Abstract. This paper studies the loss of the semimartingale property of the process g(Y )at the time a one-dimensional diffusion Y hits a level, where g is a difference of two convex functions. We show that the process g(Y ) can fail to be a semimartingale in two ways only, which leads to a natural definition of non-semimartingales of the first and second kind. We give a deterministic if and only if condition (in terms of g and the coefficients of Y ) for g(Y ) to fall into one of the two classes of processes, which yields a characterisation for the loss of the semimartingale property. A number of applications of the results in the theory of stochastic processes and real analysis are given: e.g. we construct an adapted diffusion Y on [0, ∞) and a predictable finite stopping time ζ, such that Y is a local semimartingale on the stochastic interval [0, ζ), continuous at ζ and constant after ζ, but is not a semimartingale on [0, ∞).