formula f can be learned in the limit if and only if ¬f can be learned with assurance.